1

On Energy Saving of Subway HVAC System:Investigation and Autonomous Control

Abstract—As the backbone for urban public transportation,subways are also major consumers of energy and more than30% of the total energy is used to operate the heating, ventilatingand air conditioning (HVAC) subsystems. If it were possible toreduce energy consumption of HVAC subsystems a few percent,impressive quantity of electricity would be saved. From 2012to 2013, we conducted field studies and developed autonomouscontrol system for saving energy of HVAC systems in Beijingsubway stations. The energy consumption features and the loadsignatures of the HVAC systems were investigated and we haddeployed comprehensive environment monitoring, passenger flowmonitoring and run-time data logging subsystems to monitor andinvestigate above features in several metro stations. The extractedfeatures showed a broad space for optimizing the operation ofcurrent HVAC systems for energy saving. Based on the insightslearned from the field studies, we spent four months to developand deploy an autonomous HVAC control systems in three metrostations. The system design and deployment were reported. Upto now, the developed autonomous control systems have workedwell in the summer season of 2013. The energy logs show thatthe autonomous control system helped the metro stations reduceenergy in a range from 20% to 38% than the conventional controlstrategy. We also introduce key insights learned for energy savingand some future research directions.

Index Terms—Autonomous control, Sensor network, Air con-ditioning, HVAC, Energy Efficiency, Sustainability, Power meter

I. INTRODUCTION

The Beijing Subway is a backbone transportation networkthat serves the urban and suburban districts of Beijing mu-nicipality. It has 17 lines, 227 stations and 456 km of trackin operation [2], running by two companies, Beijing MTR(operates 14 lines) and Hong Kong MTR (operates 3 lines). Itranks the third in length in the world after Seoul and Shanghai,and serves more than 7.5 millions passengers a day, which isthe busiest in China. It served totally more than 2.46 billionpassengers in 2012 .

Because of its large scale and heavy load, Beijing subwayconsumes a great amount of energy everyday. A survey onthe energy consumptions of Beijing subway was conducted in2008 [13]. At that time, Beijing subway had only five lineswhose daily energy consumptions in April 2008 and in July2008 are summarized and compared. It showed that the dailyenergy consumptions of the subway lines ranged from 0.6million to near 5 million kW·h per day. There was a steepmore than 30%increment of energy consumption from Aprilto July in each line, indicating the potential energy consumedby air conditioning subsystems for cooling and ventilation inthe summer season.

In the last year (2012), we took a more detailed study to theenergy consumptions of Beijing Subway to examine theelec-tricity consumptions in a finer granularity. We disaggregate

the overall consumptions into consumptions of train-draggingsubsystem, lighting subsystem, air conditioning subsystem andother subsystems. The investigation was conducted in line Aoperated by Hong Kong MTR. For the sake of informationprivacy, the names of the line and the names of the stationswill be set anonymous in this paper. Our investigation resultsshowed that the consumptions of air conditioning subsystemsranged from 31% to 40% in the investigated subway stations.In particular, the air conditioning subsystem of the line con-sumed more than 18,190,000 Kwh electricity per month in thesummer season. For energy saving purpose, if it were possibleto reduce the energy consumption of the air conditioningsubsystems a few percents, impressive quantity of electricitywould be saved.

Particularly in three stations (S1, S2, S3) in line A, weconducted further research by deploying environment moni-toring sensors, smart meters and temperature and state sensorsinto the HVAC. The passenger flow is recorded by the ticketchecking systems. Based on these data, we investigate therun-time features, load and supply signatures of the HVACsubsystems. The study showed that: there were many operatingproblems in current air conditioning systems in the subwaystations, including: 1) the mismatching of loads and supplies;2) Control is not adaptive to the variation of environmentsand passenger flows; 3) control problems regarding the pumpsand valves etc. To address these problem, we developed anautonomous control system to save energy for HVAC systems.

From April 2013 to July 2013, we spent four months todesign, construct and deploy the autonomous HVAC controlsystems for the three metro stations based on the insightlearned from the field studies. The design emphasized adap-tivity of the HVAC system to match its cooling supply to thevariation of the heating loads. For practical considerations ofsystem reliability and the information limitation, we designedrule-based adaptive control strategies and implemented thesestrategies by deploying control units, frequency adapters, andautomatic valves etc. The systems have worked well throughthe summer season of 2013. Up to now, by the energy logscollected in the last summer, the results showed that theautonomous control system not only provided satisfactoryindoor temperature and humidity, but also reduced the energyconsumptions of the three stations in the range from 20% to38%. Some further research directions and insights learnedfrom current deployments are discussed.

The remained sections are organized as following. Relatedworks are introduced in Section II. Field studies and investi-gation to the features of subway HVAC systems are presentedin Section III. Load signature investigation is presented inSection IV. Development and deployment of autonomousHVAC systems are presented in Section V and VI respectively.

2

The energy saving performances are introduced in Section VII.The paper is concluded with discussions in Section VIII.

II. RELATED WORK

The autonomous, optimal control for HVAC systems hasattracted great research attentions in the studies of smart andsustainable buildings[12], [17], [6], [18], [10], [22], [14]. Theoptimal control for HVAC systems in smart building is todetermine the optimal solutions (operation mode and setpoints)that minimize overall energy consumption or operating costwhile still maintaining the satisfied indoor thermal comfortand healthy environment[19]. This goal is the same in thesubway HVAC control systems. Because the HVAC systemscontain different types of subsystems, such as gas-side andwater-side subsystems, the optimal control problems of HVACare extremely difficult. One of the difficulties is the lack ofan exact model to describe the internal relationships amongdifferent components. A dynamic model of an HVAC sys-tem for control analysis was presented in [17]. The authorsproposed to use Ziegler-Nichols rule to tune the parame-ters to optimize PID controlle. A metaheuristic simulationEP(evolutionary programming) coupling approach was developedin [6], which proposed evolutionary programming to handlethe discrete, non-linear and highly constrained optimizationproblems. Multi agent-based simulation models were studiedin [3] to investigate the performance of HVAC system whenoccupants are participating. In [22], swarm intelligence wasutilized to determine the control policy of each equipment inthe HVAC system.

