You are on page 1of 17

Electrical Power and Energy Systems 64 (2015) 7187

Contents lists available at ScienceDirect

Electrical Power and Energy Systems


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

A new power system transient stability assessment method based


on Type-2 fuzzy neural network estimation
Amir Sharian, Saeed Sharian
Department of Electrical Engineering, Amirkabir University of Technology, P.O. Box 15914, Tehran, Iran

a r t i c l e

i n f o

Article history:
Received 10 July 2013
Received in revised form 1 July 2014
Accepted 8 July 2014
Available online 1 August 2014
Keywords:
Type-2 fuzzy neural network
Transient stability assessment
Critical clearing time
Type-2 fuzzy system
Multilayer perceptron neural network
MLP NN-based sensitivity analysis

a b s t r a c t
Transient stability assessment (TSA) of large power systems by the conventional method is a time consuming task. For each disturbance many nonlinear equations should be solved that makes the problem
too complex and will lead to delayed decisions in providing the necessary control signals for controlling
the system. Nowadays new methods which are devise articial intelligence techniques are frequently
used for TSA problem instead of traditional methods. Unfortunately these methods are suffering from
uncertainty in input measurements. Therefore, there is a necessity to develop a reliable and fast online
TSA to analyze the stability status of power systems when exposed to credible disturbances. We propose
a direct method based on Type-2 fuzzy neural network for TSA problem. The Type-2 fuzzy logic can properly handle the uncertainty which is exist in the measurement of power system parameters. On the other
hand a multilayer perceptron (MLP) neural network (NN) has expert knowledge and learning capability.
The proposed hybrid method combines both of these capabilities to achieve an accurate estimation of
critical clearing time (CCT). The CCT is an index of TSA in power systems. The Type-2 fuzzy NN is trained
by fast resilient back-propagation algorithm. Also, in order to the proposed approach become scalable in a
large power system, a NN based sensitivity analysis method is employed to select more effective input
data. Moreover, In order to verify the performance of the proposed Type-2 fuzzy NN based method, it
has been compared with a MLP NN method. Both of the methods are applied to the IEEE standard
New England 10-machine 39-bus test system. The simulation results show the effectiveness of the
proposed method in compare to the frequently used MLP NN based method in terms of accuracy and
computational cost of CCT estimation for sample fault scenarios.
2014 Elsevier Ltd. All rights reserved.

Introduction
Nowadays, the continues trend to increase in load demands
along with economic and environmental constraints for building
new power plants and transmission lines, have lead power systems
to operate closer to their limits which increases the occurrence
probability of transient stability problem [1,2].
The analysis and methods that are used to determine if a system
is safe or unsafe (based on pre-established criteria) is typically
referred as power system security assessment. An electric power
system might have many changes in the system operating
conditions or conguration; therefore, planning phase transient
stability studies, would not be reliable for an operational system,
so continuous system analysis is necessary for operators to take
Corresponding author.
E-mail addresses: Amir.sharifyan@gmail.com (A. Sharian), sharian_s@aut.ac.ir
(S. Sharian).
http://dx.doi.org/10.1016/j.ijepes.2014.07.007
0142-0615/ 2014 Elsevier Ltd. All rights reserved.

proper preventative control actions if insecure system conditions


occurred.
The primary objective of transient stability analysis (TSA) in a
power system is to determine the capability of power system to
remain in stable and safe operating condition when a large disturbance such as severe lightning strike, loss of heavily loaded transmission line, loss of generation station, or short circuit on buses [3]
inuences the system. CCT is a well-known indicator that can be
used to measure power system transient stability. The CCT is the
maximum time duration by which the disturbance may act on
the power system without losing its capability to recover to a
steady-state (stable) operation.
We can broadly classify security analysis depending on modeling and used technique into static and dynamic category [3,4]. Static security assessment is related to an equilibrium point of system,
where voltage and thermal limits are observed. Generally static
security assessment is done using computational tools based on
load ow algorithms. The contingencies events must be considered

72

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187

to ensure an acceptable steady-state condition, even if one element


of the system is lost.
Evaluation of the ability of a power system to withstand a nite
set of contingencies and to survive the transition to an acceptable
steady-state condition is dened as dynamic security assessment
(DSA) [4]. As illustrated in Fig. 1, DSA consists of three main categories: rotor angle stability, voltage stability and frequency stability. Also the rotor angle stability is divided into two sub categories
which are small signal stability and transient stability [3,5]. In this
paper we focus on TSA which involves the evaluation of the ability
of a power system to maintain synchronism under severe but credible contingencies. The DSA studies are usually conducted in a time
range between 3 and 5 s for small power systems. For large systems with dominant inter-area swings this time may extend to
10 s [5].
Two main categories of TSA methods are time domain simulation (or numerical integration) method and direct method. Currently, the widely used method by power utilities and most
accurate method for TSA is time domain simulation method [5,6].
This method is implemented by solving the differential equations
of power network while the direct method involves in calculation
of the transient energy margins which shows the system stability
limits. This method gives an accurate information about state variables and can be applied to any level of detail of power system
models [1,4,7]. In Ref. [37] the concept of lyapunov exponents
(LEs) is used to analyze the transient stability of power systems.
Also in Ref. [40]; a stochastic-based approach to evaluate the probabilistic transient stability index of the power system incorporating the wind farm is proposed.
However, use of such a method requires numerical solution to
nonlinear equations of system which has high online computation
cost and involves intensive and time-consuming numerical
integration efforts. Also, the difculty of designing good energy
functions for multi-machine power systems may lead to computational inefciency and inaccuracy [5,6]. So, it does not provide
information regarding the degree of stability and the degree of
instability in a power system.
In addition, TSA of large sized power systems has become a very
complex process due to the exponential expansion of complexity in
power system topology. For each disturbance many nonlinear
equations should be solved that makes the problem too complex
and will lead to delayed decisions in providing the necessary control signals for controlling the system. Therefore, there is a necessity to develop a reliable and fast online TSA to analyze the stability
status of a power system when exposed to credible disturbances.
On the other hand, direct method techniques require less online
computation efforts and can provide a quantitative measure of the
degree of system stability, but it has some challenges and limitations involved in the practical applications for power system TSA

