Expert Systems with Applications 37 (2010) 5023–5031

Contents lists available at ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa

Artificial Intelligence application to Malaysian electrical powersystem

Ahmed M.A. Haidar a,*, Azah Mohamed b, Aini Hussain b, Norazila Jaalam a
a
Faculty of Electrical and Electronics Engineering, University Malaysia Pahang (UMP)-Pahang, Lebuhraya Tun Razak, 26300 Kuantan, Pahang Darul Makmur, Malaysia
b
Department of Electrical, Electronic and Systems Engineering, National University of Malaysia (UKM)-Selangor, Malaysia

a r t i c l e i n f o a b s t r a c t

Keywords: Vulnerability assessment and control of a power system is important to power utilities due to the black-
Vulnerability assessment outs in recent years in many countries which indicate that power systems today are vulnerable when
Power system loss exposed to unforeseen catastrophic contingencies. A fast and accurate technique to assess the level of
Neural network system strength or weakness is some of the essential requirements for maintaining security of modern
Fuzzy controller
power systems, particularly in competitive energy markets. This paper presents intelligent artificial tech-
niques for vulnerability assessment of Malaysian power system and recommends preventive control
measures. Accurate techniques for vulnerability assessment and control of power systems are developed.
In vulnerability assessment, power system loss index is used as a vulnerability parameter, neural network
weight extraction is employed as the feature extraction method and the generalized regression neural
network is used to predict vulnerability of a power system. As for vulnerability control, load shedding
is considered by using the neuro-fuzzy technique. Finally, the paper presents and discusses the results
from this research with recommendations.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction Kezunovic, 2005). It is now generally recognized that the proper

framework for security assessment should be vulnerability assess-
Security of electricity supply networks has always been a key ment that incorporates information about both the security level
point in the development of the power industry. Several cascading as well as its trend in the security status.
failures and large area blackouts occurring in the USA, some Euro- Analysis of recent widespread outages demonstrates that black-
pean and Asian countries highlighted the need for vulnerability outs rarely happened and are usually caused by a sequence of low-
assessment of power systems. Nowadays, power systems have probability disturbance which is generally not expected by system
evolved through continuing growth in interconnection, use of operators. If fast control actions such as load shedding and gener-
new technologies and controls. Due to the increased operations ation rejection are not taken proactively, the system may cascade
which may cause power system to be in highly stressed conditions, and separate into unplanned islands. Normally, in emergency and
the need for vulnerability assessment of power systems is impor- abnormal conditions, a power system operator has to deal with a
tant so as to determine its ability to continue proving service in large amount of data and apply the most appropriate remedial con-
case of any unforeseen catastrophic contingency such as power trol actions. However, due to emotional and psychological stress,
system component failures, communication system failures, an operator may not be able to adequately respond to critical con-
human operator error and natural calamity (Kim, El-Sharkawi, & ditions and make correct decisions. Mistakes can damage very
Marks, 2005). Therefore, accurate vulnerability assessment is very expensive power equipment or worse still lead to the major emer-
important and a vulnerability index is significantly needed to help gencies and catastrophic situations. Clearly, there is a strong need
power system operators to steer the system to viable conditions. for automated corrective actions that can assist operators in vul-
Traditionally, power system security assessment is concerned with nerability control (Franco, 2001). For power systems which are
estimating the security level of a system and it encompasses the operated closer to their stability limits, it is desirable to use load
vulnerabilities resulting in voltage insecurity, static insecurity shedding when there is a lack of adequate spinning reserve margin
and dynamic insecurity. However, in vulnerability assessment, it and a shortage of tie line capacity.
has to take into account not only the security information but Power systems encompass many input information about the
the information of the whole system considering generation loss system state which includes load flow information such as voltage,
due to the generation outage, power loss due to the line outage, real and reactive power flows, voltage angle, generated and
increase in total load and amount of load disconnected (Song & demand powers. High dimensional datasets are becoming more
abundant when solving classification problems and therefore, the
* Corresponding author. data information needs to be reduced to a smaller number of infor-
E-mail address: ahaidar67@yahoo.com (A.M.A. Haidar). mation (Jensen, El-Sharkawi, & Marks, 2001). The most common

