Page 1
Wide Area Monitoring in Power Systems Using
Cellular Neural Networks
Bipul Luitel, Student Member, IEEE and Ganesh K. Venayagamoorthy, Senior Member, IEEE
Abstract—The demand of power and the size and complexity
of the power system is increasing. Wide area monitoring and
control is an integral part in transitioning from the traditional
power system to a Smart Grid. However, wide area monitoring
becomes challenging as the size of the electric power grid, and
consequently the number of components to be monitored, grows.
Wide area monitor (WAM) designed using feedforward and
feedback neural network architectures do not scale up to handle
the growing complexity of the Smart Grid. In this paper, cellular
neural network (CNN) is presented as a way to provide scalability
in the development of a WAM for Smart Grid. The CNN based
WAM is compared with multilayer perceptrons (MLP) based
WAM on two different power systems. The results show that the
CNN has better or comparable performance with, and scales up
much better than, MLP.
Index Terms—Backpropagation, CNN, Cellular Multilayer
Perceptron, MIMO, Power system, Wide Area Monitor
I. INTRODUCTION
Stability of electric power system depends on proper
functioning of various power system components. Power
system is a massively distributed network. Therefore, constant
remote monitoring is necessary to assess the current state of
these components. Based on this assessment, related control
action is taken on the power system components in order
to keep the system in stability. Wide area monitoring and
control system (WAMCS) has, therefore, become critical
for the power grid. However, with the addition of more
distributed resources and microgrids to a smart grid, the
number of variables to be monitored will increase and wide
area monitoring and control of such complex dynamic system
will become a challenge.
Applications of wide area monitoring systems (WAMS) in
power system for state estimation, disturbance identification
and wide area PSS have been reported in literature [1],
[2]. Various design aspects of WAMCS are studied in [3].
Unlike traditional methods of data acquisition and control,
typically supervisory control and data acquisition (SCADA)
methods that relied on remote terminal units (RTU) for
data, WAMS utilize phasor measurement units (PMUs) for
collecting data from the power system on a faster timescale
and hence can be used to monitor transient and dynamic
The authors are with the RealTime Power and Intelligent Systems Labora
tory, Missouri University of Science & Technology, Rolla, MO 65409 USA,
Contact: {iambipul, gkumar}@ieee.org
The funding provided by the National Science Foundation, USA under
the grants CAREER ECCS #0348221 and EFRI #0836017 is gratefully
acknowledged. The research was also partially supported by IEEE CIS Walter
Karplus graduate student research grant.
response of the system [1]. A substation based dynamic
state estimator has been used as WAMS in [4] that provides
abilities to predict instabilities before they occur. Although
these various techniques are being used and developed for
wide area monitoring, there are still major challenges in their
use for control. These challenges are related to extracting
dynamics of the system without knowing the system model,
mining and interpreting huge amount of data available from
monitoring devices and assessment of the overall dynamics
of the system based on wide area information [1]. It is even
bigger challenge to make reliable control decisions under
realtime constraints.
Computational intelligence (CI) techniques have shown
promises in the field of wide area monitoring and control
[5]. Since neural networks (NN) can be used to represent the
dynamics of the system by training on the historical data of
the system without having to know its actual model, they
have shown promises in predictive control applications. NNs
have been successfully implemented as state predictors and
neurocontrollers [6] in the areas of wide area monitoring and
control. Simultaneous recurrent neural network (SRN) and
echo state network (ESN) based wide area monitor (WAM)
has been demonstrated to be quite effective in performing
predictive neuroidentification of distributed power systems
for the purposes of accurate control [6], [7]. Radial basis
function networks have been used for wide area monitoring
with an adaptive critic designs based control in [8]. However,
these feedforward and feedback neural network architectures
do not scale up to handle the growing complexity of the
smart grid for wide area monitoring and control. As the
number of variables increases, the number of neurons in the
NN increases and so does the computational complexity.
Therefore, it becomes challenging for the NN training
algorithms to correctly learn the nonlinear system dynamics.
A cellular neural network (CNN) overcomes this problem
of scalability by dividing a huge network into subnetworks
among different cells where each cell consists of a neural
network that deals with fewer variables and hence fewer
neurons and lesser computational complexity. These cells are
interconnected in such a way that the connectivity and the dy
namics of the actual power system is preserved. In this paper,
WAM is developed using a multilayer perceptron (MLP) based
CNN, also known as a cellular MLP (CMLP). The design
is applied to two benchmark power systems for predicting
the speed deviations of the generators. The development and
Page 2
training mechanisms for CMLP are described and the results
are compared with a WAM developed using a multipleinputs
multipleoutputs (MIMO) MLP. The remaining sections of the
paper are arranged as follows: development of WAM using
CMLP is described in Section II. CMLP training approach is
described in Section III. Results and discussions are presented
in Section IV, and conclusions in Section V.