[5]. In recent years, machine learning and computational intelligence techniques, such as articial neural networks (ANNs), have
been proposed as promising approaches to solve some complex
power system protection and control problems instead of simulating the power system equations for TSA in power systems [5,6,8
17]. These approaches can quickly obtain a nonlinear mapping
relationship between the input data and the output and can
approximate solutions of power systems differential equations
[6]. There are two ways in using ANN for power system TSA, one
way is using ANN as a regression function to predict transient stability degree[813], such as CCT and system stability margin;
another way is using the ANN as a classier to directly classify
the system into either stable or unstable states [14]. There are
many different types of NN such as MLP NN and radial basis function (RBF) NN which can be used in different applications.
The feed-forward NN, also best known as MLP NN, was the rst
and most simple type of NN devised. It was developed in early
1970s and is the most popular topology in use today. This NN consists of an input layer, an output layer, and one or more hidden layers. In this NN the information only moves in forward direction.
Data ows into the NN through the input layer, passes through
the hidden layer and nally ows out of the NN through the output
layer. There are no cycles or loops in the network. These networks
can be constructed from different types of units such as binary
McCulloch-Pitts neurons. But frequently are devised as continuous
neurons, with sigmoidal activation function in the context of back
propagation of error. The MLP NN can be considered as simple
interpolation of inputoutput model, with NN weights as free
parameters. Such NN conguration can model functions of almost
any arbitrary complexity. The function complexity is determined
with the number of layers and the number of neurons in each
layer.
Another frequently used NN in the literatures is RBF NN [15,16].
RBF NN is powerful method for interpolation in multidimensional
space. The RBF can be replaced by the sigmoidal hidden layer in
MLP NN. The structure of the RBF NN consists of three layers
namely, the input layer, the hidden (or RBF) layer, and the output
layer. The nodes within each layer are fully connected to the previous layer. The input nodes are directly connected to the hidden
layer neurons. Usually a Gaussian function is used in each node
in RBF layer to determine distance of inputs with respect to the
mean of the Gaussian function. A linear combination of hidden
layer values that represents mean predicted output is generated
in the output layer when RBF NN is used in regression problems.
When RBF NN is used in the classication problems, the output
layer is representing a posterior probability. The output is typically
a sigmoid function of a linear combination of RBF layer values.
In RBF NN each input datum is associated with a RBF kernel
function such as support vector machine method. In this approach

Fig. 1. Taxonomy of power system stability methods.

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187

a non-linear kernel function is used to project the input data into


another space. Where in the new space the learning problem can
be solved using a linear model. RBF NN is typically trained by maximum likelihood framework. RBF NN is outperformed in most classication applications by SVM. But they can be competitive in
regression applications when the dimensionality of the input space
is relatively small [15].
There are many previous papers which are using NN based
direct methods to predict the CCT. Sobajic and Pao [8] were used
NNs to predict the CCT for a small test power system. Pao and Sobajic [9] used both unsupervised and supervised learning for the TSA
problem. In Ref. [10], Aboytes and Ramirez used NNs to predict the
stability of generators system. In Ref. [11], Bahbah and Girgis proposed a recurrent RBF NN and a MLP NNs to model system dynamics and the generators angles and angular velocities are predicted
to solve TSA problem. In Ref. [17], Sawhney and Jeyasurya
employed NNs to predict a transient stability index, which was
obtained by the extended equal area criterion method.
In Ref. [18], a procedure has been described for extracting
rules from a trained MLP NN for reasoning power systems CCT.
In Ref. [19], a generalized regression NNs based classication
method has been proposed for transient stability evaluation in
power systems. Also in Refs. [20,21], two methods are introduced
TSA which are based on adaptive resonance theory (ART) NNs. To
reduce the number of calculations and online computational time
in Ref. [22], the major portion of power system transient stabilitys mathematical calculations has been replaced by an estimation procedure. In Ref. [6], a MLP NN and a RBF NN are used to
estimate the CCT as an index for power systems TSA according
to the classical models. In Ref. [12], two MLP NNs are used to estimate the CCT and a transient stability time margin. Only a particular fault scenario is considered and the detail models of
synchronous machines are presented. In Ref. [39] classication
and regression tree (CART) based power system transient stability
preventive controls are proposed and the results are compared
with MLP based method. The nding shows that preventive controls developed by both approaches think alike. They are even
complementary.
Although ANN is the most popular computational intelligence
method to classify patterns, it requires a complicated design procedure and an extensive training process. Moreover, ANN is good in
interpolation but not so good in extrapolation which decreases
its generalization ability [23]. MLP NN is a classic solution for
CCT function approximation problem. Its accuracy can be increased
by addition of nodes and hidden layers.
Also RBF NN is shining in the problems that input parameters
can be classied into clusters (input parameters are correlated).
Another advantage of RBF NN is that the hidden layer is easier to
interpret than the hidden layer in a MLP NN. Although the RBF
NN has fast training capability, when the training is nished and
it is being used for testing; it is slower than a MLP NN [6]. So where
speed is a factor MLP NN may be more appropriate. MLP NN suffering from local minima but RBF NN is not. This is because only the
linear mapping from hidden layer to output layer is adjusted in the
learning process. Linearity ensures that the error surface is quadratic, therefore it has a single minimum. The MLP NN requires
an iterative procedure to compute the network weights. But, the
RBF NN requires an iterative procedure for clustering the data to
determine the number of nodes in RBF layer. A disadvantage for
RBF NN is that the radial basis functions should have a good coverage of the input space.
Simulation results using the New England test power system in
Ref. [6] indicates that both of NNs (RBF and MLP) could be
employed to estimate the CCT with a good degree of accuracy.
However, better test results were obtained using the MLP NN. Both
of the NNs based solutions (MLP and RBF) can accurately estimate

