Artificial Intelligence Applications in Power Systems
Om P. Malik
IEEE SA & NC Sections, May 7/8, 2018
What is Artificial Intelligence?
Artificial Intelligence (AI) has recently emerged as a science even though it may still be considered in its early stages of development. Depending on the goals and methods employed in research, its definition varies. As a broad description, it may be described as the science of making machines do things that would require intelligence if done by humans.
AI applications are now being considered in a very wide variety of disciplines, ranging from humanities to natural and applied sciences. In the context of power systems, application of artificial neural networks (ANNs) and fuzzy logic is commonly referred to in the literature as AI applications in power systems.
Over the past 25 years or so, feasibility of the application of AI for a variety of topics in power systems has been explored by a number of investigators. Topics explored vary from load forecast to real-time control and protection, and even maintenance.
Artificial Neural Networks
Natural Nerve Cell
Artificial Nerve Cell
Networks Based on Artificial Nerve Cell Model
- Multi-layer feed-forward
perceptron
- Recurrent
- Radial basis function
- Adaline
- Bayesian
- Hopfield
- Boltzman
- Kohonen
- Generalized Regression network
Types of Neuron Models
•Artificial Neuron Cell Model
•Multiplicative neuron
Reacts to product of activation of pairs of synapses
•Generalized neuron
Contains both summation and aggregation functions with sigmoid and Gaussian activation functions
Output Opk
s_bias
OutputInput
Aggregation
Thresholding
Function Function
Bias
Output OpkInputs,
Xi
Biass
1 f1
2
f2
Simple and Generalized Neuron Models
10
- ANN
(Conventional ANN)
Input
layer
-hidden
layer
-output
layer
P - ANN Input
layer
P-hidden
layerP -output
layer
-P -ANNInput
layer -hidden
layer
P -output
layer
P - -ANNInput
layerP-hidden
layer
-output
layer
DEFERENT ANN MODELS
Training of Neural Networks
Neural networks need to be trained. Based on the type of network, it may be:
• Supervised learning
•Unsupervised learning
•Competitive
Although most networks are trained off-line using available data, in some cases the weights can be up-dated on-line in real-time to track the system operating conditions.
Neural Network Controllers
Copying an existing controller with a network.Inverse plant modeling using a network.
Back propagating through a forward model of the plant.
Bayesian Networks
A Bayesian network (BN), also known as a Bayesian belief network, is a
graphical model for probabilistic relationships among a set of variables.
They have a qualitative component represented by the network
structure and a quantitative component represented by the assignment
of the conditional probability (CP) distributions to the nodes of the
network.
BNs can learn from observations. Learning of BNs can be parameter
learning and structure learning. With parameter learning, the structure
of the BN is given and only the CP parameters are learned. With
structure learning, the BN structure itself is learned. Bayesian learning
calculates the probability of each of the hypotheses given the data.
Insulation Deterioration Estimation of a Transformer
Using a Bayesian Network
Insulation LoL estimation by BN versus other methods for unit #64
Classical Direct Torque Control of an Induction Motor
Udc
abc to Switching Table
1
23
65
4
N
Sector calculator
Torque and Flux Estimation
+
+-
-
IP+
-
ANN Based DTC of an Induction Motor
Udc
abc to
1
23
65
4
N
Sector calculator
Torque and Flux Estimation
+
+-
-
SMC+
-
ANN
PI-DTC versus ANN-SMC-DTC
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 20
20
40
60
80
100
120
140
160
180
Time (sec)
Rot
or S
peed
(ra
d/se
c)
0 0.2 0.4 0.60
50
100
150
X: 0.3459
Y: 149.2
X: 0.1507
Y: 149.2
0 0.5 1 1.5 2150
152
154
156
158
160
DTC-ANN-SMC
Reference
DTC
Block Diagram of an Adaptive Controller
Controller Structure with MLFF NNs
Neuro-Adaptive PSS
Table 1:Dynamic Stability Margin* for Different Stabilizers.