II. DEVELOPMENT OF A CNN BASED WAM
Two test systems are considered for this study. Test System I
is the 12bus benchmark system shown in Fig. 1 [9]. It consists
of three generators, one in each area. Test System II is the two
area fourmachine system shown in Fig. 2 [10]. It consists of
four generators, two in each area. The WAM is developed to
predict the speed deviations (∆ˆ ω) of each generator in the
system at time instant k + 1 based on speed deviations (∆ω)
and deviation of the reference voltage (∆Vref) (shown in Fig.
3) of the generators at time instant k as the inputs.
)(
2k VrefG
∆
)(
3kVrefG
∆
)(
2k
G
ω∆
)(
3k
G
ω∆
)(
4kVrefG
∆
)(
4k
G
ω∆
) 1
+
(
2
∆
∧
k
G
ω
) 1
+
(
3
∆
∧
k
G
ω
) 1
+
(
4
∆
∧
k
G
ω
)(
2kVrefG
∆
)(
2k
G
ω
∆
)(
4kVrefG
∆
)(
4k
G
ω∆
)(
3k
G
ω∆
)(
3kVrefG
∆
Fig. 1.CNN based WAM for Test System I (12bus system).
The WAM is implemented using a CMLP where each
cell of the network consists of an MLP and represents one
generator of the power system. The cells are interconnected
based on ‘nearestn neighbors’ topology, which means
previous sample outputs of n nearest neighbors of each cell
are connected to the inputs of that cell. The “nearness” is
defined as the electrical distance between the generators and
is measured based on the length of the transmission lines
??
???
???
??? ??
?
?
? ? ?
?
V
?? ?
?
? ? ?
?
?? ?
?
???
?
?
???
?
?
??
?
?
??
)(
1kVrefG
∆
)(
2kVrefG
∆
)(
4kVrefG
∆
)(
3kVrefG
∆
)(
1k
G
ω∆
)(
2k
G
ω∆
)(
4k
G
ω∆
)(
3k
G
ω∆
) 1
+
(
1
∆
∧
k
G
ω
) 1
+
(
2
∆
∧
k
G
ω
) 1
+
(
3
∆
∧
k
G
ω
) 1
+
(
4
∆
∧
k
G
ω
? ??
)(
1k
G
ω∆
)(
1k VrefG
∆
)(
2kVrefG
∆
)(
2k
G
ω∆
)(
4kVrefG
∆
)(
4k
G
ω∆
)(
3k
G
ω∆
)(
3kVrefG
∆
? ?? ? ??
? ??? ??
?
?
??
??
?
?
Fig. 2. CNN based WAM for Test System II(twoarea fourmachine system).
ref
V
ref
∆
Fig. 3.
∆Vref).
Generator excitation system (showing application of PRBS and
separating the two generators. In this study, two nearest
neighbors are considered for developing CMLP based WAM.
For example in Fig. 2, two nearest neighbors of generator G1
are generators G2 and G4. This is represented in the CMLP
by connecting the outputs of the cells C2 and C4 to the
inputs of the cell C1. Similarly for G4, two nearest neighbors
are G2 and G3 and hence outputs of the cells C2 and C3
are connected to the inputs of the cell C4. This topology
allows for the scalability of the CMLP by keeping the size
of the MLP in each cell to a minimum. The MLP in each
cell consists of an input layer with four neurons, a hidden
layer with six neurons and an output layer with a single
neuron. This choice of number of neurons in the hidden
layer is determined by trial and error and this paper does not
compare and contrast against different number of neurons in
the hidden layer. Therefore, it is not claimed to be optimal.
The four inputs to the MLP in each cell consist of ∆Vref(k)
and ∆ω(k) associated with the generator represented by the
Page 3
cell and ∆ˆ ω(k) associated with the generators represented by
the two nearest neighboring cells. The output of the CMLP
is ∆ˆ ω(k +1) of the generator associated with the cell, where
k is the sample index of the signal. This is explained in Fig.
2. A CMLP for the 12bus system consists of an identical
architecture with three cells and is shown in Fig. 1.
Fig. 4 shows the implementation of the WAM using a three
layered feedforward MLP for predicting the speed deviations
of the three generators in the 12bus system. It consists of six
neurons in the input layer, 10 neurons in the hidden layer and
three neurons in the output layer, one output representing the
stepahead predictions of speed deviation for each generator.
The six inputs to the network are the two inputs (∆ω,∆Vref)
going into the WAM from each generator. The second test
system is formulated similarly with eight input, 15 hidden and
four output neurons for predicting the speed deviations of the
four generators in the twoarea fourmachine system. and is
shown in Fig. 5.