73

CCT when the network inputs fallow exactly the training patterns.
On the other hand when an input data is noisy or has uncertain
value; both the MLP NN and the RBF NN failed to estimate an
accurate CCT due to the sensitivity of NN based methods to
network inputs. By considering these drawbacks, it becomes
necessary to devise a more robust solution for TSA problem in
power systems.
The Type-2 fuzzy sets have been introduced as an expansion of
the type-1 fuzzy sets by Zadeh [24]. The Type-2 fuzzy logic systems
can handle uncertainties which are associated to information in
the knowledge base of the process. The Type-2 fuzzy sets have various applications in solving many problems in the power system
area [2527]. Recently neuro-fuzzy systems have been used in
many areas of science and engineering [2833]. A Type-2 fuzzy
NN combines the learning capability of NNs and the linguistic
interpretation feature of fuzzy classier to solve various problems
such as predication, control and identication [3035]. A major
problem in adaptive fuzzy system is that its complexity is exponentially increased by the number of inputs to the network. So
many efforts have been done to reduce the number of inputs to
these networks. In Ref. [36], an adaptive fuzzy classication technique is used with normal fuzzy technique to solve the power systems TSA problem. The results are demonstrating the advantages
of using adaptive fuzzy technique. Moreover, a NN and a principal
component analysis (PCA) method are employed to reduce the
number of inputs by sensitivity analysis technique. In Ref. [14], a
neuro-fuzzy system is applied for power system DSA focusing on
the transient stability. The power system security state is classied
by the neuro-fuzzy system into three categories named as
secure, doubtful security and insecure. In Ref. [38]; a binary
SVM classier with combinatorial trajectories as inputs was
trained to predict the transient stability status.
None of the previous works did not address the uncertainty and
noisy nature of power system measured data which are used as
inputs to TSA system [46,814,1622,3543]. In this paper we
proposed a Type-2 fuzzy NN to address the uncertainty which is
exist in inputs. The Type-2 fuzzy layer converts uncertain and
noisy inputs to more dependable and reliable linguistic variables
which are used as inputs to the MLP NN layer. The Type-2 fuzzy
NN methodology is used to solve the on-line power system TSA
problem for a set of particular fault scenarios (contingencies)
under different system operating conditions.
In the proposed method the Type-2 fuzzy NN is trained to provide the CCT, as a measure of the power system transient stability.
The Type-2 fuzzy system is used to cope with uncertainty of the
power system model and measurements of system operating
parameters. In addition, in order to provide a scalable solution
for a large power system, the proposed approach uses a NN based
sensitivity analysis method [17,43] to reduce the number of inputs
to the Type-2 fuzzy NN, so the complexity of system decreased and
calculation time become shorter enough to convert the proposed
method to a feasible solution. It should be notied that the sensitivity analysis and training procedure are conducted ofine but the
estimation procedure is executed in online manner.
To evaluate the efciency of proposed method, it applied to TSA
of sample fault scenarios in the standard New England 10-machine
39-bus test system [7]. The simulation results show the effectiveness of the proposed method in terms of accuracy and online computation time. In summary, the main contributions of the paper are
as follows:
a. We propose a Type-2 fuzzy NN based method to accurately
estimate the CCT as an index of power system transient stability. The proposed method considers the uncertainty of
power system model and measurements of system operating parameters.

74

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187

b. A NN based sensitivity analysis method is employed to


reduce the number of inputs. So only the most effective
inputs are considered and computational costs are reduced.
c. We conduct several simulations with different operating
conditions prior to fault scenario and compare the accuracy
of estimated CCT by the proposed Type-2 fuzzy NN method
and other popular relevant method (MLP NN) with actual
CCT to demonstrate the effectiveness of the proposed
method.
The rest of paper is organized as follows: Section Power system
describes power system model and test system. In Section Methodology for power system transient stability assessment the
proposed Type-2 fuzzy NN and sensitivity analysis methodology
for CCT prediction are presented. In Section Results and discussion
we present results and discussion of power system TSA evaluation
for sample fault scenarios by the proposed method and compare
it with MLP NN. Finally, conclusions are outlined in Section
Conclusion.

Test system
In this paper an IEEE standard New England 10-machine 39-bus
test system is used for TSA as shown in Fig. 2. We consider that bus
1 in the New England test system is a slack bus whose voltage
angle and voltage magnitude values are xed at known values.
The generated active and reactive powers of the slack bus are
denoted by PG1 and QG1. The remaining generator buses (i.e., buses
210) are assumed as PV buses and their voltage magnitude and
generated active and reactive powers in the PV buses are denoted
by Vi and PGi and QGi (i = 2,3,. . .,10) respectively. In addition the
remaining nine PV buses and other slack buses in the test system
have 29 PQ buses (i.e., buses 1139) whose active and reactive
loads are denoted by PDj and QDj (j is the bus number), even
though the loads are acting only on 19 distinct buses. System
details conguration and data are given in Ref. [7].
In this paper we assume the pre-fault system operating conditions as applied in Refs. [6,12,13] which is used as input to the propose Type-2 fuzzy NN to estimate the CCT. The inputs are
considered as follows:

Power system
Many transient disturbances can occur in a power system such
as: loss of generation, faults, loss of load, and loss of system components such as transformers or transmission lines [3]. Assessment
of the rotor swing angles can be used to determine stability or
instability condition of a power system due to transient disturbance. Following a transient disturbance, if the relative generator
rotor angles in the system remain in synchronism with each other,
we conclude that a power system is stable. On the other hand,
when the relative generator rotor angles go out of step and lost
its synchronism a power system is considered as unstable. In this
section rst we give an introduction to the classical model of
power system and then describe our standard test system.
Power system model
Assume a power system consists of n synchronous generators.
The classical model of system and the internal center of inertia
(COI) [7,44] can be formulated by Eqs. (1)(9) as follows:

dhi
~i
x
dt
Mi

i 1; 2; . . . ; n

~i
dx
Mi
Pmi  P ei 
PCOI
dt
MT

Pei

n
X

C ij sin hij Dij cos hij E2i Gii

1
2

j1i

PCOI

n
X
Pmi  Pei

i1

MT

n
X
Mi

i1

C ij Ei Ej Bij

Dij Ei Ej Gij

Y ij Gij jBij

hij hi  hj

The notations of parameters in system Eqs. (1)(9) are provided


in Table 1.

Voltage magnitudes of all the 9 PV buses (V2V10).


