DYNAMIC STABILITY EVALUATION OF SINGLE-MACHINE INFINITE

POWER SYSTEM USING RBF NEURAL NETWORK

E. A. Feilat
Department of Electrical Power Engineering
Yarmouk University
Irbid, Jordan
E-mail: eafeilat@yu.edu.jo
ABSTRACT
This paper presents a neural network based approach to
evaluate the dynamic stability of single-machine infinite
bus power system. Dynamic stability evaluation based on a
radial basis function (RBF) neural network is used to
associate real and reactive powers (P-Q) patterns with the
synchronizing and damping torque coefficients (Ks and Kd)
and the local mode damping coefficient and damped natural
frequency ( and ). The performance of the RBF neural
network was also compared in this paper with that of the
back-propagation (BP) neural network. Test results show
that the proposed approach can be used as an effective tool
for fast dynamic stability evaluation over wide range of
operating conditions.
KEYWORDS: Neural networks, RBF, BP, dynamic
stability, synchronizing and damping torques
INTRODUCTION
Power systems experience low-frequency oscillations due
to disturbances. These low frequency oscillations, between
1 and 2 Hz, result from the existence of one or several
poorly damped, or even undamped, low frequency modes.
The study of these oscillations is very important to power
system planning, operation, and control. The stability of
these oscillations is a vital concern and essential for safe
power system operation. Eigenvalue analysis is one of the
conventional methods in dealing with the dynamic stability
problem (Rouco and J. Pereze-Arriaga 1993; Gross et al.
1982; Klein et al. 1991). The stability is evaluated by
calculating the eigenvalues of the system matrix in the
linearized dynamic operation. Despite the robustness of this
technique, the method however has a drawback that it
requires a lot of computations as the size of the power
system becomes larger. In other words, the method is not so
practical in a sense that it has to evaluate all the eigenvalues
of the system matrix although the power system operator is
interested in the most critical eigenvalue. In addition, the
method is not suitable for online application. As a result,
alternative approaches have been developed to overcome
the problem of dynamic stability evaluation (DeMello and
Concordia 1968; DeOliveria 1994; Padiyar and Varma
1991; Alden and Shaltout 1979). An example, of these

approaches that is suitable for both offline and online

applications is the calculation of the synchronizing and
damping torques (Ts and Td) developed by the synchronous
machine (Padiyar and Varma 1991; Alden and Shaltout
1979; Yu 1983). The synchronizing and damping torques
are usually expressed in terms of the torque coefficients Ks
and Kd.
Several methods have been devised to calculate Ks and
Kd. An efficient least-square error technique (LSET) has
been developed (Alden and Shaltout 1979). The technique
is based on using time responses of the changes in rotor
angle, rotor speed, and electromagnetic torque, (t), (t),
and Te(t). These responses can be obtained by computer
simulation or from measured data. The significance of this
method is that it permits the calculation of the torque
components for a machine of any degree of complexity and
takes into consideration the effect of all system parameters.
It is known that power system operating conditions
change with time in real-time situations. Therefore, a
dynamic stability analysis must be repeatedly conducted in
system operation and control to provide estimates of
dynamic stability indices and provide new estimates as new
data are received. This requires significant time and
computational efforts. To avoid this burden, an online
scheme should be developed which can quickly provide
estimates of the indices on basis of data obtained by
measurements.
In general, for online estimation, the application of
artificial neural networks (ANNs) seems to be very
promising as demonstrated by many investigators in several
areas. Their applications in power systems included
security assessment, load forecasting, fault location, and
dynamic stability analysis (El-Sharkawi and Niebur 1988).
Most of these applications have been implemented using
multilayer back propagation (BP) and radial basis functions
(RBFs) neural networks. ANNs have several advantages
over conventional computing methods. These advantages
include the ability to handle situations of incomplete and
corrupted data, ability to learn from examples, and the
ability to memorize. Based on these properties, and the
several successful applications, ANNs seem to be good
candidate for online estimation of power system dynamic
stability.