One of the most closely related work is the SEAM4US(Sustainable Energy mAnageMent for Underground Stations)project established in 2011 in Europe[1]. It studies the metrostation energy saving mainly from the modeling and control-ling aspect. Multi-agent and hybrid models were proposed tomodel the complex interactions of energy consumption in theunderground subways[16], [15]. Researchers in this projecthave also proposed adaptive and predictive control schemesfor controlling ventilation subsystems to save energy [8]. Butthis project lacks the practical implementation.

!

1 2 3 4 5 6 7 8 9 10 11 12 13 14 Index  of  stations  in  Line  A

Annually  disaggregated  consumption  (kW

h)  

dragging lighting

HVAC Others

Fig. 1. Disaggregated annual energy consumption of subway stations in LineA (kWh)

Another most related work reported the factors affecting therange of heat transfer in subways [11]. They show by numer-ical analysis that how the heat is transferred in tunnels andstations. Reference [4] studied the environmental charactersin the subway metro stations in Cairo, Egypt, which showed

the different environment characters in the tunnel and on thesurface.

Further than these existing work, we conducted a morethorough and in deep studies on the HVAC systems of metrostations. Not only conducting a comprehensive field study toinvestigate the distinct features of subway energy consump-tion, but also are autonomous control systems developed anddeployed for the HVAC system in metro stations.

III. FIELD STUDY AND INSIGHTS FOR ENERGY SAVING

We firstly report our field study results and the insightslearned for energy saving. In 2012, we collected the electricitybilling logs and the logs of AC meters from line A toinvestigate the energy consumption characters of the subway.

A. Disaggregation of Annual Energy Consumption

To better understand where has the energy gone in thesubway, we conducted a disaggregation approach to the energyconsumption, i.e, to divide the overall energy consumption intothe consumptions of the dragging system, HVAC, Lighting,and others. Since the investigated line is a newly establishedline, it provides detailed energy consumption logs of eachsubsystem in the subway, even the detailed consumptions ofthe lights in a region. Based on the data logs, the disaggregatedannual energy consumptions for the stations of the line areshown in Fig.1.

We evaluated the portions taken by the HVAC system, whichwere shown ranging from 31% to 40% in different stations.Note that this portion is the portion over a year. Since theHVAC systems work only for 4 months per year in Beijing,i.e, from June to September, the energy consumption of theHVAC systems per day in the summer season are striking.To further understand the energy consumption of HVAC, wetake a closer look at the energy consumptions of the HVACsubsystem.

B. Look Closely at the HVAC system

In the subway, the HVAC subsystem is comprised by a setof refrigerators, ventilators, pumps, cooling tower, terminalfans, and the pipes connecting these components. The typicalstructure of a central HVAC system is shown in Fig.2. Theworking routine is as following:

1) Working Routine of HVAC: The refrigerator is the coreof the system, which cools water to about 5-7 degree. Thewater will be carried out by the chilling pumps to cool thesupply air. The cooled supply air will be blowed by the supplyfans and the terminal fans into the Fan Coil Units to coolthe indoor air. After cooling the supply air, the temperatureof the chilling water increases to about 12-15 degree, whichwill be returned to the refrigerator. On the other hand thecooling pump carries cooling water at about 32 degree fromthe cooling tower to the refrigerator, and the refrigerator sendsback water at about 37 degree to the cooling tower, so thatthe intaken heat from the indoor air at the refrigerator issuccessfully exchanged from the refrigerator to the outside.Note that a new air ventilator is responsible to blow outdoor

3

Fan  Coil  Units  of  users,  can  be  a  lot

Cooling  tower

valve Cooling  pump

valve

Chilling  pump

valve

condenser

evaporator

ventilator

5-­‐7  oC

12-­‐15  oC

compressor

Refrigerator  

T

T+(4-­‐5)  oC

Fig. 2. Major energy consuming facilities in a HVAC system

new air into the subway station, which on one hand to maintainindoor air quality (such as reduce CO2 density) and on theother hand to cool the indoor air when the outdoor temperatureis lower than indoor temperature. Table I further illustrates thenumber of different facilities and their rated powers in a HVACsystem in a subway station. The HVAC system in differentstations has similar structure. The real energy consumptions ofHVAC can be much different from the rated power, so that wefurther investigate the disaggregated the energy consumptionsin a HVAC system.

TABLE IMAIN COMPONENTS IN A HVAC SYSTEM OF SUBWAY AND THEIR POWER

PARAMETERS

Device names Ratedpower(kW·h)

Numberofdevices

Refrigerator large refrigerator 212 2small refrigerator 68 1

Chilling water pump large pump 30 4small pump 11 2

Cooling water pump large pump 30 4small pump 11 2

Main ventilators new-air supply fan 90 2return fan 90 2

Terminal fans terminal 1-17 1.1-7.5 17

C. Disaggregate the Energy Consumption of HVAC

Energy disaggregation investigation for HVAC was con-ducted in three stations, i.e, S1, S2, S3 in Line A, in which, theenergy consumptions of HVAC were further disaggregated intothe consumptions of refrigerates, ventilators, pumps, coolingtowers and terminal fans. Fig.3 shows the average disaggre-gated daily consumptions of different facilities in the HVACsystems of seven investigated stations in August of 2012.Fig.4 summarizes the average energy consumptions of thedifferent facilities. The results indicated that the refrigeratorstook the greatest potion, which was 50%; the pumps took29%. Since the pumps worked only if the refrigerators wereworking, so that the consumptions of the cooling part, i.e.,the consumptions of the refrigerators, cooling pumps, chillingpumps, and cooling tower took more than 80% in the overallconsumption. The investigation motivated us to pay moreattention to reduce the consumption of the cooling part.

