research opportunities arising from control and ... · research opportunities arising from control...
TRANSCRIPT
Research opportunities arising from control
and optimization of smart buildings
Qianchuan Zhao
CFINS Dept Automation and TNList
Tsinghua University
zhaoqctsinghuaeducn
20170825
ISA Workshop on Frontiers in Systems and Control
CFINS
2
Yu-Chi Ho chair professor group and ldquoCenter for Intelligent and Networked Systemsrdquo (CFINS) were established in October 2001 to provide a physical and intellectual environment for the intelligent analysis design and operation of complex and networked systems such as computer and communication networks buildings power systems and supply chains by making innovative use of analytical methods and information technology
httpcfinsautsinghuaeducn
What we mean by smart for buildings
3
Potential of ICT technology
4
bull MOORErsquos Law
bull AI
What we mean by smart for buildings
5
bull Utilize information relevant to the whole building system thanks to IoT as a result of the fast drop in the cost of hardware for computing storage and communication
bull Care about individual occupant thanks to the rapid development of machine learning techniques
Energy consumption
Type faction
Building40
(68Electr )
Transportation 40
Others 20
Energy saving for buildings has been
omitted for long it has great potential 6
Energy consumption in buildings
It was estimated that 20 ~ 30
energy saving can be achieved by
optimizing the operation and
control of buildings
Office
BuildingHVAC37
28
Office Equipments
22
Elevator
3Other10
Lights
Hotel
44
254
9
18
Lights
HVAC
Office Equipments
Elevator
Other
7
System Architecture
Information fusion
Data driven modeling + prediction
Integrated building control
for energy saving
8
Control and optimization of building energy system
Energy supply in building
Distribution Battery
CHP
Wind
E-car
Fuel cellSolar
Lighting HVAC
Shading Window
Controllable devices
Elect
Heat
ComfortTemp HumidIllum CO2
Occupant demand
Minimization of energy cost
Micro-grid
9
List of possible challenges
10
bull Integrated control under full information may suffer the curse of dimensionality problem and time consuming evaluation of performance or constraints
bull Mache learning in general is a hard problem design of a good ML algorithm also include many decision variables (model structure parameters implementation input data hellip)
Ways to address the challenges
11
According to NFLT problem specific knowledge is needed to develop efficient solutions
bull Soft optimization for integrated control OO OCBA COO NP ADP EBO IPA hellip
bull Apply problem specific knowledge to reduce the search space for a good ML algorithm
Illustration of COO
G
S N
12
13
Below we will use individual thermal comfort model as an example of ML in smart building applications
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
CFINS
2
Yu-Chi Ho chair professor group and ldquoCenter for Intelligent and Networked Systemsrdquo (CFINS) were established in October 2001 to provide a physical and intellectual environment for the intelligent analysis design and operation of complex and networked systems such as computer and communication networks buildings power systems and supply chains by making innovative use of analytical methods and information technology
httpcfinsautsinghuaeducn
What we mean by smart for buildings
3
Potential of ICT technology
4
bull MOORErsquos Law
bull AI
What we mean by smart for buildings
5
bull Utilize information relevant to the whole building system thanks to IoT as a result of the fast drop in the cost of hardware for computing storage and communication
bull Care about individual occupant thanks to the rapid development of machine learning techniques
Energy consumption
Type faction
Building40
(68Electr )
Transportation 40
Others 20
Energy saving for buildings has been
omitted for long it has great potential 6
Energy consumption in buildings
It was estimated that 20 ~ 30
energy saving can be achieved by
optimizing the operation and
control of buildings
Office
BuildingHVAC37
28
Office Equipments
22
Elevator
3Other10
Lights
Hotel
44
254
9
18
Lights
HVAC
Office Equipments
Elevator
Other
7
System Architecture
Information fusion
Data driven modeling + prediction
Integrated building control
for energy saving
8
Control and optimization of building energy system
Energy supply in building
Distribution Battery
CHP
Wind
E-car
Fuel cellSolar
Lighting HVAC
Shading Window
Controllable devices
Elect
Heat
ComfortTemp HumidIllum CO2
Occupant demand
Minimization of energy cost
Micro-grid
9
List of possible challenges
10
bull Integrated control under full information may suffer the curse of dimensionality problem and time consuming evaluation of performance or constraints
bull Mache learning in general is a hard problem design of a good ML algorithm also include many decision variables (model structure parameters implementation input data hellip)
Ways to address the challenges
11
According to NFLT problem specific knowledge is needed to develop efficient solutions
bull Soft optimization for integrated control OO OCBA COO NP ADP EBO IPA hellip
bull Apply problem specific knowledge to reduce the search space for a good ML algorithm
Illustration of COO
G
S N
12
13
Below we will use individual thermal comfort model as an example of ML in smart building applications
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