In this paper, a useful ANN based technique is used to

evaluate the dynamic stability indices using RBFNN. The
technique is based on evaluation of Ks and Kd and the real
and imaginary parts ( and d) of the low-frequency mode
from measured patterns of the generator real power and
reactive power (P and Q), which represent the generator
loading condition. A number of test results were simulated
and analyzed. A comparison between RBFNN and BPNN is
also presented in this paper. It is shown that the RBFNN is
very effective in estimating the dynamic stability indices.
DYNAMIC STABILITY CONCEPT
The dynamic response of a single machine connected to
an infinite bus comprises various modes of oscillations as
given by
x (t )=

(e )U V x
it

(1)

i i o

where i = i ji are the distinct eigenvalues with a

corresponding set of right and left eigenvectors, Ui and Vi,
respectively. xo is the initial state vector. These modes of
oscillations can be classified into three groups, namely,
stator and network transient modes, field and rotor circuits
modes, and low-frequency electromechanical modes. The
frequency band of each group is distinct from the others
and, thus, they are fairly decoupled. The oscillation of the
electromagnetic torque and, consequently, the rotor
oscillations are dominated by the low-frequency
electromechanical modes, = j. Various methods
have been proposed to break the electromagnetic torque
variations into two components; the synchronizing torque
component (Ts) is in phase and proportional with (t), and
the damping torque (Td) is in phase and proportional with
(t) (DeMello and Concordia 1968; DeOliveria 1994;
Padiyar and Varma 1991; Alden and Shaltout 1979). Ts is
responsible for restoring the rotor angle excursion, while Td
damps out the speed oscillations.
Accordingly, the estimated torque can be written as
~
Te = K s (t ) + K d (t )

(2)

For the readers convenience, calculation Ks and Kd using

LSET is briefly summarized. Following a small
disturbance, the time responses of (t), (t), and Te(t)
are obtained. The LSET is then used to minimize the sum
of the square of differences between the electromagnetic
~
torque Te(k) and the estimated torque Te (k ) over the
interval of oscillation t, where, t = NT ( N is the number of
samples, T is the sampling period, and k is the sample
index). The sum of the square of the error is defined as
N

E=

(T (k ) K (k ) K (k ))
e

(3)

k =1

Minimizing E yields the corresponding values of Ks and Kd.

SYSTEM REPRESENTATION
A nonlinear seventh-order model is used to represent a
synchronous generating unit connected to an infinite bus
through a tie line. The parameters of the power system
under study are given in Appendix A. A schematic diagram
of the system is shown in Fig. 1. For dynamic stability
studies, the nonlinear differential equations as given in
Appendix B have to be first linearized around a nominal
operating condition to obtain the desired state-space
equations
x&(t)= Ax(t)+Bu(t)
(4)
y(t) =Cx(t)
where
x(t) = [fd(t) d(t) kd(t) q(t) kq(t) (t)
(t)]T
is the state vector, y(t) = [(t) (t) Te(t) ]T is the output
signals vector, u is the disturbance vector such as the
mechanical torque, the excitation voltage, or the infinite bus
voltage. A, B, and C are constant matrices whose values
depend on the system operating conditions, parameters, and
the input signal.
V

VG
Re + j Xe
G

Figure 1: Single-Machine Infinite Bus System

ARTIFCIAL NEURAL NETWORK
Multilayer feedforward ANNs have been applied
successfully to solve some complex practical problems for
which conventional approaches have proven ineffective. In
its basic form a feedforward NN consists of an input layer,
an output layer, and one or more hidden layers. Each layer
consists of a set of neurons or nodes that are fully
connected to the neurons in the next layer. The connections
have multiplying weights associated with them. The
number of neurons and hidden layers is problem dependent.
Feedforward ANNs are divided into two groups depending
on the type of activation functions and learning methods.
The first type is the well-known multilayer perceptron
(MLP), and the second one is the radial basis functions
(RBF) network (Haykin 1999). These types of ANNs are
good candidates for function approximation, pattern
classification, and nonlinear mapping,. A typical ANN with
one hidden layer is depicted in Fig. 2.
The process of determining the weights is called training
or learning process. In the training process, sets of inputoutput patterns are associated by properly adjusting the
weights in the network such that an error measure (the
difference between the target and the predicted outputs of

the network) is minimized. A sum of squared error function

is commonly used
1
(p) 2
E=E p=
( t (p)
(5)
k - Ok )
2 p k

suitable values for the width and center of the function. The
main advantage of the RBFNN over the MLPNN is fast
learning.

where tk is the target output, and Ok is the predicted output

of the neural network. Various training algorithms have
been developed to adapt the weights in ANNs to reduce the
error defined in Equation (5) (Haykin 1999).