D. Monitor Environment and Working States of HVAC

1) Load and Supply: In addition to the power disaggrega-tion approach, we conducted deliberately sensor deployment

Disaggregated  energy  consumption  of  HVAC  systems  in  the  source  part  of  the  line

refrigerator Chilling  pump Cooling  pump Cooling  tower Ventilators Terminal  fans

Power  consum

ption  (kWh)

1 2 3 4 5 6 7

Fig. 3. Disaggregated daily energy consumptions of the main facilities inHVAC subsystems in seven stations

Fig. 4. Energy partition of HVAC subsystem in summer seasons.

and data analysis to uncover the impacts of the outdoor envi-ronments and the passenger flows to the indoor temperature.There are mainly two ways for heat to be transferred into thesubway station: 1) the heat transferred from outdoor and 2) theheat brought in by the passengers. We call the sum of themthe heating load of the HVAC system. To keep the indoortemperature at a desired setting point, the HVAC system mustsupply appropriate cooling capacity, called cooling supply toresponse to the heating load. Therefore, the signatures ofthe heating loads is of great importance for designing energysaving control policies.

TABLE IINOTATIONS DEFINED FOR THE LOAD AND SUPPLY MODELS

Notations DefinitionsL(t) the quantity of thermal imported from outside to inside at t.T (t) the indoor temperature at t.To(t) the outdoor temperature at tReq heat transferring resistance from outside to inside.Mair the volume of outdoor air input into the subway stationc the heat capacity of per cube air.Tp the body temperature of people.n(t) the the number of passengers at time t.Mmix volume of mixed airMnew volume of new airMac volume of cooling airα the proportion of new air in the mixed air.Tac temperature of cooling air at the outlet of refrigerator.Tmix temperature of the mixed air.eac efficiency of of the cooling air transportation.Mz the volume of air inside the subway station.

2) Sensor Deployments: In S1, S2 and S3, we deployed dif-ferent kinds of sensors and smart meters to he indoor, outdoortemperatures, passenger flows and power consumptions of theHVAC systems in real-time. We give notations in Table II tolist the parameters to characterize the heating load. Some ofthe parameters are directly monitored and some of them willbe inferred by our proposed load signature model.

We installed temperature sensors at four points inside thesubway station and two points outside the subway to monitorthe indoor and outdoor temperatures T (t) and To(t) respec-tively. Note that T (t) is calculated by averaging the readingsof indoor temperature sensors, so as To(t). CO2 sensors are

4

installed inside the subway to measure the indoor air quality.The passenger flow is recorded by the ticket checking system,which is denoted by n(t). Note that n(t) is calculated by thesum of the checked-in and checked-out passengers from t− 1to t.

To monitor the working state of the HVAC system, tem-perature sensors were installed at the inlets and the outletsof the refrigerators to measure the temperature of the returnair T (t) and the cooling air Tac(t). Temperature sensors arealso installed at the new air pipes and mixed air pipes of theventilator to measure the temperatures of the new air To(t)and the mixed air Tmix(t). Note that the mixed air is themixing of return air and new air. The energy consumptionsof different components of the HVAC system, i.e, refrigerator,ventilator, water tower, pumps, fans etc are measured in real-time by the embedded power meters of the HVAC system. Weprocessed the collected data from sensors to investigate theworking signatures of the HVAC system.

3) Monitored Patterns of the Passenger Flow: In stationS1, from the data of ticket checking system, Fig. 5 shows thevariation of passenger flow as a function of time during a weekfrom Sep. 15 to Sep. 21. The passenger flow shows differentstructure in working days and weekends. In the working daysthere are two obvious peaks in the rush hours in the morningand in the evening. In week ends, the passenger flow wasalmost uniformly distributed from 8:00 AM to 8:00 PM.

0  

1000  

2000  

3000  

4000  

5000  

6000  

7000  

8000  

9000  

Numb

er o

f Pa

ssen

gers

时刻

海淀黄庄一周客流

Fig. 5. Pattern of passenger flow over a week.

Outdoor  tem

perature  o C

indoor  temperature  o C

CO2  Density

Fig. 6. How the indoor temperature was affected by the outdoor temperatureand passenger flow when the HVAC system was running.

4) Observed Patten of the Load: Fig. 7 shows how theindoor temperature was jointly affected by the outdoor tem-perature and the passenger flow when the HVAC systemwas running. It is on a sunny working day. Fig. 7(a) showsthe outdoor temperature in that day. Fig. 7(b) shows theCO2 density, which can be used to infer the passenger flowovertime. Fig. 7(c ) shows the variation process of the indoortemperature. We can see that: the indoor temperature variedbetween 22 centigrade and 27 centigrade during the day andthere are four peaks in the temperature curve, which can beexplained according to following reasons:

• The first peak at 4:00 AM is because the HVAC systemwas off in the night, so the indoor temperature increasesslowly.

• The second peak at 8:00 AM is due to quick heat inputby passengers in the rushing hours, which is higher thancooling capacity of the HVAC system.

• The third peak is at 2:00 PM due to the hot outdoortemperature. But this peak is not obvious, because whenthe outdoor temperature increases slowly, the HVAC hadenough time to cool down the indoor air.

• The last peak at 18:00 PM is due to the rushing hour inthe evening.

The measurements show intuitively the joint impacts of theenvironments and passengers to the indoor temperature. How-ever, we still don’t know the significance of the passenger-introduced load and the environment- introduced load.

IV. INVESTIGATE THE LOAD SIGNATURES OF HVAC

We investigated a quantitative model to more accuratelycharacterize the impacts of environment and the passengersto the thermal load.

A. Load and Supply Models

Definition 1 (load model): We define the quantity of heatimported from outdoor environments and the passengers intothe subway station in a time unit as the load of the HVACsystem in the subway station.

L (t) =To(t)−T (t)

Req+ n(t) (Tp−T (t)) + cMair (To(t)−T (t))

(1)The HVAC system responses the loads to control the indoor

temperature at desired temperature. By assuming the indoor airis fully mixed, the variation of indoor temperature is mainlycaused by the thermal difference of the load and the supply:

L(t)− S(t) = cMz∆(t) (2)

where Mz is the volume of air in the subway station, whichcan be calculated by the geometrical information of thestation, such as the length, width, height of the station andthe tunnels. ∆(t) = (T (t+ 1)− T (t)) is the temperaturedifference changed from time t to time t+ 1.