0957-4174/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2009.12.010
5024 A.M.A. Haidar et al. / Expert Systems with Applications 37 (2010) 5023–5031

approach to feature management is by means of feature extraction. the output layer. The GRNN implements the Bayesian decision
A major challenge in feature extraction is to extract a set of fea- strategy to classify input vectors. A schematic diagram of GRNN
tures, as small as possible, that accurately classifies the learning is depicted in Fig. 1 in which it consists of four layers, namely, in-
examples. The extracted set of features must represent the entire put layer, pattern layer, summation layer and output layer. The
system, since a loss of information in the reduced set results in loss number of input units in the first layer is equal to independent fac-
of performance and accuracy (Mori, 2006). In the literature, there tors, xi . Only the hidden layer has biases. The first layer is fully con-
are many different feature extraction methods applied to power nected to the pattern layer, whose output is a measure of the
systems. The common feature extraction methods are such as the distance of the input from the stored patterns. Each pattern layer
use of principle component analysis (Sawhney & Jeyasurya, 2004; unit is connected to the two neurons in the summation layer,
Simon, 2005) and neural network based feature extraction tech- known as S summation neuron and D summation neuron. The S
nique (Kim, 2002). summation neuron computes the sum of the weighted outputs of
The term Artificial Intelligence (AI) generally refers to a group of the pattern layer while the D summation neuron calculates the un-
techniques that in some sense attempt to mimic human intelli- weighted outputs of the pattern neurons (Kim, Junki, & Byung,
gence. These paradigms include several technologies such as Arti- 2006). For D-summation neuron, the connection weight is set to
ficial Neural Network (ANN), Fuzzy Logic (FL), Genetic Algorithms unity. The output layer merely divides the output of each S-sum-
and Particle Swarm Optimization. AI techniques including ANN mation neuron by that of each D-summation neuron, yielding the
and FL have shown great promise as ways for solving certain diffi- predicted value expressed as:
cult problems in power systems, where the conventional methods Pn
^ yi exp½Dðx; xi Þ
have not achieved the desired speed and accuracy. ANN approach yi ¼ Pi¼1
n ð1Þ
for solving power system problems considers various ANN models i¼1 exp½Dðx; xi Þ

but the multi-layer perceptron is the earliest and most widely used where n is the number of independent input variables, yi is the tar-
model. In recent years, another ANN model known as radial basis get output value corresponding to the ith input pattern and the
function neural network (RBFNN) which has non-linear mapping Gaussian D function is defined as:
capability, has become increasingly popular due to its structural
p 
X 2
simplicity and training efficiency. RBFNN has been used for solving xj  xij
Dðx; xi Þ ¼ ð2Þ
power system problems such as for protection of transmission j¼1
rj
lines (Dash, Pradhan, & Pandaa, 2001), locating faults in transmis-
sion lines (Joorabian, Taleghani, & Aggarwal, 2004) and transient where p is the total number of training patterns and rj is generally
stability assessment of power systems (Dash, Mishra, & Panda, referred to as the smoothing parameter (width or spread), whose
2000). Other techniques that are widely used for solving power optimal value is often determined experimentally.
system problems are fuzzy logic and fuzzy expert system. Fuzzy lo- The GRNN has a special property in which no iterative training
gic was applied for safety analysis of power system protection and of the weight vectors is required. That is, any input–output map-
automation system (Manana, Toader, & Anatoli, 2004) and fuzzy ping is possible by simply assigning the input vectors to the cen-
expert system for voltage instability control to calculate the opti- troid vectors and fixing the weight vectors between the radial
mum amount of load shedding (Sallam & Khafaga, 2002). In addi- basis function units and outputs identical to the corresponding tar-
tion, a fuzzy logic stabilizer was developed for stability control of get vectors. This training algorithm is much better than the back
a laboratory scale model of a power system (Al-Osaimi, Abdennour, propagation training which involves long and iterative training as
& Abdullaziz, 2005). well as facing the problem of local minima. Moreover, a special
In this paper, generalized regression neural network (GRNN) is property of GRNN is that it enables users to flexibly configure the
used for assessing vulnerability of Malaysian practical power sys- network suitable for real hardware implementation, by adjusting
tem based on the proposed vulnerability indices using power sys- only two parameters which are the center and width. Since the ra-
tem loss (PSL) indices. GRNN models belong to the RBFNN group is dial basis function acts as a detector for different input vectors, the
considered a developed model of RBFNN architecture, it was cho- weight vectors are computed accordingly and there is no need to
sen because it is indicated to classifying problems and for the fact train the network. Thus, the GRNN is more straightforward and
that it has never been used before to create pattern recognition does not require a training process.
systems on vulnerability assessment. A new feature extraction
technique named as neural network weight extraction is used for 3. Neuro-fuzzy architecture
reducing the input features so as to speed up the neural network
training process. For vulnerability control of Malaysian practical Fuzzy logic system needs rules to be defined first, but one may
power system, an intelligent load shedding scheme is proposed have no knowledge about a power system for the formation of
using neuro-fuzzy controller. The proposed controller considers rules. Therefore, automated parameters tuned by a neural network
two inputs and one output in which the inputs are the calculated embedded inside a fuzzy system can replace the need for prior
vulnerability index based on PSL and the bus voltage magnitudes
whereas the output is the amount of load shed. This paper is orga-
Input layer Pattern layer
nized as such that in Sections 2 and 3, a brief background of GRNN
architecture and neuro-fuzzy are described. Section 4 describes the
neural network implementation for vulnerability assessment and
x1 Summation layer