Generated active powers of all the 9 PV buses (PG2PG10).
Active load powers of all the 19 loads acting on different buses
(PD1, PD2,. . ., PD37).
Reactive load powers of all the 19 loads acting on different
buses (QD1, QD2,. . ., QD37).
We conduct a sample TSA on the New England standard test
system with a set of four different fault scenarios. Scenario 1 is a
three phases to ground fault injected to bus 32 and is cleared by
removing the line connected between bus 32 and bus 31. Scenario
2 is a three phases to ground fault injected to bus 14 and is cleared
by removing the line connected between bus 14 and bus 33. Also
scenario 3 is a three phases to ground fault injected to bus 17
and is cleared without removing any line in the post fault systems.
And nally scenario 4 is a three phases to ground fault injected to
bus 34 and is cleared by removing the line connected between bus
34 and bus 35. The faults locations are shown in Fig. 2.
Here, its assumed that the occurrence probabilities of the sample fault scenarios are high and these fault scenarios are required
to have on-line TSA. Although only four fault scenarios are considered for verication, the proposed method is completely general
and it can be used for any fault scenario that applied at any location of the test system. Also other test systems can be used in similar way [6,12].
It should be noted that the operating conditions of the test system are changed here without compelling any limits on maximum
and/or minimum generated reactive powers of the all generators
(i.e., QG-limits are not taken into account), leading us to use above
mentioned independent system operating conditions as inputs to
the Type-2 fuzzy NN. However, other system variables can be
applied as inputs to the Type-2 fuzzy NN. To consider the violations of generated reactive power of system generators, that may
take place in practice, and also to include extra information regarding the system total loading conditions, In this paper same as Ref.
[12], we have applied the following additional inputs for the Type2 fuzzy neural network:
Generated active power of the slack bus (PG1).
Generated reactive power of the slack bus (QG1).
Generated reactive powers of all the 9 PV buses (QG2QG10).
These operating conditions are obtained by performing an AC
load-ow analysis on power system prior to fault. The New England test system consists of nine PV buses and 19 loads. Hence,

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187


Table 1
System equations parameter notations.
Parameter

Description

hi
~i
x
Mi
Pmi
Ei
Yij
Gij

Rotor angle in reference to the COI


Angular velocity of rotor in reference to the COI
Inertia constant of the ith generator
Mechanical input power of the ith generator
Internal generator voltage magnitude for the ith generator
The ijth elements of the reduced system admittance matrix
Conductance of the ijth elements of reduced system admittance
matrix
Susceptance of the ijth elements of reduced system admittance
matrix
Generated active power of the slack bus
Generated reactive power of the slack bus
Generated active power of the ith PV bus
Generated reactive power of the ith PV bus
Active load of the jth bus
Reactive load of the jth bus
The bus number
Index of generators
Number of generators in system
Electrical output power of the ith generator
COI accelerating power

Bij
PGi
QG1
PGi
QGi
PDj
QDj
j
i
n
Pei
PCOI

Fig. 2. Diagram of the New England standard test system.

for each fault scenarios, the CCT is function of the


(9 + 9 + 19 + 19 + 1 + 1 + 9 = 67) operating conditions that are used
as inputs to the Type-2 fuzzy NN.
Methodology for power system transient stability assessment
In this section, we present our proposed Type-2 fuzzy NN based
approaches to estimate the CCT as an index for power systems TSA.
First a general overview of entire system is presented, next we give
a short introduction of fuzzy neural systems that nally leads us to
the proposed Type-2 fuzzy NN. In the next sub section training and
prediction of CCT procedures are described. Finally MLP NN based
sensitivity analysis methodology is presented.
Overview of Type-2 fuzzy NN for CCT prediction
In this section we are going to describe a general overview of
power system TSA infrastructure as shown in Fig. 3. With the
development of wide area measurement system, the operational
condition of a power system can be directly measured by
geographically distributed phasor measurement units (PMUs)
before and following a disturbance. All the power system operating
conditions are transferred to an energy control center (ECC). The

75

aggregated operating conditions are easily accessible in ECC. Also


this historical data are stored in log data base for further ofine
power system analysis. These measured operating conditions contain all the real time information of the system, such as models,
arguments and disturbance. How to predict the transient stability
status post-disturbance using these measured operational condition is an important research topic in power system stability
assessment that is the focus of this paper.
According to the Ref. [12], different fault scenario causes to different CCT value. Moreover, the systems operating conditions
prior to the fault and also the power system network topologies
both have signicant effect on the CCT. Hence, the CCT is a nonlinear and complex function of the system conguration after and
before fault occurrence, system operating conditions prior to fault
and also fault type and its location in the network.
Many different source of fault can be considered in a power system such as loss of load, loss of generation and loss of system components such as transformers or transmission lines. These sources
have different level of importance and different probability of
occurrence. So we should calculate CCT for each fault scenario separately. The anomaly detection block (Fig. 3) is an online system
that inspects the measured operational conditions to nd anomaly
patterns related to each fault scenario. After detection of a fault,
the anomaly detection block reports fault scenario index to scenario selection block.
Two different NNs are associated with each fault scenario index
which are activated by the scenario selection block. One of them is
a Type-2 fuzzy NN that online approximates the CCT nonlinear
function instead of solving power system differential equations.
Since the CCT is only the function of system operating conditions
prior to fault [6,8], the proposed Type-2 fuzzy NN is used to
approximate this complex function. The proposed Type-2 neuro
fuzzy method can estimate CCT for each fault with an acceptable
accuracy. Different sets of training patterns are required to train
the CCTs for each fault scenario. Once the training phase of the
Type-2 fuzzy NN is completed; the CCT for each fault scenario
can be quickly computed.
Another NN that associated with each fault scenario index is a
MLP NN that used for sensitivity analysis of input parameters in
ofine manner. For each fault scenario we run a MLP NN based sensitivity analysis to determine more effective operating conditions
that inuences CCT estimation accuracy for each fault. The MLP
NN uses ofine log database to determine more effective input
operating conditions in calculation of CCT and extract them as a
vector associated with each fault scenario. It reduces the training
time and the effort of measuring features online while declines
the complexity of Type-2 fuzzy neural network. When the scenario
selection block reports the index of fault scenario; the selector
block, lters input operating conditions according to the most
important feature vectors which is provided by MLP NN in the sensitivity analysis procedure. More details about sensitivity analysis
are presented later.
Generally; we should consider a separate Type-2 fuzzy NN for
each fault scenario in the power system TSA; like works done in
papers [6,12,13]. When we have limited computing capability,
we can only consider more important faults. On the other hand
some fault scenarios have similar effects on power system variables (e.g., voltage, active and reactive power). In working systems,
the combination of dependent faults instead of using separate fault
can be used to reduce the number of Type-2 neuro fuzzy models
and in consequence reduction of computation cost of the proposed
model. In this paper we consider a sample combination of fault
scenarios which is consist of 4 different uncorrelated and independents faults (the worst case scenario) and shows the proposed
Type-2 neuro fuzzy method can estimate CCT for each fault with
an acceptable accuracy. The capability of proposed method can