What we mean by smart for buildings
3
Potential of ICT technology
4
bull MOORErsquos Law
bull AI
What we mean by smart for buildings
5
bull Utilize information relevant to the whole building system thanks to IoT as a result of the fast drop in the cost of hardware for computing storage and communication
bull Care about individual occupant thanks to the rapid development of machine learning techniques
Energy consumption
Type faction
Building40
(68Electr )
Transportation 40
Others 20
Energy saving for buildings has been
omitted for long it has great potential 6
Energy consumption in buildings
It was estimated that 20 ~ 30
energy saving can be achieved by
optimizing the operation and
control of buildings
Office
BuildingHVAC37
28
Office Equipments
22
Elevator
3Other10
Lights
Hotel
44
254
9
18
Lights
HVAC
Office Equipments
Elevator
Other
7
System Architecture
Information fusion
Data driven modeling + prediction
Integrated building control
for energy saving
8
Control and optimization of building energy system
Energy supply in building
Distribution Battery
CHP
Wind
E-car
Fuel cellSolar
Lighting HVAC
Shading Window
Controllable devices
Elect
Heat
ComfortTemp HumidIllum CO2
Occupant demand
Minimization of energy cost
Micro-grid
9
List of possible challenges
10
bull Integrated control under full information may suffer the curse of dimensionality problem and time consuming evaluation of performance or constraints
bull Mache learning in general is a hard problem design of a good ML algorithm also include many decision variables (model structure parameters implementation input data hellip)
Ways to address the challenges
11
According to NFLT problem specific knowledge is needed to develop efficient solutions
bull Soft optimization for integrated control OO OCBA COO NP ADP EBO IPA hellip
bull Apply problem specific knowledge to reduce the search space for a good ML algorithm
Illustration of COO
G
S N
12
13
Below we will use individual thermal comfort model as an example of ML in smart building applications
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Potential of ICT technology
4
bull MOORErsquos Law
bull AI
What we mean by smart for buildings
5
bull Utilize information relevant to the whole building system thanks to IoT as a result of the fast drop in the cost of hardware for computing storage and communication
bull Care about individual occupant thanks to the rapid development of machine learning techniques
Energy consumption
Type faction
Building40
(68Electr )
Transportation 40
Others 20
Energy saving for buildings has been
omitted for long it has great potential 6
Energy consumption in buildings
It was estimated that 20 ~ 30
energy saving can be achieved by
optimizing the operation and
control of buildings
Office
BuildingHVAC37
28
Office Equipments
22
Elevator
3Other10
Lights
Hotel
44
254
9
18
Lights
HVAC
Office Equipments
Elevator
Other
7
System Architecture
Information fusion
Data driven modeling + prediction
Integrated building control
for energy saving
8
Control and optimization of building energy system
Energy supply in building
Distribution Battery
CHP
Wind
E-car
Fuel cellSolar
Lighting HVAC
Shading Window
Controllable devices
Elect
Heat
ComfortTemp HumidIllum CO2
Occupant demand
Minimization of energy cost
Micro-grid
9
List of possible challenges
10
bull Integrated control under full information may suffer the curse of dimensionality problem and time consuming evaluation of performance or constraints
bull Mache learning in general is a hard problem design of a good ML algorithm also include many decision variables (model structure parameters implementation input data hellip)
Ways to address the challenges
11
According to NFLT problem specific knowledge is needed to develop efficient solutions
bull Soft optimization for integrated control OO OCBA COO NP ADP EBO IPA hellip
bull Apply problem specific knowledge to reduce the search space for a good ML algorithm
Illustration of COO
G
S N
12
13
Below we will use individual thermal comfort model as an example of ML in smart building applications
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
What we mean by smart for buildings
5
bull Utilize information relevant to the whole building system thanks to IoT as a result of the fast drop in the cost of hardware for computing storage and communication
bull Care about individual occupant thanks to the rapid development of machine learning techniques
Energy consumption
Type faction
Building40
(68Electr )
Transportation 40
Others 20
Energy saving for buildings has been
omitted for long it has great potential 6
Energy consumption in buildings
It was estimated that 20 ~ 30
energy saving can be achieved by
optimizing the operation and
control of buildings
Office
BuildingHVAC37
28
Office Equipments
22
Elevator
3Other10
Lights
Hotel
44
254
9
18
Lights
HVAC
Office Equipments
Elevator
Other
7
System Architecture
Information fusion
Data driven modeling + prediction
Integrated building control
for energy saving
8
Control and optimization of building energy system
Energy supply in building
Distribution Battery
CHP
Wind
E-car
Fuel cellSolar
Lighting HVAC
Shading Window
Controllable devices
Elect
Heat
ComfortTemp HumidIllum CO2
Occupant demand
Minimization of energy cost
Micro-grid
9
List of possible challenges
10
bull Integrated control under full information may suffer the curse of dimensionality problem and time consuming evaluation of performance or constraints