* Dynamic Stability Margin is defined as the maximum power output at which the generator loses synchronism while input torque reference is gradually increased
Response to a three phase to ground fault, p=0.7 pu, pf=0.62
OPEN CPSS NAPSS
Maximum Power 2.65 pu 3.35 pu 3.60 pu
Maximum Rotor Angle 1.55 rad 2.14 rad 2.36 rad
ADALINE Network as an Identifier
Radial Basis Function Network
RBF-Identifier & Pole-Shifting Controller
Stability Margin Test
APSS CPSS APSS
Experimental Power System Model
Plant
GN
IdentifierUnit Delay
Σ
Disturbance
u-vector
w vector
w(t)+
-PLANT
wGN
Controller
Learning Algorithm GN-identifier
Output wi(t+T)
u_vector
ω_vector
Identifier Controller
28
Performance of GN identifier
Results of GN identification for a 3-Phase to Ground fault at generator bus for 100 ms at P=0.7, Q=0.3 (lag).
Experimental Results of GN identification under 23 % step change in torque reference and trained on-line.
29
Performance of GNPSS and GNAPSS under three phase to ground fault for 100ms at the middle of one line in a double
circuit system at P=0.7pu and Q=0.3 pu (lag) .
Performance of GNPSS and GN based adaptive PSS when one line is removed at 0.5 sec. and re-energized at 5.5 sec and
then again same line is removed at 10.5 sec. and re-energized at 15.5 sec. at P=0.8 pu and Q=0.4 pu (leading).
Fuzzy LogicGeneral Concept
Fuzzy Logic Membership Functions
18-May-8
Examples of Membership Functions distributions
-1 0 10
1
NB NM NS ZO PS PM PB
(a) Initial
-1 0 10
1
(b) Contracted
-1 0 10
1
(c) Expanded
-1.5 -1 -0.5 0 0.5 1 1.50
0.5
1
(a) =2.5
-1.5 -1 -0.5 0 0.5 1 1.50
0.5
1
(b) =1.0
-1.5 -1 -0.5 0 0.5 1 1.50
0.5
1
(c) =0.6
Fig. 2. Linear scaling
32/15
Fig. 3. Nonlinear scaling
.
Initial
Contracted
Expanded
Initial
Fuzzy Rules Table
We
NB NM Z PM PB
∆We
NB NB NB NB Z Z
NM NM NB NS Z Z
Z NS NS Z PS PS
PM Z PS PM PM PM
PB Z PM PB PB PB
Conventional and Fuzzy PID Algorithm
Fuzzy Logic Self-tuning PI Algorithm
Hybrid Micro-Grid Configuration
Synchronous
Generator
Diesel
Engine
Clutch
Diesel Generator
PDiesel
PHydro
PWind
PDumpLoad
PLoad
Pump-Turbine
Dynamics
Synchronous
Generator
Tunnel and
Penstock Water
Dynamics
Head
Flow
Hydro-Pumped Storage Facility
Induction
Generator
Wind Turbine
Dynamics
Wind Generator
Dump Load
ControllerResistor Bank
Dump Load
Consumer Load
Dump Load Frequency Control
Fuzzy IF-THENInference engine
Δf
dΔfdt
ΔP
23/42
DL frequency control
Previous fuzzy frequency control
[4]
Proposed fuzzy controller
1 Large membership functions reduce the regulation time
2 Small membership functions reduce oscillations around settling point24/42
HPS Turbine Mode Controller
Simulink Fuzzy PID Controller Model
Linear Fuzzy PID Controller Non-Linear Fuzzy PID Controller
HPS Pumping Mode Controller
Simulink Fuzzy Logic Controller Schematic
Fuzzy Logic Controller Control Surface
Generator Loads
Power system diagram including SVC
Device
SVC
Transmission Lines
Generator Loads
+
-
Δω / ΔPeSVC
Controller
Vref
Vm
u(t)
Transmission Lines
Classical control approach for
a FACTS device
Proposed Solution
Proposed adaptive control system structure for SVC device
In a SMIB system
Generator Loads
+
-
ΔPSVC
SVC
Adaptive Controller
System Identification
Vref
Vm
u(t)
Transmission Lines
Single Machine Infinite Bus System Simulation Results
μ (
e)
Example of membership function before and after adaptation
………. Before adaptation After adaptation
Consequent Parameters
Multi-machine System Simulation Results
G3
1
G2
1
G5
1
G1
1
G4
1Load 1
Load 2
Load 3
1 7 96 3
8
4
2
5
Schematic model of a multi-machine power system with an SVC
device installed at the middle of the tie-line
Sensorless Control of a Switched Reluctance Machine
Fuzzy Logic Controller for SRM
Speed Tracking SRM
0 5 10 15
0
50
100
150
200
250
300
350 a
Sp
eed
(ra
d/s
)
Time (s)
Wmes
Wref
0 5 10 15-10
-8
-6
-4
-2
0
x 10-6 b
S
peed
err
or
(rad
/s)
Time (s)
Estimated and Real Load Torque SRM
0 5 10 15-2
0
2
4
6
8
10
12 a
T
orq
ue (N
.m)
Time (s)
Tl
Tl-obs
0 5 10 15-0.05
0
0.05 b
To
rqu
e e
rro
r (
N.m
)
Time (s)
Permanent Magnet Synchronous Generator WECS
Field Oriented Control of Stator Side Converter of PMSG
•D-Q components of the stator reference voltages, that ultimately control the rectifier firing angle, are generated by two PI controllers with d-q components of the stator currents as inputs.