TEST RESULTS

O1

O2

p1

O3
Pm
On

Input Layer

Hidden Layer

Output Layer

Figure 2: Two-Layer Feedforward Neural Network

Backpropagation Neural Network
Multilayer perceptron neural networks have been applied
to nonlinear system modeling and function approximation
problems. It has been proven that a network with one
hidden layer can perform any nonlinear mapping and no
more than two hidden layers are needed for most
applications (Haykin 1999). Normally, the transfer function
for the hidden layers is the sigmoid, and the linear for the
output layer. Several training techniques have been
developed to adapt weights in the MLPNN. In the present
paper, the backpropagation (BP) training technique will be
used. The BPNN uses a gradient descent rule to adapt the
weights, and the error is calculated and propagated
backwards from the output to the hidden layer to the input.
Radial Basis Functions Neural Network
The RBF neural network has a topology of one hidden
layer. Each output neuron is a linear combination defined
by

The proposed RBFNN and BPNN for estimating the

dynamic stability indices each consists of three layers. The
input layer contains two neurons for the real power P and
reactive power Q. The output layer has four neurons for Ks,
Kd, , and . For the RBFNN, the number of neurons in the
hidden-layer is determined by the LS learning procedure to
reach the satisfactory error value, and is equal to the
number of epochs. A compromise criterion between
accuracy and training speed were used to select the number
of hidden neurons of the BPNN. Training sets consisting of
70 input-output patterns were used to represent the dynamic
stability behavior over wide range of loading conditions of
P and Q from 0.1 to 1.0 pu, and for small mechanical
torque disturbance of 0.2 pu for 20 ms. Both the networks
were trained using the Matlab Neural Network Toolbox
routines such as the RBF and BP with momentum.
The convergence rates of the two networks with related
sum of squared errors (SSE) are listed in Table 1. The error
convergence of the two networks is depicted in Fig. 3. The
RBFNN possesses 60 neurons and reached to the
satisfactory SSE after 60 epochs of iterations, which is
equal to the number of hidden neurons. For the BPNN,
three cases of 20, 40, and 60 hidden neurons were
examined. The results in Table 1 show that the performance
of the RBFNN in terms of the training accuracy is much
better than that of the BPNN. Although, RBFNN has a
larger size than the BPNN, the RBFNN is still a good
candidate for high-speed online and real-time applications.
Table 1: Error Convergence of RBFNN and BPNN
Network
Nh
Epochs
SSE
RBF
60
60
0.0018
20
20000
2.48
BP
40
5629
1.55
60
2270
0.56

where j ( x p ) is a radially symmetric function. The most

Figure 4 show the target and predicted values of Ks, Kd,

, and obtained by the RBFNN. The figures show that
the neutral network was successful in predicting the four
stability indices. The SSE is about 0.0018, which was
obtained after 60 epochs, with one epoch corresponding to
one cycle through the presentation of all the training pattern
pairs.

common basis function is the Gaussian

( x c j ) 2
Ok ( x p ) = exp
(7)

2
2 j

where j and cj are the width and center of the radial basis
function, respectively. The weights wkj can be determined
by using the least-squares (LS) method after selecting

Figure 5 show the target and predicted values of Ks, and

Kd obtained by the BPNN. The SSE reduced to 0.5599 after
2270 epochs of iterations. Examining the performance of
both networks, one can see that the RBFNN takes less CPU
time and it is more stable when compared to that of the
BPNN.