bull Mache learning in general is a hard problem design of a good ML algorithm also include many decision variables (model structure parameters implementation input data hellip)
Ways to address the challenges
11
According to NFLT problem specific knowledge is needed to develop efficient solutions
bull Soft optimization for integrated control OO OCBA COO NP ADP EBO IPA hellip
bull Apply problem specific knowledge to reduce the search space for a good ML algorithm
Illustration of COO
G
S N
12
13
Below we will use individual thermal comfort model as an example of ML in smart building applications
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Energy consumption
Type faction
Building40
(68Electr )
Transportation 40
Others 20
Energy saving for buildings has been
omitted for long it has great potential 6
Energy consumption in buildings
It was estimated that 20 ~ 30
energy saving can be achieved by
optimizing the operation and
control of buildings
Office
BuildingHVAC37
28
Office Equipments
22
Elevator
3Other10
Lights
Hotel
44
254
9
18
Lights
HVAC
Office Equipments
Elevator
Other
7
System Architecture
Information fusion
Data driven modeling + prediction
Integrated building control
for energy saving
8
Control and optimization of building energy system
Energy supply in building
Distribution Battery
CHP
Wind
E-car
Fuel cellSolar
Lighting HVAC
Shading Window
Controllable devices
Elect
Heat
ComfortTemp HumidIllum CO2
Occupant demand
Minimization of energy cost
Micro-grid
9
List of possible challenges
10
bull Integrated control under full information may suffer the curse of dimensionality problem and time consuming evaluation of performance or constraints
bull Mache learning in general is a hard problem design of a good ML algorithm also include many decision variables (model structure parameters implementation input data hellip)
Ways to address the challenges
11
According to NFLT problem specific knowledge is needed to develop efficient solutions
bull Soft optimization for integrated control OO OCBA COO NP ADP EBO IPA hellip
bull Apply problem specific knowledge to reduce the search space for a good ML algorithm
Illustration of COO
G
S N
12
13
Below we will use individual thermal comfort model as an example of ML in smart building applications
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Energy consumption in buildings
It was estimated that 20 ~ 30
energy saving can be achieved by
optimizing the operation and
control of buildings
Office
BuildingHVAC37
28
Office Equipments
22
Elevator
3Other10
Lights
Hotel
44
254
9
18
Lights
HVAC
Office Equipments
Elevator
Other
7
System Architecture
Information fusion
Data driven modeling + prediction
Integrated building control
for energy saving
8
Control and optimization of building energy system
Energy supply in building
Distribution Battery
CHP
Wind
E-car
Fuel cellSolar
Lighting HVAC
Shading Window
Controllable devices
Elect
Heat
ComfortTemp HumidIllum CO2
Occupant demand
Minimization of energy cost
Micro-grid
9
List of possible challenges
10
bull Integrated control under full information may suffer the curse of dimensionality problem and time consuming evaluation of performance or constraints
bull Mache learning in general is a hard problem design of a good ML algorithm also include many decision variables (model structure parameters implementation input data hellip)
Ways to address the challenges
11
According to NFLT problem specific knowledge is needed to develop efficient solutions
bull Soft optimization for integrated control OO OCBA COO NP ADP EBO IPA hellip
bull Apply problem specific knowledge to reduce the search space for a good ML algorithm
Illustration of COO
G
S N
12
13
Below we will use individual thermal comfort model as an example of ML in smart building applications
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
System Architecture
Information fusion
Data driven modeling + prediction
Integrated building control
for energy saving
8
Control and optimization of building energy system
Energy supply in building
Distribution Battery
CHP
Wind
E-car
Fuel cellSolar
Lighting HVAC
Shading Window
Controllable devices
Elect
Heat
ComfortTemp HumidIllum CO2
Occupant demand
Minimization of energy cost
Micro-grid
9
List of possible challenges
10
bull Integrated control under full information may suffer the curse of dimensionality problem and time consuming evaluation of performance or constraints
bull Mache learning in general is a hard problem design of a good ML algorithm also include many decision variables (model structure parameters implementation input data hellip)
Ways to address the challenges
11
According to NFLT problem specific knowledge is needed to develop efficient solutions
bull Soft optimization for integrated control OO OCBA COO NP ADP EBO IPA hellip
bull Apply problem specific knowledge to reduce the search space for a good ML algorithm
Illustration of COO
G
S N
12
13
Below we will use individual thermal comfort model as an example of ML in smart building applications
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Control and optimization of building energy system
Energy supply in building
Distribution Battery
CHP
Wind
E-car
Fuel cellSolar
Lighting HVAC
Shading Window
Controllable devices
Elect
Heat
ComfortTemp HumidIllum CO2
Occupant demand
Minimization of energy cost
Micro-grid
9