•Conventional PI controllers are replaced by trained ANFIS with d-q axes stator currents error and integral of error as inputs.
• Applied to a 1.5 MW wind turbine system with PMSG
Δώ
Δω
Sector B
Sector F
Sector E
Sector C
Min Axis
Sector D
R
θ
Six sector phase plane
O
Sector A
Fig. 6 Fuzzy sets for input variable
Two Area Power System for LFC
0 10 20 30 40 50-0.035
-0.03
-0.025
-0.02
-0.015
-0.01
-0.005
0
Time (sec)
df1
Both PI
PI in Area #1 and PFC in Area #2
Frequency variation of area- 1 in a Two Area Thermal System without Reheat unit when disturbance in both areas
Frequency variation of area- 2 in a Two Area Thermal System without Reheat unit when disturbance in area - 1
-
0 10 20 30 40 50-0.016
-0.014
-0.012
-0.01
-0.008
-0.006
-0.004
-0.002
0
Time (sec)
df2
PI
PID
Fuzzy
Polar Fuzzy
Short Term Load Forecast• Statistical methods• AI based methods employing both neural networks and fuzzy logic- Neural networks need to be trained- Using heuristic optimization techniques, e.g. GA, that employ
random search and fuzzy rules to guide search, performance can be improved.
- A generalized neural network (GNN) with four wavelet components of the historic load data as input and fuzzy logic guided random search GA as a learning tool for the GNN is used for short term load forecast.
- RMS error with: back propagation training – 0.0610 GAF training – 0.0486
Short Term Load Forecast with FL and GNN
0 20 40 60 80 100 1200
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Time(Hours)
Loa
d (
kW)
Actual and predicted training and testing for winter season using GNN
actual
predicted
0 20 40 60 80 100 1200
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Time(Hours)Lo
ad (
kW)
Actual and forecasted training (0:101) and testing (102:117) for the winter season using GNN-GAF model
actual
predicted
Self-Tuning Load Forecast using GNN-W-GAF
0 20 40 60 80 100 1200
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Time(Hours)
Load
(kW
)
Actual and predicted training and testing for the winter season using GNN-W-GA-F
actual
predicted
Supervisory Control of a Cogeneration Plant
Generator Fuzzy Set-Point Control
Fuzzy Logic Self-Tuning PI Controller
Fuzzy Adaptive Control PSS
RLS identifier and a self-learning Mamdani fuzzy logic controller.
^y(k+1)
y(t) Mamdani FLC
RLS identifier
z-1
+
_
AVR
&
Exciter
TL Gridu(t)
APSS
Results
0.1 p.u. step increase in torque and return to initial condition
(power 0.30 p.u., 0.9 pf lead)
3 phase to ground fault at the middle of one transmission line and successful re-closure
-adaptive Mamdani fuzzy logic PSS (AMFLPSS
----fixed centers FLPSS
(power 0.9 p.u., 0.9 pf lag)
Adaptive Neuro-Fuzzy Inference System
General Schematic of ANFIS
Basic structure of a typical ANFIS with two inputs and two-rule fuzzy system
Adaptive Neuro Fuzzy Inference System
• An ANFIS is an integration of neural networks and fuzzy inference systems to determine the parameters of the fuzzy system.