Ok ( x p ) =

w (x
kj

) + wo

(6)

1 0

1 0

1 0

15

10

SSE

1 0

1 0

1 0

1 0

-1

-2

-3

1 0

2 0

3 0

4 0

5 0

6 0

E poch

Figure 3-a: Training Error Converge for the RBFNN

Nh = 60, Epochs = 60, SSE = 0.0018
1 0

1 0

1 0

0 .1

0 .2

0 .3

0 .4

0 .5

0 .6

0 .7

0 .8

0 .9

P (p u )

Figure 4-d: Target and Predicted Values for using

RBFNN, P=0.1-1.0 pu, Q=0.2 pu, Nh=60
4

3 .5
3

2 .5

M S E
1 0

1 0

1 0

Ks
p u /ra d

-1

-2

2
1 .5
1

-3

5 0 0

1 0 0 0

2 0 0 0

1 5 0 0

0 .5

E po ch

Figure 3-b: Training Error Converge for the BPNN

Nh = 60, Epochs = 2270, SSE = 0.5599

3 .5

0
0

0 .1

0 .2

0 .3

0 .4

0 .5
P (p u )

0 .6

0 .7

0 .8

0 .9

Figure 5-a: Target and Predicted Values for Ks using

BPNN, P=0.1-1.0 pu, Q=0.2 pu, Nh=60
0 .2

0 .1

0 .1

K s 2 .5
p u /ra d
2

0 .1
K d 0 .1
p u /ra d /s
0 .1

1 .5

0 .0

0 .0

0 .5
0

0 .0

0 .1

0 .2

0 .3

0 .4

0 .5
P (p u )

0 .6

0 .7

0 .8

0 .9

0 .0

Figure 4-a: Target and Predicted Values for Ks using

RBFNN, P=0.1-1.0 pu, Q=0.2 pu, Nh = 60
0 .2

0 .1

0 .1

0 .2

0 .3

0 .4

0 .5

0 .6

0 .7

0 .8

0 .9

P (p u )

Figure 5-b: Target and Predicted Values for Kd using

BPNN, P=0.1-1.0 pu, Q=0.2 pu, Nh=60

0 .1
0 .1

CONCLUSION

0 .1
Kd
p u /r a d /s
0 .1
0 .0
0 .0
0 .0
0 .0
0

0 .1

0 .2

0 .3

0 .4

0 .5

0 .6

0 .7

0 .8

0 .9

P (p u )

Figure 4-b: Target and Predicted Values for Kd using

RBFNN, P=0.1-1.0 pu, Q=0.2 pu, Nh = 60
0

-1
-2
-3

-4
-5
-6
-7
-8
-9
-1 0
0

0 .1

0 .2

0 .3

0 .4

0 .5
P (p u )

0 .6

0 .7

0 .8

0 .9

Figure 4-c: Target and Predicted Values for using

RBFNN, P=0.1-1.0 pu, Q=0.2 pu, Nh=60

In this paper, an RBFNN used to evaluate the dynamic

stability of a single machine connected to an infinite bus.
The proposed RBFNN consist of an input layer with two
neurons, a hidden layer, and an output layer with four
neurons. Pairs of online measurement of the generator real
power (P) and reactive power (Q) are chosen as the input
signals to the RBFNN. The outputs are the corresponding
synchronizing and damping torque coefficients, and the real
and imaginary parts of the low-frequency mode. These
indices not only reflect the relative stability of the
synchronous machine, but also can predict the possible
loading condition which render the system unstable.
The RBFNN was also compared with the BPNN.
Computer simulation for various generator loading
condition show that the proposed RBFNN approach is a
reliable and generalized idea for dynamic stability
evaluation. The proposed RBFNN could provide great
performance for high-speed identification of the dynamic
stability of the power system at various loading conditions.

It can be implemented for online or real-time estimation of

the dynamic stability indices. It can also be used in
conjunction with a neural controller to tune the parameters
of power system stabilizer.

SMA. IEEE Trans. on Power Systems, Vol.. PWRS-8, No.3

(Aug), 1257-1265.
Yu Y. N. 1983. Electric Power System Dynamics, Academic
Press, NY.

APPENDIX A

-1
-2
-3

The parameters of the synchronous generator and

transmission line are given in Table A.