Section 5 describes the neuro-fuzzy implementation for load shed-

Output layer
ding. Results and conclusions are given in Sections 6 and 7, x2
respectively.
S
y
D
2. Neural network architecture and training algorithm
xn
The GRNN which is a kind of normalized RBFNN is similar to the
RBFNN in the input and hidden layers, but is slightly different in Fig. 1. Structure of GRNN model.
 A.M.A. Haidar et al. / Expert Systems with Applications 37 (2010) 5023–5031 5025

gent techniques in vulnerability assessment and control. The test

system consists of 22 generators, 177 lines and 56 loads with a
voltage class of 275 kV as shown in Fig. 3. The bus and line data
of the test system are collected from power system control center.
The training data set is generated by carrying out simulations,
in which the system behavior at the base case condition is first
analyzed. The next step is to analyze the system behavior when
subjected to credible system contingencies such as line outage
(LO), generator outage (GO), load increase (LI) and disconnection
of loads (DL). For each set of training data, the vulnerability index
based PSL is then calculated. PSL considers total system loss, gen-
eration loss due to generation outage, power line loss due to line
outage, increase in total load and amount of load disconnected.
The rational for considering PSL is due to the fact that losses in a
power transmission system are a function of not only the system
load but also of the generation. PSL is then used as the GRNN out-
put and it is given by:
Fig. 2. ANFIS Structure with two inputs, one output, and two membership function SBCL
for each input. PSL ¼ Pn Pm ð3Þ
SCCL þ SIL þ SLD þ i¼1 SLGO;i W G;i þ i¼1 SLLO;i W L;i