Definition 2 (supply model): The quantity of heat cooleddown by the HVAC system in a unit time is defined as

5

the supply of the HVAC system, which is defined based ondifferent working modes of the HVAC system:

S(t) = cMnew (T (t)− To(t)) , New air mode(Twin(t)− Twout(t))V wcoolβac, Refrigerator modecMnew (T (t)−To(t))+(Twin(t)−Twout(t))V wcoolβac, Mixed

(3)where Mnew is the volume of new air blowed into the subwaystation by the new air ventilator. Twin(t)−Twin(t) is the temper-ature difference of input and output water at the refrigerator;V wcool is the volume of the cooling water; βac = cwcooleac,where cwcool is the heat capacity of the cooling water and eacis the heat transportation efficiency of the refrigerator. So that(Twin(t)− Twout(t))V wcoolβac measures the cooling supply pro-vided by the refrigerator and cMnew(T (t)− To(t)) measuresthe cooling supply of the new air.

From the fan affinity laws[7], ventilators operates under apredictable law that the air volume delivered by a ventilatoris in the one-third order of its operating power: Mv = βvE

13v .

So that, the supply model of the HVAC system in the subwaystation is rewritten into:

S(t) =cE

13v βv (T (t)− To(t)) , New air mode

(Twin(t)− Twout(t))V wcoolβac, Refrigerator mode

cE13v βv (T (t)−To(t))+(Twin(t)−Twout(t))V wcoolβac, Mixed

(4)In the load and supply model, To(t), T (t), Twin(t), Twout(t), Evand V wcool are measured in real time by the deployed sensors.We have provided a public dataset of these sensors’ readingsin [20]. Since c is a known constant. Only Mnew, Mair, Req ,and βac are unknown.

B. Identify Load Signature by Linear Regression

1) Linear Regression Model: By substituting (1) and (4)into (2), we construct a linear regression model to estimatethe unknown parameters: n(t)(Tp − T (t))

To(t)− T (t)V wcool(T

win(t)− Twout(t))

T cpα−βac

= cMz∆(t) (5)

where α = cMair + 1Req

is the coefficients of To(t)−T (t) inthe load model, which is treated as one unknown coefficient.We can rewrite (5) as A(t)θ = B(t). Then by sensor measure-ments and HVAC states from 1 to t, we can set up an overdeter-mined observation matrix A1:t = [A(1),A(2), · · · ,A(t)]T ,and an observation vector B1:t = [B(1),B(2), · · · ,B(t)]T .Then the problem of identifying the load signature is toidentify the vector θ by solving A1:tθ = B1:t, with theconstraints that cp, α, βac are nonnegative.

Data collected from S1 from a timespan of Aug 21th, 2013to Aug 23th, 2013 was selected to solve the linear regressionmodel. The dataset provides real-time T (t), Tac(t), Twin(t),Twout(t), V wcool, and Ev in one-minute resolution. The passengerflow data is in per-hour resolution. We estimated the per-minute passenger amount by linear interpolations. Based on

these data, the observation matrix A1:t is constructed and thevector B1:t are constructed.

2) Solve the Linear Regression by a Search Algorithm:Since the coefficients are required to be nonnegative, directlyapplying the least square estimation is inefficient. We proposea search algorithm to solve this constrained optimizationproblem:

θ = arg min[cp,α,βac]

t∑i=1

|Ai,1cp + Ai,2α−Ai,3βac −Bi|

t∑i=1

(Ai,1cp + Ai,2α)

subject to: cp > 0, α > 0, βac ≥ 0

(6)

Ai,j is the item in ith column and jth row of the matrixA1:t. Note that we divide the difference of the accumulatedload vector and the accumulated supply vector by the accumu-lated load vector, which is to find the coefficient vector thatcan provide the minimum relative difference between the loadvector and the supply vector. The search algorithm searchesall combinations of [cp, α] for cp < 1000 and α < 10000.The parameter set [c∗p, α

∗,−β∗ac] which provides the overall

minimum relative error is chosen as the optimal solution ofproblem (6). For the number of coefficients is limited, thecomputing complexity of the algorithm is tolerable.

3) Load Signatures of Subway HVAC: For the particulardataset of August 23, 2013 of S1[20], we calculated theoptimal parameter set θ as [83, 53703,−1290071]T . Whenvarying the scope of the data, we found the solution varywithin tolerable range of errors. By substituting the calculatedcoefficients into the load model, the derived load signaturewas plotted in Fig.7a). The real-time supplies calculated by thesupply model are plotted in Fig.7b). We can see the supplyfollows closely to the load. The relative error between theintegrated load and integrated supply is plotted in Fig.7c),which it is relative small. It indicates that the searchingalgorithm has provided a rather confident estimation to theload signatures. From the load signature, we can see that:

1) The loads contributed by the outdoor temperature takethe major portion, more than the thermal loads intro-duced by the passengers.

2) The load signature of a day is predictable according tothe variations of outdoor temperature and traffic flowpattern.

3) The heating load increases quickly in the rushing hours,causing three peaks during a day, i.e., two peaks atrushing hours and one at the hottest time at noon.

C. Learned Insights for Energy Saving

From the energy disaggregation, the sensor monitoringresults, and the identified load signature of the HVAC system,we learned valuable insights and found potential points forenergy saving in the HVAC systems in subways.

1) The cooling part including refrigerators and pumpsdominate the energy consumption of the HVAC.

2) To avoid the mismatch of load and supply, because incurrent (time-table based) control strategy, the indoortemperature is sometimes lower than the desired value

6

0 200 400 600 800 1000 12000

1

2

3

4

5x 10

6

index of minute

quan

tity

of th

erm

al

environment introduced loadpassenger flow introduced loadtotal loads

0 200 400 600 800 1000 12000

1

2

3

4

5x 10

6

index of minutes

Qua

ntity

of t

herm

al

real−time supply by HVAC plus indoor thermal variation

0 100 200 300 400 500 600 700−2

0

2

4

6

8

10

12

14

16x 106

Index of time in minute for a period when HVAC in refrigerator mode

Qua

ntiti

ty o

f the

rmal

integrated loadintegrated supplyrelative error between integrated load and supply

Fig. 7. Derived load signatures Vs. Variation Signatures of Supply Vs. the relative error between integrated load and integrated supply

due to the higher supply than the load, and sometimehigher due to the quickly increasing of the load. Anenergy efficient control strategy must be adaptive to thereal-time load.