-4

-5
-6

Table A: Synchronous Generator and Transmission Data

-7
-8
-9
-1 0
0

0 .1

0 .2

0 .3

0 .4

0 .5

0 .6

0 .7

0 .8

0 .9

P (p u )

Figure 5c: Target and Predicted Values for using

BPNN, P=0.1-1.0 pu, Q=0.2 pu, Nh=60

Xkkq = 1.52
Xakq = 1.49
Xkkd = 1.605
Xd = 1.70

Xq = 1.64
Ra = 0.001
Rkd= 0.0132
Rkq= 0.0132

Rfd = 0.0007
Re = 0.03
Xe = 0.6
H = 4.0

APPENDIX B

15

The nonlinear equations that describe the dynamics of the

synchronous-machine are given below.

10

Xffd= 1.65
Xafd= 1.55
Xfkd= 1.55
Xahd= 1.55

0 .1

0 .2

0 .3

0 .4

0 .5

0 .6

0 .7

0 .8

0 .9

P (p u )

Figure 5-d: Target and Predicted Values for using

BPNN, P=0.1-1.0 pu, Q=0.2 pu, Nh = 60
REFERENCES
Alden R.T.H. and A.A. Shaltout. 1979. Analysis of Damping and
Synchronizing Torques - Part I: General Calculation
Method. IEEE Trans. on Power Apparatus and Systems,
Vol. PAS-98, No.5 (Sept/Oct), 1696-1700.
DeMello F.P. and C. Concordia. 1968. Concepts of Synchronous
Machine as Affected by Excitation Control. IEEE Trans. on
Power Systems Apparatus, Vol. PAS-87, (March), 835-844.
DeOliveria S.E.M. 1994. Synchronizing and Damping Torque
Coefficients and Power System Steady-State Stability as
Affected by Static VAR Compensation. IEEE Trans. on
Power Systems, Vol. PWRS-9, No.1(Feb), 109-119.
El-Sharkawi M. and D. Niebur. 1988. Artificial Neural Networks
with Applications to Power Systems. IEEE Service Center,
NJ.
Gross G., C.F. Imparato, and P.M. Look. 1982. A Tool for the
Comprehensive Analysis of Power Systems Dynamic
Stability. IEEE Trans. on Power Systems Apparatus, Vol.
PAS-101, No.1 (Jan), 226-234.
Haykin S. 1999. Neural Networks, A Comprehensive Foundation.
2nd ed., Prentice Hall, NJ.
Klein M., G.J. Rogers, P. Kundur. 1991. A Fundamental Study of
Inter-Area Oscillations in Power Systems. IEEE Trans. on
Power Systems, Vol. PWRS-6, No.3 (Aug), 914-921.
Padiyar K. R. and R. K. Varma. 1991. Damping Torque Analysis
of Static VAR System Controller. IEEE Trans. on Power
Systems, Vol. PWRS-6, No.2 (May), 458-465.
Rouco L. and J. Pereze-Arriaga. 1993. Multi-Area Analysis of
Small Signal Stability in Large Electric Power Systems by

p fd = (V f i f R fd )
p d = (Vd + id ( Ra + Re ) + q ) + q
p q = (Vq + iq ( Ra + Re ) d ) d
p kd = ikd Rkd
p kq = ikq Rkq
p =

p = (Tm Te )
2H
Te = d iq qid
where:
fd X ffd
X
d afd
kd = X fkd

q 0
kq 0

X afd
Xd
X akd
0
0

X fkd
X akd
X kkd
0
0

0
0
0
Xq
X akq

0 i fd
0 id

0 ikd
X akq iq

X kkq ikq

BIOGRAPHY
E. A. Feilat obtained his B.Sc. in Electrical Engineering in
1987 from the University of Jordan, his M.Sc. in Electrical
Machines and Power Systems in 1989 from the Jordan
University of Science and Technology, and his Ph. D. in
High Voltage Engineering from Mississippi State
University in 2000. From 1990-1996 he worked with King
Fahd University of Petroleum and minerals as Lecturer at
the Electrical Engineering Department. In 2001, he joined
the Electrical Power Engineering Department at Yarmouk
University as Assistant Professor. His area of research
includes high voltage engineering and application of
artificial intelligence to power system engineering.