knowledge about a power system. Neuro-fuzzy system is a combi- where SBCL is the system power loss in MVA at base case; SCCL is the
nation of neural network and fuzzy logic in which it combines the system power loss in MVA at contingency case; SIL is the increase in
learning and adapting abilities of neural networks with the fuzzy total load in MVA; SDL is the amount of load disconnected in MVA;
interpretation of fuzzy logic system. An example of a neuro-fuzzy SLGO;i is the loss of generated MVA due to generator outage; SLLO;i is
system is the adaptive neural network based fuzzy inference sys- the loss of transported MVA due to line outage; W G;i is the weight of
tem (ANFIS) which combines the Takagi–Sugeno fuzzy inference individual generator power output; W L;i is the weight of individual
system (FIS) with neural network. The ANFIS defines five layers line power influence; n is the number of generators; and m is the
which perform the function of fuzzification of the input values, number of lines.
aggregation of membership degree, evaluation of the bases, nor- From Eq. (3), it can be noted that the proposed vulnerability in-
malization of the aggregated membership degree and evaluation dex, PSL will have values in the range of 1–0 assuming that at a
of function output values (Negnevitsky, 2002). contingency case, the losses in a power system will be greater than
A typical ANFIS structure with five layers and two inputs, each at base case. These values can be categorized by a control operator
with two membership functions is shown in Fig. 2. The five layers based on their vulnerability boundaries. If the value of PSL is close
of the ANFIS are connected by weights. The first layer is the input to 1.0, it indicates that the system is ‘Invulnerable’ whereas if the
layer which receives input data that are mapped into membership PSL value is small, that is, close to 0, it implies that the system is
functions so as to determine the membership of a given input. The ‘Vulnerable’. The assumed limits of index values can be changed
second layer of neurons represents association between input and or readjusted by a control operator based on any new system con-
output, by means of fuzzy rules. In the third layer, the output are figuration. The weight of individual generator and line is chosen
normalized and then passed to the fourth layer. The output data based on their importance considering power system operating
are mapped in the fourth layer to give output membership function practices.
based on the pre-determined fuzzy rules. The outputs are summed
in the fifth layer to give a single valued output. The ANFIS has the 4.2. Normalization
constraint that it only supports the Sugeno-type systems of first or
0th order. The system can only be designed as a single output sys- Normalization is a transformation of each feature in the data
tem and the system must be of unity weights for each rule. set. This is the step to preprocesses the network training set by
normalizing the inputs and targets so that they fall in the interval
4. Neural network implementation in vulnerability assessment [1, 1] before they are presented to the ANN. The following equa-
tion is used to normalize the input data:
Before applying GRNN for power system vulnerability assess-
2ðxi  xmin Þ
ment, the first step is to collect as many appropriate training data xiðnormÞ ¼ ð4Þ
ðxmax  xmin Þ  1
sets as possible. Such data should be both measurable in a real
power system and available from power utilities. The approach where xi is the non-normalized feature vector; xmax is the maximum
used in generating the training data set is by carrying out simula- value of feature vector; xmin is the minimum value of feature vector.
tions on a power system which is subjected to various disturbances
and gathering a set of system features along with the correspond- 4.3. Feature extraction method
ing system vulnerability index. The next step is to normalize the
inputs and targets by transforming each feature in the data set. Feature extraction is the process of mapping all available fea-
The normalized features are processed by using feature extraction tures into a composite feature set of a lower dimension. Here,
methods before presenting them to the ANN. The steps mentioned dimensionality of a feature set is reduced by combining features
above are discussed in this section. while retaining the characteristics that allow for accurate classifi-
cation. The feature extraction method proposed in this paper is
4.1. Data generation named as the neural network weight extraction (NNWE) method.
Fig. 4 shows a schematic diagram of the ANN structure used to
The practical 87 bus test system which is the Malaysian power illustrate the proposed NNWE method. The procedure of the pro-
system is used for validating the proposed computational intelli- posed NNWE method is described as follows:
5026 A.M.A. Haidar et al. / Expert Systems with Applications 37 (2010) 5023–5031

Fig. 3. Single line diagram of the Malaysian 87 bus test system.

i. Determine the original training data sets of the ANN for vul- iii. Train the first ANN using the sub-data sets for the vulnerable
nerability assessment (second ANN). cases. The first ANN is trained by considering different num-
ii. From the original training data sets, select sub-data sets ber of hidden neurons at a fixed accuracy.
whose vulnerability index based on PSL shows that the sys- iv. Obtain the weight matrix of the trained first ANN which
tem is vulnerable. These sub-data sets are applied as inputs maps the input and the hidden neurons. The weight (Simon,
for training the first ANN that is used for feature extraction. 1999) matrix is represented by the following equations:
 A.M.A. Haidar et al. / Expert Systems with Applications 37 (2010) 5023–5031 5027

The first ANN for NNWE R is the set of data with reduced features. In other words, the
number of reduced features depends on the number of hid-
den neurons in the first ANN. This number of hidden neurons
Input Layer Hidden Layers
has to be less than the number of the input variables to the
first ANN in order to get the reduced feature sets for the sec-
wnm vm
xn ond ANN for vulnerability assessment.

vi. Using the different R sets of data from (v), train the second
Output Layer ANN. The outputs of the second ANN are evaluated for differ-
ent R sets of data. The most accurate second ANN output will
PSL determine the optimum number of reduced input feature
sets.

The NNWE method uses a simple approach to find the optimal

number of reduced input feature sets for the second ANN. From the
x1 optimal number of reduced input features, the dimension of the
original input features can be reduced significantly, thus, speeding
up the training and convergence of the ANN for vulnerability
Matrix of weights W assessment of power systems.