• Automatically realize the fuzzy system by using the neural network methods.
• Fuzzy Sugeno models are involved in the framework of adaptive system to facilitate learning and adaptation.
• Permit combination of numerical and linguistic data.
• Requires structural and parameter learning algorithms.
The Proposed Adaptive Neuro-Identifier
• A Multilayer Perceptron (MLP) network is constructed to represent the plant
Architecture of adaptive neuro-identifier
Adaptive Simplified Neuro-FuzzyController
Proposed control system structure
18-May-8
NFC architecture
NNN xxsignxf
73/15
Nonlinear Function (NLF):
Nx
18-May-8
Control system structure
74/15
Online Adaptive Neuro Fuzzy Controller for Nonlinear
Functions in the Input Layer for Damping Power
System Oscillations
A fuzzy PSS is usually made adaptive by adjustment of input membership functions (premise) and consequent parameters (CPs).
Number of controller parameters depend on the shape and number of membership functions.Scaling factors have received little attention in the adaptive fuzzy PSS design
18-May-8
System Configuration
.
w
77/15
18-May-8
Simulation results
78/15
Multi-machine power system.
Fig.14. 0.10 pu step inc-dec in torque of G3,
PSS on G3.
0 2 4 6 8 10 12 14 16 18 20-0.05
-0.04
-0.03
-0.02
-0.01
0
0.01
0.02
0.03
0.04
0.05
Time (s)
w1- w
2 (ra
d/s)
No PSS
CPSS
ANFPSS
0 2 4 6 8 10 12 14 16 18 20-0.25
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
0.2
0.25
Time (s)
w 2- w3 (
rad/s
)
No PSS
CPSS
ANFPSS
ω1-ω
2(r
ad/s
)ω
2-ω
3(r
ad/s
)
0 2 4 6 8 10 12 14 16 18 20-0.05
-0.04
-0.03
-0.02
-0.01
0
0.01
0.02
0.03
0.04
0.05
Time (s)
w1- w
2 (ra
d/s)
No PSS
CPSS
ANFPSS
0 2 4 6 8 10 12 14 16 18 20-0.25
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
0.2
0.25
Time (s)w
2- w3 (
rad/
s)
No PSS
CPSS
ANFPSS
Fig.15. 0.10 pu step inc-dec in torque of G3,
PSS on G1, G2 and G3.
1.5 MW VSWECS
Fig.2 Field oriented control scheme with speed sensor at generator
PI PI+_+_
PI+_
++
+_
SVMPWM
gen.
dq
dq
abc
PMSG
s
1P
min/
sec/
rot
rad --
isq
isd
0
i sd
i sq
v sd
v sq
isq
isdisa
isb
isc
vDC
v
w de
w qe
rn
*
rn
mechn
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
Time (Sec)
Genera
tor
Speed (
p.u
.)
Generator Speed with PI Controller
Generator Speed with ANFIS Controller
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
Time (Sec)A
ctiv
e P
ow
er (
p.u
.)
Active Power with PI Controller
Active Power with ANFIS Controller
•Applied to the 1.5 MW wind turbine system.
• The wind speed starts at 11m/s, is changed to 9 m/s after 12 s
Experimental Results of Applying the
ASNFC in a Real-Time System
Generator speed deviation in response to a 15% step increase
in the torque reference (P=0.80 p.u. and 0.75 p.f. lag)
200 km Transmission Lines
Generator speed deviation in response to a three-phase to ground
short circuit test at the middle of a 200 km transmission line with
an unsuccessful re-closure (P=0.97 p.u. and 0.93 p.f. lag)
Concluding Remarks
• A wide spectrum of AI applications in power systems, from load forecast to maintenance, is being explored.
• A general survey of the type of AI applications that have been and are being explored for application in power system has been attempted.
• This is not an exhaustive survey and some other applications are also being pursued.
• Actual application of AI techniques, particularly for real-time applications, is lagging. One application that seems to have been adopted by the utilities is neural network based load forecast algorithms.