3) The conventional HVAC control strategy is not takingthe variations characters of the passenger flow intoconsideration.

4) The amount of fresh-air supply is not adapting to theindoor-outdoor temperature difference. In conventionalcontrol strategy, the fresh-air supply is adjusted onlyonce a season, whose control unit is not connected tothe automation system.

5) The pumps are over used. Each refrigerator is designedto work with one chilling pump and one cooling pump.But in the field study, we found that the number ofrunning pumps are general larger than the designednumber. The reason was found that the valves of therefrigerator pipes were not jointly controlled with therefrigerator.

V. AUTONOMOUS HVAC CONTROL SYSTEM

Based on the learned insights for energy saving and the loadsignatures of the HVAC systems, we designed and developedautonomous HVAC control system from January 2013 toMay 2013, in a pilot project supported by Hongkong MTR,particularly in three metro stations in line A. We introduce thesystem design by using S1 station as an example.

A. Design Principles

The principle of our control system design is that: 1) to usethe existing facilities in the old system as much as possible toreduce the updating cost; 2) the autonomous control systemshould coexist with the old control system and supports easyswitch for the purpose of system reliability; 3) We exploitthe idea of distributed control, centralized management, i.e.,the ventilators, pumps, and refrigerators are controlled dis-tributively for subsystem reliability. A central controller isresponsible to monitor overall system states and set controlparameters to the distributed controllers. 4) reliability: allthe control devices need to work well under the hostileenvironments in the subway station;

B. System Architecture

The architecture of our developed autonomous HVAC con-trol system is shown in Fig.8. The key parts are: central

!

Fan  valve  controller

Fan  Frequency  Adapter  &  controller

Fan  controller

Ventilator  controller Cooling  

tower  controller

Pump  controller

Data  collector

Central    collector

Remote  debugger

Energy  saving  module

BAS  station

BAS  system

BAS  controller

Ventilator  controller

Ventilator  valve  

controller

Cooling  tower  

controller

pumps  controller

Pumps  valve  controller

Smart  meter

Temp&humi

Press CO2

Fig. 8. System architecture of autonomous HVAC control systems.

controller, distributed controller, sensors, electric valves, andfrequency adapters to enable autonomous control of the HVACsystems.

i). Controllers: There are five distributed control cabinets(DCC): i.e., ventilator cabinet, cooling pump cabinet, freez-ing pump cabinet, refrigerator cabinet, and cooling towercabinet, and one central control cabinet(CCC). All DCCsare connected to the CCC via Profibus. Each DCC containsinformation collection unit, logic unit, and failure protectionunit, working individually in run-time to control its HVACfacility (such as ventilator). Such a distributed design is toavoid the system failures in case the communication failureshappen to improve the system reliability. The central controllermakes higher level decisions and assigns control parameters tothe DCCs. The CCC is connected to the building automationsystem (BAS), in which, a graphical user interface is providedto render the run-time states of the HVAC system to thesubway station managers.

ii). Sensors and Data collection Unit: As mentioned in Sec-tion III-D2, we deployed temperature, humidity, CO2 sensorsto monitor environment states. We also deployed temperaturesensors and smart meters to monitor the working states of theHVAC system. The passenger flow data is obtained from theticket checking system. The sensor data are reported to thecentral controller via profibus links.

iii). Electric valves and frequency adapters. We addedelectric valves to the pumps, so that the valves can be jointlycontrolled with the refrigerators and the ventilators. Frequency

7

TABLE IIICONTROL POLICIES FOR DIFFERENT COMPONENTS OF THE HVAC SYSTEM

Components AutonomousControl Function

Added sensorsand control unit

Control Policies

Refrigerator Autonomouslycontrolled stopand start

Indoortemperature,humidity,outdoortemperature,humidity

Start-stop Policy: Start a refrigerator when outdoor airs enthalpy value[5] > (52KJ/kg) +(3KJ/Kg); Stopall refrigerators when the outdoor airs enthalpy value< (52KJ/kg) -(3KJ/Kg);Add-decrease Policy: when cooling water temperature >7oC + 1oC and when the load of runningrefrigerators is higher than 95%, start an additional refrigerator. When the cooling water temperature<7oC and when the average loads of the refrigerators is lower than 50%, stop one refrigerator.

Valves ofpipes

Autonomous stopand start

Electrical valves Start-stop policy: before starting a refrigerator, open its valves of the cooling water pipes and chillingwater pipes; Stop the valves of the pipes after stopping the refrigerator.

Coolingpumps.(Policyof theChillingpumps isthe same)

Jointly controlledwith refrigeratorand frequencyadapting

Frequencyadapter, inletand outlettemperaturesensors

Frequency Policy: Frequencies of all running cooling pumps are set equal. When the frequency of therunning cooling pumps is higher than the setting frequency + 5Hz, start another cooling pump. Whenfrequency of the running pumps is lower than the setting frequency - 10Hz, stop a cooling pump.Jointly Control with Refrigerator: Start a cooling pump before starting a refrigerator, if the number ofrunning cooling pumps is not more than the number of running refrigerators. Turning off the coolingpumps by frequency policies after a refrigerator is turned off.

Coolingtowerventilators

Autonomous stopand start and gearadaptation

Cooling watertemperaturesensor

Gear level setting: 1 low→ 2 low → 3 low →4 low →1 high+3 low →2 high+2 low →3 high+1 low→4 highGear level adapting policy: when returned water temperature> max(32oC, wet-bulb temperature)[9]for 5 minutes, increase a level; when returned water temperature< max(20oC,wet-bulb temperature)[9]for 5 minutes, decrease a level.

Ventilators Autonomousstop and start,frequencyadapting, valuejointly control

Indoor CO2 tem-perature, humid-ity, outdoor tem-perature, humid-ity, CO2, sensors,electrical valves

Operating modes: 1) new air mode when refrigerators are off, the new air valve is fully open; 2) mixedmode, when both new air ventilators and refrigerators are used, the new air value is partially open; thesupply fan and exhausted fan are open; 3) winter ventilation mode: the new air value is partially open,the working frequency is low.Enthalpy value based running mode selection: 1) working in mode 1, when (39KJ/kg)+(3KJ/Kg)<outdoor airs enthalpy value<(52KJ/kg) -(3KJ/Kg); 2) turn to mode 2 when outdoor airsenthalpy value>(52KJ/kg) +(3KJ/Kg); 3) turn to mode 3 when outdoor airs enthalpy value<(39KJ/kg)- (3KJ/Kg);

adapters were added to the pumps, so that the pumps canadjust water flow by frequency change, which can smooththe switching of pumps when the working states of therefrigerators change.