Original 4.4. Performance evaluation

training data Set of reduced features R Input Layer
p The performance of the ANN is measured according to the clas-
sification rate (CR) (Faraoun & Boukelif, 2006) which is given by the
following expression:
Hidden Layer
 
C
CR ¼ 100 ð8Þ
U
where C is the number of testing sets classified correctly and U is
Output Layer the total number of testing data sets.
As for the accuracy of the testing sample, it can be gauged by
calculating absolute error (AE) of the actual and desired test data.
PSL
Absolute error is defined as,
The second ANN for vulnerability assessment
AEn ¼ DOn  AOn ð9Þ
Fig. 4. ANN structures used in the proposed NNWE method and vulnerability where n is the test data number, DO is the desired output obtained
assessment.
from VA, and AO is the actual output obtained from ANN.

5. Neuro-fuzzy implementation in load shedding

2 3
v 1 ¼ x1 w11 þ x2 w21 þ x3 w31 þ x4 w41 þ    þ xn wn1
6v ¼ x w þ x w þ x w þ x w þ  þ x w 7 During steady state operation of a power system, the power bal-
6 2 1 12 2 22 3 32 4 42 n n2 7
6 7 ð5Þ ance is always maintained. However, such balance of power may
4 : : : : : 5
be disturbed by sudden changes in load or loss of generation. If
v m ¼ x1 w1m þ x2 w2m þ x3 w3m þ x4 w4m þ . . . þ xn wnm at any stage, it is found that the current operating state of a power
where v m is the value of hidden neuron and xn is the input system is insecure where some or all of the system constraints are
variable. violated, then fast corrective actions need to be taken so as to bring
The weights matrix w extracted from (5) for different the system back to a secure operating state. Initially, generators are
number of hidden neurons is given as, re-dispatched optimally in a manner so as to satisfy the generation
2 3 security constraints. If re-scheduling of generators fails to provide
w11 w21 w31 w41 . . . wn1 a feasible solution for secure operation, load shedding is restored.
6w w22 w32 w42 . . . wn2 7
6 12 7 It is noted that load shedding is only used as a last resort so as
w¼ 6 7 ð6Þ
4 : : : : : 5 to avoid a complete system failure.
w1m w2m w3m w4m . . . wnm The objective of solving the load shedding problem is to mini-
mize the difference between the generated power at base case
and the generated power at contingency case. The difference can
v. Using the weights obtained by Eq. (6) with the original num- be expressed mathematically as:
ber of input variables, determine the values of the reduced
feature sets by using the following equation: DS ¼ SCC  SBC ð10Þ
where SCC is the apparent generated power at contingency case and
2 3 SBC is the apparent generated power at base case.
2 3 x11 . . .
w11 w21 w31 w41 . . . wn1
6 7 The difference between the generated power at contingency
6w 6 x21 . . . 7
6 12 w22 w32 w42 . . . wn2 7
7 6 7 case and the generated power at base case, DS, will give the
R ¼ wp ¼ 6 76
6 x31 . . . 7
7 amount of load to be shed so that a power system can remain in
4 : : : : : 5 6 7
4 x41 . . . 5 a secure state. This statement implies that:
w1m w2m w3m w4m . . . wnm
xn1 . . . DS ¼ SNFLC ð11Þ
ð7Þ
where SNFLC is the amount of load shed estimated by a controller.
5028 A.M.A. Haidar et al. / Expert Systems with Applications 37 (2010) 5023–5031

Degree of membership
0.5

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Input variable PSL

1
Degree of membership

0.8

0.6

0.4

0.2

0
0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
Input variable V

Fig. 5. Initial membership functions of the input variables.