76

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187

Fig. 3. Block diagram of the proposed fuzzy NN based TSA.

be expanded by considering a set of Type-2 fuzzy NNs for each


probable fault scenario in the power system. It is obvious that
the expansion capability of the proposed Type-2 fuzzy NNs depend
on the applied data in the training process.
As mentioned before, the CCT of a particular fault scenario in
the New England test system is a function of 67 power system
operating conditions. The neural networks-based sensitivity analysis method is used to select 15 top most effective operating conditions in order to reduce the number of inputs to the Type-2 fuzzy
neural network.
Fuzzy neural systems
NNs are good at recognizing input patterns, but are weak to
explaining how they reach their decisions. On the other hand,
fuzzy logic systems can reason with imprecise information and
nicely explain their decisions but they are incapable of automatically acquiring the rules they use to make those decisions. These
limitations are cause for combination of two or more techniques
into intelligent hybrid systems. This combination overcomes the
limitations of individual techniques. Also hybrid systems are
important when dealing with different application domains.
The intelligent hybrid systems are used successfully in many
applications including process control, cognitive simulation, nancial trading, medical diagnosis, and engineering design. In the
intelligent hybrid systems an inference mechanism under cognitive uncertainty is provided by fuzzy logic and advantages, such
as learning, generalization, adaptation, parallelism and fault-tolerance are offered by computational NNs. The concept of NNs is
incorporated into the fuzzy logic in order to achieve humans like
capability for the system to deal with cognitive uncertainties. Usually fuzzy neural systems are used in two different congurations
as follows [23]:
In the rst conguration as illustrated in Fig. 4 the fuzzy interface block accepts linguistic statements and provides the perception as an input vector to a NN. After the training phase of NN, it
is adapted to yield desired command outputs or decisions.
Second conguration as illustrated in Fig. 5 use different structure. In this conguration a NN drives the fuzzy inference mechanism. The NNs have the roll of tuning membership functions of
fuzzy systems. In this conguration fuzzy systems are employed
as decision-maker for controlling equipment.
Fuzzy logic can use rules with linguistic labels which are
directly encoded by expert knowledge, but designing and tuning

of the membership functions which quantitatively dene these linguistic labels is a time consuming task. By using the learning capability of NN we can automate designing and tuning of the
membership functions and substantially reduce development time
and cost while improving performance.
In this paper we used the rst conguration of fuzzy neural system. The Type-2 fuzzy layer converts uncertain and noisy inputs to
more dependable and reliable linguistic variables which are used
as inputs to the MLP NN layer.

Type-2 fuzzy NN
The concept of the Type-2 fuzzy set was introduced by Zadeh
[24] as an extension to the type-1 fuzzy set. It can handle the
uncertainties associated with process and input variables. The idea
of the Type-2 fuzzy logic is shown in Fig. 6. Type-2 fuzzy logic consists of three main components named as: fuzzication (Type-2
fuzzier), inference (rule base and inference engine) and output
processing (type-reducer and difuzzier) [45]. The Type-2 fuzzier
transforms an input crisp variable into a Type-2 fuzzy set it uses in
circumstances where it is difcult to determine an exact membership function of input variables. Hence it is very useful for incorporating uncertainties.
The rule base block in Type-2 fuzzy logic consists of a set of
fuzzy IfThen rules which can handle rule uncertainties. The fuzzy
inference engine gives a mapping from the input Type-2 sets to the
output Type-2 sets. The type-reduction block converts the Type-2
output sets of inference engine to a type-1 set that is called the
type-reduced set. These type-reduced sets are then defuzzied
to obtain crisp outputs. As a result, Type-2 fuzzy logic systems
are very powerful paradigm to handle uncertainty in the real world

Decisions
Neural Network

Fuzzy Interface
Perception as
neural inputs

Neural
outputs
Linguistic
statements

Learning
algorithm

Fig. 4. The rst model of fuzzy neural system.

77

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187

Knowledge-base

Neural
Inputs
Neural Network

Neural
outputs

Decisions
Fuzzy Interface
Neural
outputs

Learning
algorithm
Fig. 7. Overview block diagram of Type-2 fuzzy NN.
Fig. 5. The second model of fuzzy neural system.

Step 1: Fig. 10 illustrates the degrees of membership to Aii and


l
Aii which are obtained for each of the input variable xi that is an
l
l
input to Type-2 fuzzy neural network. Where Aii and Aii are the
li th UMF and LMF of ith input variable respectively.
Step 2: after the degree of membership for total input variables
are obtained, vector Z and V that represents degree of membership set can be written as:


 A1 xi ;
Vi l
i

t

lA1i xi ; . . . ; l Ali xi ; lAli xi ; . . . ; l AK i xi ; lAK i xi


i

10

Fig. 6. The idea of a Type-2 fuzzy logic system.