List of possible challenges
10
bull Integrated control under full information may suffer the curse of dimensionality problem and time consuming evaluation of performance or constraints
bull Mache learning in general is a hard problem design of a good ML algorithm also include many decision variables (model structure parameters implementation input data hellip)
Ways to address the challenges
11
According to NFLT problem specific knowledge is needed to develop efficient solutions
bull Soft optimization for integrated control OO OCBA COO NP ADP EBO IPA hellip
bull Apply problem specific knowledge to reduce the search space for a good ML algorithm
Illustration of COO
G
S N
12
13
Below we will use individual thermal comfort model as an example of ML in smart building applications
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
List of possible challenges
10
bull Integrated control under full information may suffer the curse of dimensionality problem and time consuming evaluation of performance or constraints
bull Mache learning in general is a hard problem design of a good ML algorithm also include many decision variables (model structure parameters implementation input data hellip)
Ways to address the challenges
11
According to NFLT problem specific knowledge is needed to develop efficient solutions
bull Soft optimization for integrated control OO OCBA COO NP ADP EBO IPA hellip
bull Apply problem specific knowledge to reduce the search space for a good ML algorithm
Illustration of COO
G
S N
12
13
Below we will use individual thermal comfort model as an example of ML in smart building applications
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Ways to address the challenges
11
According to NFLT problem specific knowledge is needed to develop efficient solutions
bull Soft optimization for integrated control OO OCBA COO NP ADP EBO IPA hellip
bull Apply problem specific knowledge to reduce the search space for a good ML algorithm
Illustration of COO
G
S N
12
13
Below we will use individual thermal comfort model as an example of ML in smart building applications
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Illustration of COO
G
S N
12
13
Below we will use individual thermal comfort model as an example of ML in smart building applications
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
13
Below we will use individual thermal comfort model as an example of ML in smart building applications
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Motivations
bull HVAC system
ndash First invented to serve the machine manufacturing process etc --Set point oriented control
bull When HVAC serves peoplehellip
ndash Set point oriented control like what they did on the machine
20
22
24
26
28
30
32
2008
119
2008
124
2008
129
2008
23
2008
28
2008
213
2008
218
2008
223
2008
228
2008
34
2008
39
2008
314
2008
319
2008
324
2008
329
2008
43
Set p
oint (oC)
0
2
4
6
8
10
12
14
Energy co
nsu
mptio
n (kW
)S et pointE nergy consum ption
Tokyo Univ 2008 survey data
17
19
21
23
25
27Set point
Day 1
Day 2
Day 3
Day 4
Day 5
FIT Tsinghua Univ 2011 survey data
14
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Motivations(contrsquod)
ndash Intelligent thermostat (Perry D et al 2011)
bull User-oriented control system
ndash User only inputs sensations
ndash Personalized and self-learning
12
51
020
50
10
020
05
00
Task 1 Set to Heat
Thermostats
Tim
e (
s)
WEB TCH SMT BTN HYB
05
12
34
(min
ute
s)
Completed TasksIncomplete Tasks
Human perception
Indoor environment
Control
Perceive
15
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Existing modelsbull The chamber study model
ndash Predicted Mean Vote-Predicted Percent Dissatisfied (PMV-PPD) model
bull quantifies the thermal comfort concept as a mapping from the environmental factors and personal factors to a 7-level comfort value scale based on an average over a large data set
air temperature
radiant temperature
relative humidity
air velocity
clothing level
metabolic rate
activity level
Environmental factors Personal factors
PMV-PPD Model
Thermal sensation
cold cool slightly cool
neutral slightly warm
warm hot
PMV value -3 -2 -1 0 1 2 3
A 7-level thermal sensation scale
16
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Existing modelsbull The models based on the human body physiology
ndash The two-node (core and skin) model
ndash The multi-segment mathematical model of human body
ndash The sensation and comfort model for human segments and the whole-body
bull Field study comfort modelndash The original models were presented by Humphreys and Nicol which
described a strong relationship of the comfortable temperatures inside a building to the mean temperatures prevailing inside the building
ndash Classified by de Dear and Brager as physiological behavioral and psychological
ndash The ASHRAE adaptive model ASHRAE standard 55-2004
ndash SCATS European adaptive comfort standard EN15251
17
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Challengesbull The main challenges
ndash All these works focus on average thermal comfort models instead of personalized comfort models
ndash There exist less related literature and research on personalized comfort models
ndash The cases for the group are more complicated and challenging
18
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Terminal Control Strategies for Energy and Comfort
Adaptive HMIOccupants
Sensors