In this study, a Takagi–Sugeno FIS is adapted to the ANFIS as it is checking data are presented to the ANFIS model, it is considered
more effective for system identification. For load shedding estima- that the model has parameters associated with minimum error.
tion using neuro-fuzzy logic controller (NFLC), the ANFIS output is
associated with the amount of load shed in MVA p.u., the inputs to 6. Results and discussion
the NFLC are the PSL and lowest voltage magnitude. Fig. 5 shows
the initial membership functions of the input variables in which From the total datasets generated by simulations using Power
the membership function parameters are selected to satisfy the de- System Analysis and Toolbox (Milano, 2006), testing datasets have
sired output. A multi-layer feed forward neural network trained by deliberately been chosen outside the region of the training or
using the back propagation algorithm is used to adjust the mem- learning datasets so as to test the generalization and extrapolation
bership function parameters according to the input–output charac- capability of the neural network after learning. The development of
teristic of the training patterns. The neural network computation
time depends on the number of rules, which on the other hand de-
pend euphonically on the number of the membership function and Table 1
inputs. The parameters associated with the membership function GRNN vulnerability assessment results for the 87 bus test system.

can change through the training process. The adjustment of these Contingency GRNN PSL Desired PSL Absolute error
parameters is facilitated by a gradient vector, which provides a LI-10% 0.538 0.527 0.012
measure of how fuzzy inference system models the input/output LI-12% 0.539 0.401 0.138
relations. Once the gradient vector is obtained, any of the optimi- LI-15% 0.214 0.269 0.054
zation routines can be applied to adjust the parameters so as to LI-18% 0.214 0.193 0.021
LI-20% 0.133 0.159 0.026
reduce an error usually defined by the sum of the squared differ- LI-30% 0.075 0.069 0.006
ence between actual and desired outputs. The overall output of LI-32% 0.05 0.059 0.009
the ANFIS is the estimated amount of load shed by NFLC which LI-33% 0.05 0.054 0.005
calculates the sum of outputs of all defuzzification neurons and LI-35% 0.046 0.046 0.0
GO-2158 0.805 0.872 0.067
is given by,
GO-2182 0.85 0.929 0.08
GO-2298 0.409 0.38 0.029
X
n
GO-2306 0.803 0.819 0.016
SNFLC ¼ l i ðki0 þ ki1 x1 þ ki2 x2 Þ ð12Þ GO-2308 0.538 0.672 0.134
i¼1
GO-2182, 2298 0.247 0.257 0.01
GO-2182, 2298, 2436 0.114 0.062 0.052
where ki0 ; ki1 and ki2 is the sets of consequent parameters of rule i GO-2182, 2298, 2552, 2740 0.114 0.087 0.028
and barli is the normalized firing strength. GO-2182, 2298, 2684 0.114 0.122 0.008
After modeling the ANFIS, it is validated with a set of testing GO-2510, 2511 0.284 0.299 0.015
GO-2510, 2511, 2158 0.283 0.287 0.003
data and checking data set. The reason for using a checking data GO-2510, 2511, 2158, 2306 0.284 0.259 0.024
set for model validation is to avoid the point in the training process LO-21, 22 0.418 0.436 0.018
when the model begins to over fit the training data. In principle, LO-21, 22, 33, 34 0.085 0.098 0.013
the model error for checking data set tends to decrease during LO-21, 22, 81, 82 0.085 0.098 0.012
LO-23, 24 0.648 0.534 0.113
training until the point when over fitting begins. This is when
The average absolute error 0.038
the model error for the checking data suddenly increases. When
 A.M.A. Haidar et al. / Expert Systems with Applications 37 (2010) 5023–5031 5029

AI for vulnerability assessment and control of the test system is PSL. The desired or actual PSL values are obtained from vulnerabil-
implemented using Matlab software version 7 on an Intel Pentium ity analysis simulations. From the testing results, it is noted that in
2.13 GHz with 496 Mb of RAM. The ANN and neuro-fuzzy coded terms of absolute errors, the GRNN is considered to assess vulner-
originally in Matlab were utilized and modified to suit the vulner- ability of the test power system accurately in which the average
ability assessment data. There are 234 input features selected for absolute error is 0.038.
each set of data comprising of real and reactive power flows and The criteria for determining system vulnerability is based on
total generated real and reactive powers. By using the proposed the vulnerability index calculated at base case in which a system
feature extraction method, the extracted features are reduced to is said to be invulnerable if the PSL value is close to 1.0. Referring
32 which are 15% of the original features. to Table 1 and Fig. 6, it is noted that some of the contingencies give
The results of vulnerability assessment on a practical 87 bus comparatively low values of PSL. These contingencies are recog-
test system using GRNN are tabulated as shown in Table 1 and nized by PSL indices as vulnerable because the indices values are
graphically shown in Fig. 6 for some contingency cases only. The close to 0 and such values have been classified as causing the sys-
GRNN testing results are evaluated in terms of absolute error, tem vulnerable. If such contingencies occur, the system is said to
which is the difference between the desired PSL and the GRNN be vulnerable and may cause interruption of power supply.