Z V 1 ; V 2 ; . . . ; V i ; . . . ; V n t
applications and environments where there are uncertainties that
are difcult to predict.
Structure design of inference and the output processing algorithms in Type-2 fuzzy systems is difcult. Therefore, we use a
NN to model operation inference (rule base and inference engine)
and output processing (type-reducer and difuzzier) in Type-2
fuzzy system. Such a system which is named Type-2 fuzzy NN
combines the learning capability of NNs with the linguistic interpretation ability of fuzzy classiers. Overview block diagram of
the proposed Type-2 fuzzy NN system is illustrated in Fig. 7.
The block diagram is composed of two parts; a Type-2 fuzzier
(Type-2 fuzzy sets) and a neural network. The Type-2 fuzzier
enables efcient modeling of the linguistic and numerical uncertainties in the inputs and expert knowledge. Inputs to Type-2 fuzzier is crisp values while output vector consists of the fuzzied
inputs values which are called linguistic variable [23].
The NNs are designed in an attempt to mimic the human brain
and inspired from the biological world. These networks can be
trained and used for different types of problems such as function
approximation, mapping (pattern association and pattern classication) and clustering. In this paper we use a MLP NN trained by
the back-propagation algorithm [15] in the proposed Type-2 fuzzy
NN where inputs to the MLP NN is linguistic variable. In fact, the
MLP NN uses fuzzied inputs instead of crisp values that embed
uncertainty of measurement in input parameters. The components
of the proposed Type-2 fuzzy NN is given in Fig. 8. Referring to
Fig. 8, for each input variable xi, i = 1, 2, . . ., n, a set of Ki Type-2
l
fuzzy member ship functions are dened as Aii ; li 1; 2; . . . ; K i .
Here, we use Type-2 Gaussian membership functions for the
Type-2 fuzzy set. Fuzzifying inputs by a non-linear Type-2 fuzzy
membership function enables modeling the uncertainty of inputs.
As shown in Fig. 9. Each Type-2 membership function Ai is represented by an upper membership function (UMF) and a lower
membership function (LMF) that are denoted as Ax and Ax
 Ai xi and lAi xi are the member[21]. As it can be seen in Fig. 9; l
ship degree of input variable xi to the upper membership function
Ai and the lower membership function Ai respectively. The details
of data processing method in the proposed Type-2 fuzzy NNs are
described in the following steps:


 A1 x1 ;
Z l
1

11

lA11 x1 ; . . . ; l AK 1 x1 ; lAK1 x1 ; l A12 x2 ;


1

lA12 x2 ; . . . ; l AK 2 x2 ; lAK2 x2 ; . . . ; l A1i xi ; lA1i xi ; . . . ; l AK i xi ;


2

lAKi xi ; . . . ; l A1n xn ; lA1n xn ; . . . ; l AKn n xn ; lAKn n xn

t

12

 li xi and l li xi are respectively the membership degree of


where l
A
A
i

input variable xi to the UMF Aii and the LMF Aii .


The L is size of Z and can be obtained by the following equation:

L2

n
X
Ki

13

i1

Step 3: the fuzzied input vector Z is considered as input to NN.


A MLP consists of an input layer, an output layer and one or
more hidden layers. L represents the size of the input layer
where size of the output layer is determined by number of
NN outputs. Number of the hidden layer and the size of hidden
neurons at each hidden layer are carefully selected to provide
best results in training phase of system. One of the popular
training algorithms for MLP NN is error back propagation algorithm [15], which is based on the gradient descent technique.
The standard back propagation method is too slow in MLP NN
training [46], so it is superseded by fast resilient back-propagation technique for training the MLP NN.
Training and testing of Type-2 fuzzy NN to predict CCT
In this section, we give a general overview of training procedure
of Type-2 fuzzy NN which is designed in previous section. As mentioned before the structure of Type-2 fuzzy NN is composed of two
parts, a Type-2 fuzzier and a NN. In order to Type-2 fuzzy NN can
be used in estimation of CCT, it should rst be trained. The training
inputs should be carefully selected in a range related to possible
variation of input parameters. According to Section Test system,
the CCT for a particular fault scenario in New England 39-bus test
system is function of 67 (19 + 19 + 9 + 9 + 1 + 1 + 9 = 67) power
systems operating parameters. This conguration can be applied

78

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187

Fig. 8. The components of the proposed Type-2 fuzzy neural network.

2. A random value with uniform distribution is assigned independently to each of the variables mentioned in step 1as turbulence. The random turbulence described by the following
equations:

Fig. 9. Upper and a lower membership functions in Type-2 Gaussian membership


function.

Fig. 10. The structure of the Type-2 fuzzier for input variable xi.

to estimate the CCT for every valid fault scenario in a given power
system. As described in Refs. [6,12,13], and illustrated in Fig. 11,
the training and testing patterns of Type-2 fuzzy NN can be
obtained from the following procedures:
1. It is assumed that the following operating conditions can vary
randomly over some specied ranges of their nominal values:
 Voltage magnitudes of all the nine PV buses are bounded
between 0.9 and 1.1 times their corresponding nominal values.
 Active and reactive powers of all the 19 loads are varied in the
range 0.61.1 times their corresponding nominal values.



PDj k PDj0 0:6 0:5ejPD k

14



QDj k QDj0 0:6 0:5ejQD k

15



V i k V i0 0:9 0:2eiV k

16

where PDj(k) and QDj(k) are, the active power and the reactive
power of load at the jth load bus, respectively. And PDj0 and QDj0
are their nominal load at the jth load bus. Also Vi(k) is the voltage
magnitude at the ith PV bus for the kth training pattern and Vi0 is
the nominal voltage magnitude at the ith PV bus. Also e denotes a
uniformly distributed random number within range [0, 1].
3. In the next step, a load-ow analysis is performed on the above
mentioned training data to make sure that each of the scenarios
is a feasible power ow solution. Moreover, other operating
conditions including generated active and reactive power of
the slack buses and generated reactive powers of all the nine
PV buses which are assumed as input to Type-2 fuzzy NN are
obtained from the results of the load-ow program.
4. A time-domain simulation technique is employed to compute
the CCT for all the sample fault scenarios by solving power system differential equations. The solving method is 4th order
RungeKutta with time scale resolution of Dt = 0.001 s.
5. As shown in Fig. 11, in parallel to the time-domain simulation,
the MLP NN based sensitivity analysis is used to reduce number
of input operating conditions and chooses more effective ones
in order to simplify the design and training procedure of
Type-2 fuzzy NN.