TempHumidityAir speedCO2
Acoustic levelIlluminance
Controller
T
R
H
CO2
Dynamic
Comfort
Region
InterpreterEstimated
comfort zone
1e
oe T
RH
Optimization
CO2
HotColdDryHumidNoisyhelliphellip
00
100
200
300
400
500
600
700
800
900
1 9 17 25 33 41 49 57 65 73 81 89 97 105 113 121 129 137 145 153 161
Ro
om
lo
ad W
m2
metered
simulatedEnergy metering
EnergyTemphelliphellip
bullPsychology
bullEngineering
bull industrial design
Human factors
Building manager
energy requirement
Lights Blind Window AC
Tsinghua-UTC Building Energy Energy Safety and Control System Research Center(CFINS DBS IE CPSR)
19
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Sensation votes based model
Voting software Sensors
Setup 1 Every one hour the software will pop up to let the user vote2 The sensor box will record the environment measurements store them in local
computer through COM and further upload to the server database
20
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
PDTC -- PMV framework
bull Heat balance equation of human
ndash Mapping from the environment to the human thermal vote
ndash Heat balance of human body
0M W C R E S
NeuralCoolCold Warm Hot
21
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
PDTC -- the proposed model
bull Personalized Dynamic Thermal Comfort(PDTC)
ndash Perception thermal vote
ndash Considering the dynamics of human thermal perception
0 1 2 3( ) ( ) ( ) ( ) ( )( )a aPDTC k m k m k P m k t m k R C
0 1 2 3( )a aPTV m m P m t m R C
22
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Parameter estimation
bull Parameter estimation ndash Least squares
bull Recursive least squares estimation with forgetting factorsndash Time-variant forgetting factors
0 1 2 3
0 1 2 3
1 20 1 2 3 0 1 2 3
1
1
1
ˆ ˆ ˆ ˆ arg min ( ) ( ( ) )
arg min ( ) ( )
NN k
km m m m k
N
m m
N
m m k
k
m m m m k PDTC m m m m r
k k
eal vote
( ) ( )( )PDTC kk X k
( ) ( ) ( )k PDTC k truevote k
( ) [1 ( )]a ak P t R C
1 if the k and k-1 are in the same day( )
otherwisek
23
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Results and validations
Office layout
Time From Nov 2009 ndash Jan 2010
12151217 1224 15 112 115-2
0
2
4
6
Subject A Recursive Results
m0
m1
m2
m3
12151217 1224 15 112 115-2
0
2
4
6
Month and Date
Para
mete
r V
alu
es
Subject B Recursive Results
24
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Results and validations
bull Model validation ndash accuracy
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject A
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject B
-20 -10 0 10 20-04
-02
0
02
04
Time offset
R
i
Subject C
-20 -10 0 10 20-04
-02
0
02
04
R
i
Subject D
1
2
3
4
Bias and MSE Correlation coefficient of residuals and inputs
SubjectPDTC
R-MSE
PDTC
R-Bias
PDTC
P-MSE
PDTC
P-BiasPMV
P-MSE
PMV
P-Bias
A 07230 0009 09703 007 24916 006
B 05319 -0015 05980 -0034 12999 0575
C 01442 -0058 01363 0026 05885 0058
D 05182 0064 05356 -005 04327 0272
E 07860 0064 09019 025 34994 -014
F 02860 0036 02684 00214 0713 -0047
G 03607 -0061 03634 01370 04633 -026
H 07167 -0087 08088 -0139 06777 0249
I 02371 -0025 02209 0023 0264 0932
25
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
A study case of applications
bull Personalized energy saving potentials
East Outside
External Wall
6m
6m
ow oWQ Q
Heat transfer of the
external wall and
window
iWQ
Heat transfer of the
interior walls
Sensible and latent heating
load for warming and
humidifying outside air
fa S fa LQ Q
Lamps heat emission
ltQ
Appliances heat emission
eqpQ
occQHuman body heat emission
0 1 2 3
( ) ( ) ( ) ( )
( ) (
(
)
)
a k a k
a k
t h all
k
a down up a down
a k
up
Min Q
s t m k m k P m k t m threshod
h h h t t
k R C
t
-10 -8 -6 -4 -2 0 2 40
1
2
3
4
5
6
7
8
Increase of heating load relative to PMV based results ()
Se
ns
itiv
ity
re
lati
ve
he
ati
ng
lo
ad
d
ec
rea
se
(
)
A
B
C
D
E
F
G
H
I
PMV sensitivity
Higher energy cost higher sentivity in comfor
and energy saving tradeoff
( ( ) ( )) ( ) 100PDTC a a PMV a a PMV a aR Q t h Q t h Q t h
| | 100PDTC threshold PDTC threshold PDTC thresholPDTC dS Q Q Q
26
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
bull Limitations of the previous work in real application
ndash Require the user to vote every one hour
ndash Nonlinear comfort constraint when online implemented
bull Can we be more user-friendly
27
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Complaint driven more user-friendly
bull Settingsndash Users only complain whenever they felt necessary
bull Advantagesndash Less demanding for users
ndash No interruption for users
ndash Close-loop control
Human Machine Interface
YJiang et al ldquoA Human Machine Interface for Building Indoor Environment Controlrdquo Chinese Patent ZL 2010 2 02929811
28
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Complaint driven more user-friendly
bull Challenges
ndash No intensity information in complaints binary variables
ndash No comfort samplesbull No-complaint periods have many possible explanations
ndash Few information of inner complaint region bull Environmental parameters are set around the comfort region boundary(Closed-
loop test-bed effects)
29
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Problem formulation
bull Problem formulation
ndash Only given the samples of target class ie a set of samples of a type of complaint 120594 = 1199091 1199092 hellip 119909119899 119909119894 isin1198772 ie in the temperature and relative humidity plane how to obtain a boundary description of the complaint region 119891(119908 119909) only based on the complaint samples 120594