LO-23,24 Absolute Error

Desired PSL
LO-21,22,81,82
GRNN PSL
LO-21,22,33,34

LO-21,22

GO-2510,2511,2158,2306

GO-2510,2511,2158

GO-2510,2511

GO-2182,2298,2684

GO-2182,2298,2552,2740

GO-2182,2298,2436

GO-2182,2298
Contengiences

GO-2308

GO-2306

GO-2298

GO-2182

GO-2158

LI-35%

LI-33%

LI-32%

LI-30%

LI-20%

LI-18%

LI-15%

LI-12%

LI-10%

0 0.2 0.4 0.6 0.8 1

PSL

Fig. 6. GRNN vulnerability assessment results for the 87 bus system.

5030 A.M.A. Haidar et al. / Expert Systems with Applications 37 (2010) 5023–5031

Degree of membership
1

0.5

0
0 0.2 0.4 0.6 0.8 1
Input variable PSL
Degree of membership

0.5

0
0.7 0.75 0.8 0.85 0.9 0.95 1
Input variable V

Training data
Output variable LS

ANFIS output
200

-200
1
0.9 0.8 1
0.8 0.4 0.6
0.7 0 0.2
Input variable V Input variable PSL
Fig. 7. Final membership functions of the input variables.

Table 2
Results of load shedding using NFLC for the 87 bus system.

Contingency case Amount of load shed Amount of load shed Ranking of weak buses
MVA p.u. (NFLC) MVA p.u. (Simulation)
LI-20% 40.887 40.960 2130, 2196, 2326, 2412, 2754
LI-30% 82.814 83.057 2016, 2128, 2130, 2142, 2166, 2170, 2196, 2220, 2248, 2250,
2276, 2326, 2396, 2412, 2470, 2636, 2674, 2722, 2754, 2814, 2880
GO-2182, 2298, 2436 40.362 40.232 2130, 2196, 2250, 2276, 2326, 2396, 2412, 2754
GO-2182, 2298, 2552, 2740 24.124 24.141 2130, 2250, 2339, 2396, 2754
GO-2182, 2298, 2684 14.528 14.537 2130, 2338, 2340, 2396, 2554, 2684, 2698, 2754, 72, 73
GO-2510, 2511, 2158, 2306 2.868 2.688 2196, 2412
LO-21, 22, 33, 34 26.125 26.377 2250, 2334, 2339, 2880, 2952
LO-21, 22, 81, 82 26.752 26.482 2250, 2334, 2339, 2880, 2952

As for vulnerability control results using the ANFIS, the results the voltage magnitudes are violated. After determining the weak
are presented in terms of the ANFIS output which is the amount buses, it is considered that load shedding is to be applied only at
of load shed and the final membership functions of the input vari- these buses. The determination of the weak buses can assist sys-
ables obtained after training the ANFIS as shown in Fig. 7. The re- tem operators in determining the appropriate load shedding loca-
sults of load shedding using the NFLC for various contingency cases tion. Referring to Table 2, the critical contingency that causes
are summarized as shown in Table 2. From the table, it is shown greater amount of load shed is due to multiple outages of genera-
that for each contingency case, the amount of load to be shed is tors 2182, 2298 and 2436 (GO-2182, 2298, 2436) and increase in
determined by NFLC in terms of per unit MVA. Comparing the re- total load LI-20% and LI-30%.
sults NFLC and simulation in terms of the amount of load shed, it
can be concluded that the NFLC can accurately determine the
amount of load shed, in which the accuracy of NFLC is in the range 7. Conclusion and future works
of 93–99%.
An important factor to consider in load shedding is to determine This research presented the application of AI for vulnerability
the location, where load is to be shed. Therefore the weak buses are assessment and control of Malaysian practical power system. Neu-
identified by considering buses with low voltage magnitudes in the ral network weight extraction method is used as the feature
range of (0.94–0.83) per unit. At these weak buses, it is noted that extraction method and the generalized regression neural network
 A.M.A. Haidar et al. / Expert Systems with Applications 37 (2010) 5023–5031 5031