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187

79

Fig. 11. Block diagram of the proposed Type-2 fuzzy NN based approach.

6. Results of CCT from the time-domain simulation block is used


as the desired output of Type-2 fuzzy NN that is associated with
selected input operating conditions which are used in training
procedure of Type-2 fuzzy neural network.

did not have noticeable change in error are omitted from Type-2
fuzzy NN inputs.

It should be noted that in this paper the MATPOWER software


[47] is applied for load-ow analysis. The presented pattern generation procedure produces 2000 random data patterns for sample
fault scenarios. Among them 1500 patterns (about 75% of the total
produced patterns) were chosen for training and the remaining
500 patterns (25% of the total produced patterns) were applied
to test the Type-2 fuzzy neural network.

In this section, we evaluate the accuracy of proposed Type-2


fuzzy NN in CCT estimation for sample fault scenarios in New England 39-bus test system. Also, a MLP NNs based sensitivity analysis
is used to reduce the number of input parameters to Type-2 fuzzy
NN as illustrated in the block diagram of Fig. 13. So the total number of inputs is reduced from 67 operating conditions to only 15
more effective operating conditions for New England 39-bus test
system. Then the design is trained by 15 inputs to estimate CCT
for four sample fault scenarios.

MLP NN based sensitivity analysis method


Sensitivity analysis is the study of how the variation in the output of a model can be allocated, qualitatively or quantitatively to
different sources of variation [17,43]. The main idea behind the
sensitivity analysis method is that it reduces the number of input
features from original set of features without losing the important
information. According to Section Test system, we have 67 systems operating parameters as input to the Type-2 fuzzy neural
network. These operating parameters have direct impacts in calculation of CCT. However some of these operating parameters have
more information and higher inuence in nal CCT than others.
Therefore, we are going to rank the operating parameters in order
of their importance and determine those operating parameters
(inputs) for each fault scenario that have higher impact on the nal
value of CCT. So we use a MLP NN based sensitivity analysis
method as proposed in Refs. [12,17,43], to nd the operating
parameters (Type-2 fuzzy NN inputs) those have more impact on
CCT. The benet of reduction in inputs of Type-2 fuzzy NN is that
it reduces the training time and effort of on-line measurement of
features while declines in complexity of Type-2 fuzzy NN.
As illustrated in Fig. 12, in this analysis, one input parameter at
a time is selected and varied by a xed percentage while rest of the
inputs remains constant. Then input data is fed to the NN and the
mean absolute error (MAE) is calculated. These steps are repeated
for all the input parameters and MAE is calculated. After all the
inputs have been examined, ranking decision is made based on
the change in MAE. Those inputs whose change in value makes a
noticeable change in the value of output MAE are selected as favorite candidate as input parameters. Remaining parameters which

Results and discussion

Application of Type-2 fuzzy NN to predict the CCT


In this section, rst we design a Type-2 fuzzy NN then the
design is trained by 15 more effective operating conditions to estimate CCT for four sample fault scenarios. Here, we select 15 operating conditions among of total 67 inputs that have more effect on
CCT of four sample fault scenarios. These 15 selected operating
conditions are: VG2, PG5, VG9, VG3, VG10, PG2, VG5, PD1, PG7,
VG8, PG4, PG6, VG7, VG4, and PG1. The details of sensitivity analysis based parameter selection method is presented in the next
section. Design and train procedures of the proposed Type-2 fuzzy
NN are described in the following steps:
Step 1: the maximum and minimum variation of 15 input operating conditions are extracted from 1500 training patterns as
shown in Table 2.
Step 2: for each of the 15 inputs, three sets of Type-2 fuzzy are
dened as S, M and L. Here, Gaussian membership functions are chosen for inputs as shown in Fig. 14. Moreover the
number of membership functions and UMF and LMF variances
are determined by an expert knowledge.
Step 3: after Type-2 fuzzier is designed, we are going to design
NN. Here, we use a single layer perceptron NN and train it by
the resilient back-propagation method. The number of input
neurons is obtained by the following equation:

L2

n
X
K i 2  15  3 90
i1

17

80

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187

Fig. 12. Block diagram of the MLP NN based sensitivity analysis.

Fig. 13. Block diagram of the proposed Type-2 fuzzy NN based approach.

The perceptron NN has 90 neurons in input layer and four neurons in output layer without any hidden layer. A linear transfer
function is applied for the neurons in output layer. The MSE threshold to stop the training procedure was set to 0.005 s. It took about
8 s with 1159 epochs on average to train the proposed Type-2
fuzzy NN.
In order to evaluate how well the trained Type-2 fuzzy NN
reacts on the sample fault scenarios; it was tested by 500 test patterns. Performance of the proposed Type-2 fuzzy NN is evaluated
by Root Mean-Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE) between real and estimated values. The denition is given by Eqs. (18) and (19) as follows:

v
u N
u1 X
2
RMSE t
actual CCTp  estimated CCTp
N p1

MAPE 100 


N

1X
actual CCTp  estimated CCTp


N p1
actual CCTp

Fig. 14. Membership functions of the input xi.

Table 3
The RMSE and MAPE for 500 total testing patterns for each fault scenario.

18

19

RMSE
MAPE

Fault scenario
1

Fault scenario
2

Fault scenario
3

Fault scenario
4

0.0102
2.49%

0.0061
1.69%

0.009
2.41%

0.0061
2.11%

Table 2
The maximum and minimum value for each 15 input.
Input

Minimum value

Maximum value

Input

Minimum value

Maximum value

Input

Minimum value

Maximum value

VG2
PG5
VG9
VG3
VG10

0.8838 pu
3.7415 pu
0.9239 pu
0.8848 pu
0.9428 pu

1.0802 pu
5.1073 pu
1.1291 pu
1.0814 pu
1.1522 pu

PG2
VG5
PD1
PG7
VG8

4.6651 pu
0.9112 pu
6.6246 pu
4.1245 pu
0.9250 pu

6.3680 pu
1.1134 pu
12.1439 pu
5.6301 pu
1.1306 pu

PG4
PG6
VG7
VG4
PG1

4.6548 pu
4.7874 pu
0.9572 pu
0.8975 pu
6.3707 pu

6.3540 pu
6.5350 pu
1.1698 pu
1.0969 pu
8.9442 pu

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187

Fig. 15. Comparison between the estimated output and the actual output for selected fault scenarios of New England 39-bus test system.