) 1
+
(
2
∆
∧
k
G
ω
) 1
+
(
3
∆
∧
k
G
ω
) 1
+
(
4
∆
∧
k
G
ω
)(
4k
G
ω∆
)(
3k
G
ω∆
)(
2kVrefG
∆
)(
4kVrefG
∆
)(
3kVrefG
∆
)(
2k
G
ω
∆
Fig. 4.
system).
Implementation of WAM using MLP for Test System I (12bus
III. CNN TRAINING
Theneuralnetworksaretrainedonlineusing
backpropagation algorithm [11]. In this approach, weights of
the neural network are updated after every sample is passed
through the network. After all the samples are covered, this
process is repeated for as many passes through the network as
required to achieve better convergence, as explained in [11].
Values of various parameters involved in training are listed in
Table I. The training data is collected from the test systems
designed on RSCAD and simulated on a Realtime Digital
Simulator [12]. During the forced training, all of the generators
are simultaneously perturbed using a pseudorandom binary
signal (PRBS) (shown in Fig. 6) applied to the excitation
system of the generators. The deviation of the generator
) 1
+
(
2
∆
∧
k
G
ω
) 1
+
(
3
∆
∧
k
G
ω
) 1
+
(
4
∆
∧
k
G
ω
)(
4k
G
ω
∆
)(
3k
G
ω
∆
)(
2kVrefG
∆
)(
4kVrefG
∆
)(
3k VrefG
∆
)(
2k
G
ω
∆
)(
1k
G
ω∆
)(
1kVrefG
∆
) 1
+
(
1
∆
∧
k
G
ω
Fig. 5.
fourmachine system).
Implementation of WAM using MLP for Test System II (twoarea
0510
−0.2
0
0.2
0510
−0.1
0
0.1
0510
−0.2
0
0.2
0510
−0.2
0
0.2
PRBS magnitude
Time (s)
0510
−0.2
0
0.2
0510
−0.2
0
0.2
0510
−0.1
0
0.1
05 10
−0.2
0
0.2
∆Vref (pu)
Fig. 6. PRBS signals applied to the four generators and the resulting ∆Vref.
speed as a result of the PRBS perturbation is recorded
along with the reference voltage applied to the generator
excitation system (∆Vref in Fig. 3). The MIMO MLP is
trained using these two signals of each generator as the inputs.
Page 4
TABLE I
PARAMETERS USED FOR TRAINING MLP
Trials
Number of passes
Learning Rate (µ)
Momentum Gain (δ)
50
100
0.005
0.001
In case of CMLP, each cell is treated as an “object”
and therefore, all of the cells are simultaneously trained
with similar parameters. Since no parallel hardware/software
platform is used in this study, the cells are trained sequentially,
one cell after the other. However, the property of their parallel
implementation is still maintained. The parallel training
approach of each cell object of the CMLP is explained
further.
Each cell consists of four inputs viz. actual reference voltage
applied to the generator excitation system ∆Vref(k), actual
speed deviation of the generator ∆ω(k) and the predicted
speed deviations of the nearest two generators, ∆ˆ ωk1(k) and
∆ˆ ωk2(k). For every sample of the input data I(k), each cell
produces a stepahead predicted output O(k + 1). Therefore,
for any input data of size 1,2,...,k,...,K discrete samples,
and Wn and Vn be the input and output weight matrices
respectively of the MLP in nthcell, then the output of each
cell is given by:
On(k)=
=
∆ˆ ωn(k)
f (In(k − 1),Wn(k),Vn(k))
(1)
Thus, In(k) = [∆Vrefn(k) ∆ωn(k) ∆ˆ ωn1(k) ∆ˆ ωn2(k)]
uses the predicted output of the previous sample in case of
the neighboring cells n1 and n2. This helps the parallelization
of the cell objects, as long as the calculation of each sample
output is synchronized among the different cells. In MATLAB,
this is achieved by training each cell sequentially for every
sample. After the output is calculated for each cell, the weights
of each cell are updated before calculating the output for the
next sample. This process of online training of a CMLP using
backpropagation is shown in the flowchart of Fig. 7. The part
in the flowchart surrounded in dark box shows the process that
can be implemented in parallel irrespective of the number of
cells when a suitable platform is available.
IV. RESULTS AND DISCUSSIONS
A. Test System I
For Test System I, only one operating point is considered.
A MIMO MLP is trained and tested on the same data set
using the architecture explained above. The same training data
is also used to train a CMLP consisting of three cells. Fig.
8 shows the actual versus predicted speed deviations of the
three generators obtained from CMLP. The mean absolute
error (MAE) between the actual and the predicted outputs for
the two networks are calculated for comparison. The average
and standard deviation of the MAE obtained over 50 trials are
presented in Table II.
?
? ?
?
??? ?
?
?
?
? ?!
?
? " #
?$%
?&'
?(
?