C. Autonomous Control Policies for Energy Saving

Based on the hardware architecture, we designed au-tonomous control policies for controlling the HVAC system.For considerations of many practical issues, we designed anddeveloped rule-based, adaptive control policies. These rule-based policies for different HVAC subsystems are summarizedin Table III. Since the polices are depicted in detail in the table,we only explain some key points here:

1) The air enthalpy value [9] measured the amount of heatin a unit volume of air, whose formula is:

i = 1.01t+ (2500 + 1.84t) ∗ 0.001d (7)

where t = T (t) + 273.15 is the air temperature andd is the water content in 1kg air. The air enthalpywas calculated in real-time based on sensor readings toindicate the thermal condition of the air.

2) For appliances with frequency adapter, the running fre-quency is measured as an indicator of the loads to decidewhether to start an additional one or to stop a runningone.

3) The wet-bulb temperature [5] is the lowest temperaturethat can be reached under current ambient conditions bythe evaporation of water, which is the temperature feltwhen the skin is wet and exposed to moving air. It isdetermined by current temperature and humidity.

By using the air enthalpy, water temperature, running fre-quency of the devices as the indicators of the system loads, therule-based control policies have enabled an autonomous andadaptive control strategy for the HVAC system, which makesthe supply of the HVAC be adaptive to the load variations.Note that the policies in TableIII is the result after many onlinetest and evaluations. Although a predictive control strategy ispossible based on the predictable load signature of the HVACsystem, accurate load prediction is actually difficult becauseit needs much more information such as the weather, socialevents knowledges, and more powerful information fusionstrategies. At current stage, we present the rule-based, adaptivestrategy, which is reliable, needs much less information, andis practical.

VI. IMPLEMENTATION OF AUTONOMOUS CONTROLSYSTEM

From April 2013 to July 2013, we spent four months todevelop and implement the proposed the autonomous HVACcontrol systems in three subway stations in line A of Beijing.During the system development and implementation, we haveencountered and solved numerous practical problems. Sincethe subway system cannot be disturbed during the systemdevelopment, most of the development work were carriedout at mid-night when the trains were not running. We mustguarantee the original control system can work normally inevery morning after our development at the night, therefore,the autonomous control system was designed to coexist withthe original control system. The control strategy and can beeasily switched between each other.

8

Fig. 9. Central controller cabinet anddistributed control cabinets

Fig. 10. Positions and readings ofdeployed sensors in station S1

A. Hardware System

The HVAC appliances used in the subway stations are fromCarrier http://www.carrier.com.cn. We developed the centralcontroller and the distributed controllers as introduced inSection V-B and deployed Profibus DP network to connectthese control cabinets to the HVAC appliances. The figures ofthe central controller and the distributed controller are shownin Fig.9. These control units collect real-time working statesfrom the HVAC, make local decisions and deliver the controlcommands. All information are collected and transferred byProfibus DP network. We deployed sensors as introduced inSection V-B. For a specific station S1, the locations of thedeployed sensors and a snapshot of their readings are shownin Fig.10. Note that the result is from our software interface.

B. Software System

Based on the deployed sensors and the control units, wedeveloped and deployed software systems to monitor theworking states of the HVAC system and to render the systemstates to the users. The software was deployed as a plug-inof the Building Automation System (BAS), which runs onthe BAS server. It provides comprehensive, real-time environ-ment and working states information of the HVAC system,which not only provides real-time information for systemstate monitoring and performance evaluation, but also providesimportant evidence for verifying and online diagnosing thecontrol policies. Actually, the control policy in Table III wasthe result after many times of modifications according to thereal system performances.

Fig.10 shows a snapshot of the overview of the environmentstates rendered in the software interface. Fig.11 shows thewater system state of the HVAC system, including the real-time temperatures of cooling water, chilling water, the workingfrequencies of the pumps, the states of valves, the load of therefrigerators and the states of the cooling tower. Fig.12 showsthe wind system states in the HVAC system, including thetemperature of the supply air and the returned air, the runningstates (including the real-time frequency, power, current andstates etc.) of the supply fans, the exhaust fans and the newair fans. Alarming, parameter setting, energy saving report andcontrol policy adjustment are also supported in the softwareinterface.

VII. RUNNING RESULT AND ENERGY SAVINGPERFORMANCE

The developed autonomous HVAC control system hasworked well during the summer of 2013. The control perfor-

Fig. 11. water system state in the HVAC system

Fig. 12. wind system state in the HVAC system

mances and the energy saving performances were evaluated.

A. Control Performance

The goal of the HVAC is to keep the indoor temperatureand humidity varying within desired intervals during a day,called comfort intervals, to assure people feel comfort in thesubway station.

1) Comfort Interval: Because passengers stay only shorttime in the subway station, the comfort interval in the subwaystations is generally looser than that in home or in commercialbuildings. The interval between 25oC to 29oC is set asthe comfort temperature interval and 40% − 60% is set asthe comfort humidity interval. Fig.13 shows the temperaturevariations during a day in station S1. The four curves show thereadings of four temperature sensors in the subway station. Wecan see the indoor temperatures from 5:00AM and 11:00PMare well control in the comfort interval. Fig.14 shows thehumidity variations of the same day in the same station, inwhich the three curves show the readings of humidity sensors,which shows that the indoor humidity was also well controlledto be within the comfort interval.

2) Temperature peaks: The temperature variations in Fig.13show that when the HVAC system was stopped at nightthe indoor temperature increased quickly. There were slightlytemperature peaks at 8:00 AM because when the heating loadincreased quickly at the rushing hours, the indoor tempera-ture increased before the HVAC system could take sufficientresponses. Anyhow, the overall performances for temperature

9

Temperature  of  region  A

Temperature  of  region  A

Temperature  of  region  B

Temperature  of  region  B

Fig. 13. Temperature variations during a day on August 23, 2013

Humidity  of  region  A

Humidity  of  region  A

Humidity  of  region  B

Fig. 14. Humidity variations during a day in August 23, 2013

and humidity control are satisfactory.