Target class the cold or hot complaints which are from single subject
30
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Important properties
Properties of the complaint regionThe complaint region in the environment parameter space (in normal environment parameter range) for a given complaint is connected Additionally some of the parameters are unidirectional
1 Existing researches conclude both the human comfort zone and discomfort zone are connected areas
2 Unidirectional parameter in human perception generally exists Some of the parameters are not clear
eg temperature in hot and cold complaints is unidirectional relative humidity is not clear
31
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
bull Pareto-frontier set of the complaint samplesndash A sample 119909119894 isin 1198772is in the pareto-frontier set with respect to the
generalized inequality le119878 iff there is no sample 119909119895 119895 ne 119894 such that
119909119895 le119878 119909119894
ndash where 119878 is a proper cone and 119909119895 le119878 119909119894 means 119909119895 minus 119909119894 isin 119878
Temperature
Relative humidityComplaint samples
2 2 | (1 0) 0(2) Rx x xS
The cone (direction) of
Pareto frontier set in the
direction 2S
No samples in this region
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press 2004
Multi-linear one-class classifier model
32
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Multi-linear one-class classifier model
bull Multi-linear one-class classifier learning
ndash Least square linear estimation is performed for each of the pareto-frontier set 119881119896 and obtain a set of linear equations (classifiers)
bull Pareto-frontier set plays the role of support vector in support vector description method
bull Multi-linear approximation of the nonlinear boundary
bull The complain region can be described by
2min ( ) 12
k
j
w
x
T
k
V
w x c k
1kc
0 0 if ( ) 12
0 otherwise
T
T k i k i k
k k k
k
xw x c Vf x w x c k
33
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Multi-linear one-class classifier model
bull Performance metricsndash False Negative Rate (Missing detection rate) the rate of
complaints that were missed
ndash False Positive Rate (False detection rate) the rate of complaints that were mistaken as comfort
Empirical RuleIf the subject has not complained for 20 minutes and heshe will not complain for next 20minutes the current environment conditions are regarded as ldquocomfort samplesrdquo
1
1 comfort
iC y C
comf
N
o irt
FPR IN
1 The empirical rule is based on the results of transient thermal comfort research2 The higher FPR the more conservative of the classifier is
34
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Experiment settings
Experiment test-bed Touch screen Human Machine Interface
Dedicated HVAC and other terminals
Integrated sensors and computers
Closed-loop operation mode in test-bed
Sensors Radiant ceilingHuman Machine Interface
Zhuo Mao Fulin Wang Teng Gao Yunchuang Dai Qianchuan Zhao Yin Zhao Biao Sun Jing Guo and Fan Zhang Research of the room occupant complaining behavior pattern for the indoor environmental control Advanced Materials Research Vols 374-377 (2012) pp 1064-1067
35
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Results of the experiment data
24 26 28 30 3220
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject A
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 26 28 3010
20
30
40
50
60
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject C
23 24 25 26 27 28 2940
45
50
55
60
65
Temperature 0C
Rela
tive h
um
idity
Subject D
FPR =031FPR =08
FPR =077 FPR =065
Green polygon presents the parameter region of the experimentFPR is estimated as FPRC
36
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Results of the experiment data
1 Cold complaints usually occur in the lower temperature part and while hot complaints in the higher part
2 Data are collected in 3-4 continuous days during their experiments3 Ambiguous region which both hot and cold complaint had occurred exists
23 24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject B
24 25 26 27 2840
45
50
55
60
65
Temperature 0C
Re
lati
ve
hu
mid
ity
Subject D
FPR =024FPR =038
37
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Results of the experiment databull Comparison with the PMV model
1 Complaint-based comfort model may have a larger complaint area than the PMV model which indicates that indoor environment control based on PMV may cause complaints
2 Different regions in the learning results represent different perceptions
PMV numerical results in temperature and relative humidity plane The clothing index was chosen as 06 and air velocity was 0 which is accordance with our experiment conditions
-02
-02
0
0
002
02
02
04
04
04
06
06
06
08
08
08
1
1
1
12
12
Temperature 0C
Rela
tive h
um
idit
y
23 24 25 26 27 2840
45
50
55
60
65
70
75
80
22 23 24 25 26 27 28 29 3035
40
45
50
55
60
65
Temperature (oC)
Re
lati
ve
hu
mid
ity
(
)
Cold
Complaint
Region
1 Possible
Comfort
Region
3 Unexplored
Region
Hot Complaint
Region
2 Possible
Uncomfortable
region
Hot Complaints
Cold Complaints
38
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Performance analysis
bull Comparison with other models