?)
?
*
!+
,

#
?
" #
?$%
?&'
?(
?
?)
?
*
!.
,
?/0
12
/0
1
?/0
?
3?/4?
*
5
"? 3 3
%
3 6
?
3
2
/7 ?
(
" ?
%
5
3?
?
%
6
8
9
:
;
<
=
=+
∧∧∧
)( ),(),(),(),(),() 1(
21
kVkWkkkkVfk
nn
nn
nrefn
n
ωωωω
?
3?/>
5
?
5@
%
? ?3 6
+
,A
2
B
0
C
/+
,A
2
C
BD
+
,
.
,A
2
B
0
C
/.
,A
2
C
BD
.
,
E
3??&F?
@
2
"
)
"?& ?
?
? ?
1
@
??
@
# ??
?%
D
+
G
??
H
D
.
G
2
/
2
B
0
?/0
?/?
B
0I
?
H
J K L
MN
MN
J K L
JK L
MN
Fig. 7.Flowchart for training of CMLP using backpropagation.
B. Test System II
Different operating points shown in Table III are considered
for the Test System II. A CMLP consisting of four cells is
trained on OP1 and tested on OP1, OP2 and OP3. These
operating points are different to each other in the amount
of power transfer between the two areas of the test system.
Testing data is also obtained for operating point OP4 by
causing a 10cycle 3phase to ground fault on bus 8 of the
test system during OP1 steady state conditions. Similarly,
operating point OP5is obtained by causing a line outage on
one of the two transmission lines between the buses 7 and 8
in the test system.
Fig. 9 shows the convergence diagram for the four outputs of
the MIMO MLP. Similar convergence diagram for the CMLP
is shown in Fig. 10. These diagrams show how the mean
squared error (MSE) between the actual and the predicted
outputs decreases over multiple passes of the training data
through the network. The testing outputs obtained from the
CMLP for the five operating points are shown in Figs. 11 to
Page 5
TABLE II
COMPARISON OF MLP AND CMLP IN TEST SYSTEM I
G2G3G4
MLP
0.018403
0.001395
0
CMLP
0.017512
0.000860
1
MLP
0.016520
0.001298
0
CMLP
0.015853
0.000559
1
MLP
0.014042
0.000949
0
CMLP
0.014383
0.000677
1
Avg.
Std.
Winner
0246810
−5
0
5x 10
−4
∆ω2 (pu)
actual
predicted
0246810
−5
0
5x 10
−4
∆ω3 (pu)
0246810
−5
0
5x 10
−4
Time (s)
∆ω4 (pu)
Fig. 8.Output of the CMLP based WAM for the 12bus system
TABLE III
FIVE OPERATING POINTS CONSIDERED IN THE STUDY
OP1, OP4, OP5
950
1650
253.2
22.68
705.6
163.5
705.5
296
441.5
68.8
705.6
169.8
OP2
556
1469
302.9
57.2
573.8
117.2
537.7
234.6
309.5
49.79
537.7
140.1
OP3
950
944
80.45
38.02
579.5
53.89
579.1
81.12
314.4
31.56
578.6
59.53
Load 1 (MW)
Load 2 (MW)
Parea1⇔area2(MW)
Qarea1⇔area2(MVar)
PG1(MW)
QG1(MVar)
PG2(MW)
QG2(MVar)
PG3(MW)
QG3(MVar)
PG4(MW)
QG4(MVar)
15. The comparison of absolute errors obtained using MLP
and CMLP for OP1 to OP5 are shown in Figs. 16 to 20,
respectively. The average and standard deviation of the mean
absolute error (MAE) obtained by the two networks during
testing on five operating points over 50 trials are shown in
Table IV.
C. Analysis
Learning in CNN is a challenging task because of
the connectivity between several cells that are learning
concurrently. Since the predicted output from one cell is used
as input(s) to other neighboring cell(s), errors due to poor
training and hence false predictions of the NN in one cell can
ripple through all of the cells and deteriorate the performance
02040 60 80100
0
1
2
3
4
5
6
7x 10
−3
Epochs
Average MSE
G1
G2
G3
G4
Fig. 9. Convergence of individual outputs of the MIMO MLP during training.
0 2040 6080100
0
0.5
1
1.5
2
2.5x 10
−3
Epochs
Average MSE
G1
G2
G3
G4
Fig. 10. Convergence of individual cells of the CMLP during training.
of the CNN. On the other hand it is also arguable that the
NNs get trained even better due to the connectivity because
the errors propagate through the network and each cell is
trained actively (through its own training) and passively
(through the training of its neighbors) as training algorithm
on each cell tries to minimize the error at its output. This
way, knowledge of the actual dynamics of the system is
preserved not only on the individual neural networks at each
cell, but also on the connectivity between the different cells of