22  

23  

24  

25  

26  

27  

28  

29  

30  

31  

32  

1   24  

47  

70  

93  

116  

139  

162  

185  

208  

231  

254  

277  

300  

323  

346  

369  

392  

415  

438  

461  

484  

507  

530  

553  

576  

599  

622  

645  

668  

691  

714  

737  

760  

783  

806  

829  

852  

875  

898  

921  

944  

967  

990  

1013  

1036  

1059  

1082  

1105  

1128  

1151  

1174  

1197  

The  indoor  temperature  VS.  The  temperature  of  cooling  air  

Indoor  Temperature   Temperature  of  cooling  air  

Fig. 15. How the indoor temperature response to the cooling supply of theHVAC system

3) Respond speed: We also evaluated how the indoortemperature was affected by the cooling supply of the HVACsystem. We measured the cooling supply of the HVAC systemby the temperature of the cooling air blowed by the coolingfans. Fig.15 shows the variation of the indoor temperature in asubway station over a day following the temperature variationsof the cooling air. We can see that when the temperature of thecooling air changed, the indoor temperature changed quickly,which shows that the indoor temperature has short respondingtime to the cooling supply of the HVAC.

4) Adaptivity of the Refrigerators: We further investigatedhow the working states and energy consumptions of therefrigerators change over a day with the variations of theoutdoor temperatures. Fig.16a) shows the temperature vari-ations and indoor-outdoor temperature differences over a day.Fig.16b) shows the concurrent working states and the energyconsumptions of the two refrigerators in that day. We cansee the working loads of the refrigerators are closely and fastresponding to the indoor-outdoor temperature differences.

B. Energy Saving Performance

We evaluated the energy saving performances of the au-tonomous control system by the energy logs in August 2013,which was the hottest time in Beijing. Since the autonomouscontrol system coexists with the conventional control system,

0 200 400 600 800 1000 1200−10

0

10

20

30

40

50

minitues in a day

tem

pera

ture

outdoor temperatureindoor temperaturetemperature difference

a) Indoor, outdoor temperature variations

0 200 400 600 800 1000 12000

200

400

current of refrigerator 1

0 200 400 600 800 1000 12000

200

400

current of refrigerator 2

0 200 400 600 800 1000 12000

500

current of refrigerator 1 + refrigerator 2

b) States and currents of the refrigerators

Fig. 16. Indoor outdoor temperature differences VS. the states and theconsumptions of the refrigerators.

we selected three days to run the conventional control strategy,and ran the autonomous control strategy in the other days. Theenergy consumptions of the HVAC systems in each day in thethree subway stations are listed in column 2 to 4 in TableIV. The three days in which we ran the conventional controlpolicy were highlighted.

1) Energy consumption is dependent on the outdoor tem-perature: At first, we can see that the energy consumptionsof the HVAC system in a day were highly dependent on theaverage outdoor temperature of the day. The fifth column ofthe table shows the average outdoor temperature from 8:00AMto 10:00PM in the day and the sixth column shows the averageindoor temperature of the same duration. We can easily seethat the daily energy consumption of the HVAC system ispositively correlated to the average outdoor temperature, whichcoincides our common sense.

TABLE IVENERGY CONSUMPTION COMPARISON

Time S1 S2 S3 To(t) T(t) Mode7 4644.294 5928.829 7923.496 28.38 29.19 auto8 6859.793 5850.69 10213.82 31.52 29.21 old9 5757.094 9897.5 13580.29 33.03 29.56 auto

10 6766.396 6777.309 6484 33.85 29.84 auto11 3797.602 7475.777 6772.504 31 29.6 auto12 4950.317 7083.711 10167 30 29.3 auto13 6673.889 10554.88 12782.09 31.24 27.99 old14 5603.398 6987.937 7742.109 31.52 28.15 auto15 4829.498 6504.684 8130.896 33.53 30.15 auto16 5482.404 5518.301 6252.008 30.96 30.32 auto17 7639.799 7715.895 10464.5 36.32 28.74 auto18 7569.996 7126.309 6636.703 33.07 28.74 auto19 7142.816 7265.707 7462.195 32.63 28.48 auto20 4869.703 5487.184 7927.684 29.99 28.79 auto21 5794.017 5608.193 6580.805 29.48 28.86 auto22 10313.89 7381.176 8992.219 29.45 27.59 old23 5058.072 4115.918 4803.492 32.07 28.45 auto24 5598.16 4966.538 4825.138 29.95 28.5 auto

2) Energy saving performance: Since the energy consump-tion of the HVAC system was positively correlated to the av-erage outdoor temperature, we select the days that had similaroutdoor temperatures to compare the energy consumptions ofthe conventional and autonomous running modes. So that theenergy consumptions of August 8 and August 13 of the oldcontrol strategy were compared with that of August 11, August14, and August 16 of the autonomous control strategy. Theenergy consumption of August 22 was compared with August20, 21, and 24.

The energy saving performance of the autonomous con-trol policy over the conventional control policy is shown inTable.V. From the results, it can be seen that, for example,

10

TABLE VENERGY SAVING PERFORMANCE OF AUTONOMOUS CONTROL

Station Names S1 S2 S3

average of Aug. 8 ,13 in old mode 6766.84 8202.7 11497.95average of Aug. 11, 14 in auto mode 4700.5 7231.8 7257.30

Saved Energy 2066.341 970.9 4240.64Energy saving ratio 0.305 0.118 0.368

average of Aug. 22 in old mode 10313.89 7381.17 8992.21average of Aug. 20,21,24, auto 5420.62 5353.97 6444.54

Saved Energy 4893.26 2027.20 2547.67Energy saving ratio 0.474 0.274 0.283

Average energy saving ratio 0.389 0.196 0.326

in S1, the autonomous control system can save more than2000Kwh power per day. The average energy saving ratio inS1 is over 38%, in S2 is over 19% and in S3 is over %32.