to predict vulnerability of a power system. Based on the vulnerabil- Joorabian, M., Taleghani, A., & Aggarwal, R. (2004). Accurate fault locator for EHV
transmission lines based on radial basis function neural networks. Journal of
ity index in terms of power system loss, vulnerability of a power
Electric Power Systems Research(71), 195–202.
system can be assessed, and the vulnerabilities are controlled using Kim, M., 2002. Application of computational intelligence to power system
load shedding technique. The techniques proposed in this paper vulnerability assessment and adaptive protection using high speed
provide fast, cost effective and accurate methods which can be communication. Ph.D Thesis, University of Washington, USA.
Kim, M., El-Sharkawi, M.A., & Marks, R.J., 2005. Vulnerability indices for power
implemented for real time applications. This was confirmed by system. In Proceedings of the IEEE 13th international conference on intelligent
the results obtained in this research of which sample results are gi- systems application to power systems (pp. 335–341).
ven in this paper. Feature studies are intended to incorporate the Kim, B., Junki, B., & Byung, L. (2006). Modeling of silicon oxynitride etch
microtrenching using genetic algorithm and neural Network. Journal of
developed vulnerability assessment and control techniques with Microelectronic Engineering, 83, 513–519.
the supervisory control and data acquisition (SCADA) system and Manana, D., Toader, M., & Anatoli, P. (2004). Fuzzy logic in power system
energy management system so as to enable on-line vulnerability performability. Proceeding of the Second IEEE International Conference on
Intelligent System, 326–330.
assessment and control of modern power systems. Milano, F. (2006). An open source power system analysis toolbox. IEEE Transaction
on Power System(3), 1199–1206.
References Mori, H. (2006). State-of-the-art overview on data mining in power systems.
Proceeding of the IEEE Conference and Exposition (PSCA), 1–5.
Negnevitsky, M. (2002). Artificial Intelligence – A guide to intelligent system (1st ed.).
Al-Osaimi, S., Abdennour, A., & Abdullaziz, A. (2005). Hardware implementation of a
Addison Wesley (Chapters 4 and 8).
fuzzy logic stabilizer on a laboratory scale power system. Journal of Electric
Sallam, A., & Khafaga, A. (2002). Fuzzy expert system using load shedding for
Power Systems Research, 74, 9–15.
voltage instability control. Proceedings of the IEEE Large Engineering Systems
Dash, P., Mishra, S., & Panda, G. (2000). A radial basis function neural network
Conference on Power Engineering, 125–132.
controller for UPFC. IEEE Transactions on Power Systems(15), 1293–1299.
Sawhney, H., & Jeyasurya, B. (2004). On-line transient stability assessment using
Dash, P., Pradhan, A., & Pandaa, G. (2001). Application of minimal radial basis
artificial neural network. Proceedings of the IEEE on Large Engineering Systems
function neural network to distance protection. IEEE Transactions on Power
Conference, 76–80.
Delivery, 16, 68–74.
Simon, H. (1999). Neural network: A comprehensive foundation. Canada: Prentice
Faraoun, K. M., & Boukelif, A. (2006). Neural networks learning improvement using
Hall.
the K-means clustering algorithm to detect network intrusions. International
Simon, P., 2005. Oscillatory stability assessment of power system using
Journal of Computational Intelligence(3), 161–168.
computational intelligence. Ph.D Thesis, Universit Duishdurg-Essen, Germany.
Franco, C. et al. (2001). Operation and management of the electric system for
Song, H., & Kezunovic, M., 2005. Static security analysis based on vulnerability index
industrial plants: An expert system prototype for load-shedding operator
(VI) and network contribution factor (NCF) method. In IEEE/PES transaction on
assistance. IEEE Transaction on Industry Application, 37, 701–708.
transmission and distribution conference and exposition Asia-Pacific Dalian, China
Jensen, C. A., El-Sharkawi, M. A., & Marks, R. J. (2001). Power system security
(pp. 1–7).
assessment using neural networks: Feature selection using Fisher
discrimination. IEEE Transaction on Power System, 16(1), 757–763.