1 Leave-one-out methods were utilized to evaluate the FNR for each methods2 Comfort samples were extracted from the experiment record according to the
empirical rule in previous slide3 SVM model using the linear kernel function
Subjects Fisher Linear
discriminant model
SVM model Proposed model
Hot Cold Hot Cold Hot Cold
A 04 -- 04 -- 008 --
B 017 0235 011 03 009 005
C 0253 -- 03 -- 008 --
D 054 038 045 041 007 006
E 047 0194 039 023 008 01
F 0307 058 029 038 007 001
False Negative Rate Comparison
Richard ODuda Peter EHart and David GStork Pattern Classification2nd edition John Wiley amp Sons Inc 2001
The proposed model has low false negative rate
39
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Experimental validation
40
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Experimental valuation
41
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Group thermal comfort modelbull The group comfort zone model
ndash We introduce here is a quite natural one take the convex hull of the individual comfort zones of the group
ndash Defining group comfort region as the intersection of all group memberrsquos individual comfort regions or the intersection of those of the majority when there are conflicts
42
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Experiment resultsbull The comparison with PMV
ndash Large group in Lanzhou Testbed
bull It is obvious that the individualdifferences in thermalpreference often incurdissatisfactions in the groupThis indicates that the averagemodel such as PMV may havebias in predicting the thermalcomfort for large group
Pareto frontier set(cold) Pareto frontier set (hot)
43
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Summary
44
bull Challenges
ndash Accurate occupant counting or localization problem
See T Labeodan W Zeiler G Boxem et al Occupancy measurement in
commercial office buildings for demand-driven control applications A survey and detection system evaluation Energy and Buildings 2015 93 303-314
ndash Data Mining for integrated building control and optimization
See F Xiao C Fan Data mining in building automation systems for improving
building operational performance Energy and Buildings 2014 75 109-118
F Cheng X Fu C Yan A framework for knowledge discovery in massive building automation data and its application in building diagnostics Automation in Construction 2015 50 81-90
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Links
45
bull IEEE RAS TC on Smart Buildingshttpwwwieee-rasorgsmart-building
Q Jia Q Zhao H Darabi et al Smart building technology IEEE Robotics amp Automation Magazine 2014 21(2) 18-20
bull IFAC TC on Smart Citieshttptcifac-controlorg93
bull Q Zhao Research opportunities arising from control and optimization of smart buildings Control Theory and Technology Vol 15 No 1 pp 78ndash80 February 2017
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
References
46
bull Jiang Y Wang FL Jiang ZY Hou Y Zhao QC Liu Y Zhang F Jiang Y Human-Computer Interface of Two-Way Interactive Architectural Environment Control System International Patent WO2012019328 Application No PCTCN2010001582
bull Zhao QC Zhao Y Wang FL Wang JL Jiang Y Zhang F ldquoA data-driven method to describe the personalized dynamic thermal comfort in ordinary office environment from model to applicationrdquo Building and Environment 72(309-318) 2014
bull Zhao QC Zhao Y Wang FL Jiang Y Jiang Y Zhang F ldquoPreliminary study of learning individual thermal complaint behavior using one-class classifier for indoor environment controlrdquo Building and Environment 72(201-211) 2014
bull Zhao QC Chen ZJ Wang FL Jiang Y Ding JL ldquoExperimental study of group thermal comfort modelrdquo 2014 IEEE International Conference on Automation Science and Engineering (CASE) pp1075-1078
bull Z Cheng Q Zhao F Wang Y Jiang L Xia and J Ding ldquoSatisfaction based Q-learning for integrated lighting and blind controlrdquo Energy and Buildings vol 127 pp 43ndash55 2016
bull F Wang Z Chen Q Feng Q Zhao Z Cheng Z Guo Z Zhong ldquoExperimental comparison between set-point based and satisfaction based indoor thermal environment controlrdquo Energy and Buildings vol 128 pp 686ndash696 2016
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Thanks Prof Ho for your inspiring guidance over the years
47
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Multi-linear one-class classifier model
bull Determine pareto-frontier sets of samples
2 2
(1) (2)| 0 0S x R x x Example
By incorporating the prior knowledge the pareto-frontier set of a certain class of samples represent boundary profiles in specified direction which we interested most
Stephen Boyd Lieven Vandenberghe Convex Optimization Cambridge University Press2004
A sample is in the pareto-frontier set with respect to generalized inequality iff there does not exist any other sample such that
where S is a proper cone in Rm
m
ix R
S jx i j
Sj ix x
j ix Sx Generalized inequality means Sj ix x
4843
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Unbiased theoretically
bull Expression noise when survey or vote
Jaffe-katz and Budescu 1989
1 2 3 4 5 6 7 8 9 10 11 12 13 14
IMPOSSIBLE 93 5 3
IMPROBABLE 5 60 33 25
UNLIKELY 25 30 65 25
POSSIBLE 5 68 18 10
LIKELY 18 50 33
PROBABLE 10 33 58
CERTAIN 100
IMPOSSIBLE 85 13 25
5 13 43 40 5
IMPROBABLE 25 28 20 40 8 25
20 15 10 65 10
UNLIKELY 15 23 38 15 75 25
35 3 5 70 18 5
POSSIBLE 3 3 25 8 75 25 25 10 5 10 25
50 25 45 35 10 5 25
PROBABLE 13 15 23 28 15