3) Economic benefit: Since the HVAC system works formore than 100 days in a summer, if 2000Kwh energy is savedper day, more than 200,000 Kwh power can be saved in asummer, in one subway station. Therefore, the autonomouscontrol system can not only reduce the energy consumption ofthe subway to make the subway be more environment friendly,but also can bring tangible economic benefit to the operatingcompany via reducing the operating cost.

VIII. CONCLUSION

We have presented the energy disaggregation approach, loadsignature identification and, an autonomous HVAC controlsystem for saving energy in HVAC systems of subway stations.The field analysis shows that HVAC consumes 31%-40%energy in the overall energy consumption of a subway station.Among the consumptions of the HVAC, the refrigerators andthe pumps take the greatest portion. Further, by investigatingthe loads of HVAC systems in the subways, we show the heat-ing load has predictable signatures, according to the variationof outdoor temperature and the passenger flows. The outdoortemperature’s impact generally is more significant than thepassenger flow. These investigations show the conventionaltime-table based control strategy lacks adaptivity and maytrigger load-supply mismatch. To tackles these problems, wepresented autonomous control system by deploying sensors,distributed control unites, electrical valves and frequencyadapters into the HVAC system. We carefully designed therule-based control policies by online tests and evaluations.The new control policies improve the adaptivity of the HVACsystem and as the result save energy for 20% to 38% indifferent stations.

In future work, if more information about the weather,the social events, and the system states can be availablein real-time, the prediction of the outdoor environment andthe passenger flow can be more accurate, which may enableprediction-based HVAC control strategy. Efficient informationcollection and more accurate system state estimation is also animportant research direction, which can provide fine-grainedinsights about the energy waste in HVAC. Cost efficient,effective, autonomous control devices and standards are stillneed to be explored.

ACKNOWLEDGMENT

Thanks for Dr. Xiaoliang Zhang, Dr. Ye Liu, and Mr. YeZhang for the work of energy diagnosis and energy savingpolicy investigation. Thanks for Xiangyu Xi for data pro-cessing. This work was supported in part by the NationalBasic Research Program of China Grant 2011CBA00300,2011CBA00301, the National Natural Science Foundation ofChina Grant 61202360, 61033001, 61361136003.

REFERENCES

[1] Sustainable energy management for underground stations, 2011.[2] Beijing subway, Oct. 2013. Page Version ID: 576457338.[3] C. Andrews, D. Yi, U. Krogmann, J. Senick, and R. Wener. Designing

buildings for real occupants: An agent-based approach. IEEE Transac-tions on Systems, Man and Cybernetics, Part A: Systems and Humans,41(6):1077–1091, 2011.

[4] A. H. A. Awad. Environmental study in subway metro stations in cairo,egypt. Journal of Occupational Health, 44(2):112–118, 2002.

[5] C.-M. Chu, T.-L. Jong, and Y.-W. Huang. A study of thermal comfortcontrol using least enthalpy estimator on HVAC system. In AmericanControl Conference, 2005. Proceedings of the 2005, pages 3665–3670vol. 5, 2005.

[6] K. Fong, V. Hanby, and T. Chow. HVAC system optimization for energymanagement by evolutionary programming. Energy and Buildings,38(3):220–231, Mar. 2006.

[7] R. W. Ford. Affinity laws. ASHRAE JOURNAL, 53(3):42–43, 2011.[8] A. Giretti, A. Carbonari, and M. Vaccarini. Energy saving through

adaptive control of ventilation systems. Gerontechnology, 11(2), June2012.

[9] M. F. Hordeski. HVAC Control in the New Millennium. CRC Press,June 2011.

[10] J. House and T. Smith. Optimal control of building and HVAC systems.In American Control Conference, Proceedings of the 1995, volume 6,pages 4326–4330 vol.6, 1995.

[11] Z. Hu, X. Li, X. Zhao, L. Xiao, and W. Wu. Numerical analysis offactors affecting the range of heat transfer in earth surrounding threesubways. Journal of China University of Mining and Technology,18(1):67–71, Mar. 2008.

[12] A. Kelman, Y. Ma, and F. Borrelli. Analysis of local optima in predictivecontrol for energy efficient buildings. In 2011 50th IEEE Conference onDecision and Control and European Control Conference (CDC-ECC),pages 5125–5130, 2011.

[13] M. Lu, T. He, X. Pei, and Z. Chen. Analysis of the electricityconsumption and the water consumption of beijing subway. Journalof Beijing JIaotong University, 35(1):136–139, Feb. 2011.

[14] D. Murayama, K. Mitsumoto, Y. Takagi, Y. Iino, and S. Yamamori.Smart grid ready BEMS adopting model-based HVAC control for energysaving. In Transmission and Distribution Conference and Exposition (TD), 2012 IEEE PES, pages 1–6, 2012.

[15] R. L. Roberta Ansuini. Hybrid modeling for energy saving in subwaystations. 2012.

[16] R. Serban, H. Guo, and A. Salden. Common hybrid agent platform –sustaining the collective. In 2012 13th ACIS International Conferenceon Software Engineering, Artificial Intelligence, Networking and ParallelDistributed Computing (SNPD), pages 420–427, 2012.

[17] B. Tashtoush, M. Molhim, and M. Al-Rousan. Dynamic model of anHVAC system for control analysis. Energy, 30(10):1729–1745, July2005.

[18] M. Vasak and A. Starcic. Model predictive control of a heating, ventila-tion and air conditioning system. In 2013 36th International Conventionon Information Communication Technology Electronics Microelectronics(MIPRO), pages 913–918, 2013.

[19] S. Wang and Z. Ma. Supervisory and optimal control of building HVACsystems: A review. HVAC&R Research, 14(1):3–32, 2008.

[20] Y. Wang, H. Feng, and X. Qi. SEED: public energy and envi-ronment dataset for optimizing HVAC operation in subway stations.arXiv:1312.2632 [cs], Dec. 2013.

[21] R. Yang and L. Wang. Optimal control strategy for HVAC systemin building energy management. In Transmission and DistributionConference and Exposition (T D), 2012 IEEE PES, pages 1–8, 2012.