LIKELY 25 20 28 25 25
65 8 33 40 18 25
80 18 775 5
95 85 15
CERTAIN 25 25 10 85
Percentage of Rankings Received by Any Term Across Subjects
WW Ranks
WN Ranks
4943
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
An intuitive illustration
0-3 3
-05 550
True vote
NeuralCoolCold Warm Hot
Noise distribution
Noise distribution
5043
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Problems in the identification framework
bull Output-dependent observation noise
bull Observation noise is dependent on the system output
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
5143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Problem in the identification framework
bull Unbiased estimation of the system parameters
ndash Inconsistency of the noise at different outputs
ndash Output-dependent mean value (cannot remove the noise by average)
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
ˆE 5243
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Proposed identification methods
bull Key ideas
ndash First identify the noiseless output 119910(119906119894) using the noise model
bull Decouple the relationship between the parameters and the noise
ndash Then identify the system parameters 120579 bull Return to the normal system identification
Unknown SystemsInputs
Identification
yyu
ˆ
Observation Noise
( )iy u
Estimate the noiseless outputˆE
119910(119906119894) estimation of 119910 119906119894 120579 estimation of 120579
5343
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Noise model
bull Output-dependent bounded noise modelndash The noise is bounded and its bound is related to the noiseless output
ndash The probability density function has peak value at 0
bull Truncated distributions -- examples for different outputs in a bounded range
Truncated Normal Distribution (TN) Truncated Double Exponential Distribution (TDE)
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
08
w
No
ise
dis
rib
uti
on
de
ns
itit
y
TDE(a=-3b=3 = 1y = -25)
TDE(a=-3b=3 = 1y = -15)
TDE(a=-3b=3 = 1y = 0)
With noise parameter 120582
-3 -2 -1 0 1 2 3 4 5 60
01
02
03
04
05
06
07
w
No
ise p
rob
ab
ilit
y d
en
sit
ity
TN(a=-3b=3 = 1 y = -25)
TN(a=-3b=3 = 1 y = -15)
TN(a=-3b=3 = 1 y = 0)
With noise parameter 120590
5443
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Proposed identification methods
bull When the noise parameter (120575) is known
ndash Choose the input as
ndash Construct the following identification equation
bull This is the function of when the noise parameter is known
bull An explicit for of the equation for example TN model is
1 2 012k I i i I d k Ku
0 1 I i I i k I i iu u u u
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )iy u
0
( ) ( )( ) ( )
1ˆ( ) ( ) 1
( ) ( ) 1( ) ( )
i iK
i k I ii i k
a y u b y u
y u y u i Ib y u a y u K
Where 120593Φ are the pdf and cdf of standard normal distribution
5543
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Proposed identification methods
bull When the noise parameter (120575) is known (contrsquod)
ndash If the identification equation has unique solution
ndash The identification can be done by solving the following noiseless identification
bull Where and
0
1( ) ( ( ( ) )) 1ˆ 2( )
1
K
i
k
i k I iu u yy E w y i IK
u
( )i Ky u
T
KY
2[ ( ) ( ) ( )]T
i Iu u u 1 2[ ( ) ( ) ( ) ]T
K K K I KY y u y u y u
Note
1 The solution of identification is related to the number of repeated input
2 The inputs should satisfy the Persistent Exciting Condition
3 We name the identification method as Basic Identification Algorithm (BIA)
K 1 2iu i I
5643
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Proposed identification methods
bull When the noise parameter (120575) is unknownndash Underdetermined problem 119868 identification equations with 119868 + 1
unknown variables
bull Introduce an additional criterionndash Maximum likelihood under the constraint of identification equations
( ) 12 iy u i I
max log ( | )L D
0
1ˆ( ) ( ( ( ))) ( ) 1
1
( ) ( ) 1
K
i i k I i
k
T
i i
y u E w y u y u i IK
y u u i I
Note
1 When the system is identifiable then given 120575 there is unique 120579 and 119910 119906119894 2 The unknown parameter is usually a scalar and the optimization is converted to the
one-dimension search problem where each search step involves a procedure of
identification when the noise parameter is known3 We name the algorithm as Joint Identification Algorithm (JIA)
st
5743
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Algorithms
Theorem 1Under the condition that the identification equation has unique solution then the proposed algorithm can obtain the unbiased estimate of the unknown system parameter when K
5843
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Algorithms(contrsquod)
Theorem 2When the identification equations have unique solution for different the
results of Joint Identification converge to the true system parameter 120579 and noise parameter 120575 with in probability when K
5943
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Numerical test and application
6043
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143
Numerical test and application
bull Application in PDTC model
Y Zhao and Q Zhao ldquoSystem Identification for Output-dependent Bounded Noises and its Application in Learning Personalized Thermal Comfort Modelrdquo To appear in IEEE Proceedings of International Conference on Robotics and Automation Karlsruhe Germany 2013
6143