Fig. 16. Distribution of errors between the actual CCT and the estimated CCT in percent (Ep) for all sample fault scenarios.

81

82

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187


Table 4
Specication of MLP NNs which are used in sensitivity analysis.
Fault scenario
number
Fault
Fault
Fault
Fault

scenario
scenario
scenario
scenario

1
2
3
4

Number of NN
inputs

Number of NN hidden
neurons

Average NN training
time (s)

Average training
epochs

67
67
67
67

9
9
12
10

9
7
8
6

891
749
689
526

Table 5
The RMSE and MAPE for total 500 testing patterns of four fault scenarios.
Fault scenario 1 Fault scenario 2 Fault scenario 3 Fault scenario 4
RMSE 0.0067
MAPE 1.61%

0.0055
1.59%

0.0049
1.69%

0.0052
1.82%

where N is the total number of patterns in test set and p represents


the pattern index. The RMSE and MAPE for 500 total test patterns
are shown in Table 3 for each sample fault scenario.
The results are proving generalization accuracy of trained Type2 fuzzy NN for all sample fault scenarios. Comparison between the
actual CCT and the estimated CCT by using 30 test patterns out of
total 500 test patterns is shown in Fig. 15.
Also, distribution of errors in percent for sample fault scenarios
are shown in Fig. 16 and calculated by the following equation:

Ep 100 

actual Targetp  estimated Targetp


actual Targetp

20

where p represents index of test pattern. As it can be seen in Fig. 16,


occurrence frequency of error in percent (Ep) for all fault scenarios
follows the Gaussian distribution. In the rst fault scenario about
30.4% of total test patterns have Ep between 1% and 1%. Moreover,
only 0.2% of total 500 test patterns have Ep between 12% and
11% which is worse case error.
In the second fault scenario about 40.6% of total test patterns
have Ep between 1% and 1%. Moreover, only 0.2% of total 500 test
patterns have Ep between 12% and 11% which is worse case
error. Also in the third fault scenario about 27.6% of total test patterns have Ep between 1% and 1%. Moreover, only 0.2% of total
500 test patterns have Ep between 9% and 8% which is worse
case error. Finally in the fourth fault scenario about 34.4% of total
test patterns have Ep between 1% and 1%. Moreover, only 0.2% of
total 500 test patterns have Ep between 10% and 11% which is
worse case error.
As shown in Fig. 16 we can conclude that for all the sample scenarios, small |EP|s (less than 5%) are more frequent (about 90%)
than large |EP|s. on the other hand if a Gaussian probability function is tted to the EP distribution, it will be have small variance
and mean value near to zero. So in 90% of conditions the estimation

Fig. 17. Comparison between MLP NN CCT estimation and the actual CCT for sample fault scenarios of New England 39-bus test system.

83

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187


Table 6
15 Most important inputs and their MAE for each fault scenario.
Fault scenario 1

Fault scenario 2

Fault scenario 3

Fault scenario 4

Operating conditions

Value of MAE

Operating conditions

Value of MAE

Operating conditions

Value of MAE

Operating conditions

Value of MAE

VG2
VG9
VG10
PD1
VG5
PG7
VG8
PG1
PG6
PG10
PG3
PG9
PG4
PG2
VG6

0.1682
0.0885
0.0576
0.0548
0.0528
0.0509
0.0409
0.0329
0.0245
0.0236
0.0222
0.0196
0.0189
0.0187
0.0164

PG10
VG2
PD1
PG8
VG6
PG9
VG10
VG3
VG9
VG4
PG5
PG6
VG5
VG8
VG7

0.053
0.052
0.0516
0.0439
0.0434
0.0421
0.0418
0.0414
0.0302
0.0287
0.0279
0.0216
0.0197
0.0182
0.0146

VG9
VG10
PD1
VG2
PG10
PG8
PG4
PG2
VG8
PG1
VG4
PG7
PG9
PD19
PG3

0.1167
0.0596
0.0525
0.0432
0.0424
0.0411
0.0399
0.0396
0.0331
0.0306
0.0275
0.0151
0.0147
0.0143
0.0136

VG2
VG3
PG7
PG4
VG7
VG9
PG9
PG10
PG6
VG6
PD1
VG10
PG3
VG4
PG2

0.1636
0.0589
0.0526
0.0481
0.046
0.0439
0.037
0.0359
0.0277
0.0254
0.0242
0.0209
0.0153
0.0145
0.0139

Table 7
Comparison between MLP NN and Type-2 fuzzy NN in training phase.

Training algorithms
Training time
Average training
epochs
Training MSE
Number of inputs

Type-2 fuzzy neural


network

MLP NN

Resilient backpropagation
8s
1159

Resilient backpropagation
16 s
1842

0.005
15

0.035
20

error is less than 5%. It can be observed that the trained Type-2
fuzzy NN estimates actual CCT for all sample fault scenarios with
an acceptable accuracy. The experimental results show that proposed Type-2 fuzzy NN which is only considered higher impact
inputs give satisfactory CCT estimation for all sample fault
scenarios.
Sensitivity analysis
As described in Section MLP NN based sensitivity analysis
method a MLP NNs based sensitivity analysis is used to reduce
number of input parameters to the Type-2 fuzzy NN. This reduction can affect estimation accuracy. In this section we are going
to analyze the effect of MLP NNs based sensitivity analysis on estimation accuracy. The MLP NN is rst trained with 67 operating
conditions in each training pattern. Then the trained NN is applied
to perform sensitivity analysis. We use four MLP NNs to perform
sensitivity analysis for each fault scenario separately. Details of
four trained MLP NNs are shown in Table 4.
Here, number of neurons in hidden layer are tuned after running several experiments. Also for all the MLP NNs tangent sigmoid
transfer function and linear transfer function are applied to the

hidden layer neurons and the output layer neurons respectively.


MSE between actual and estimated values of CCT in the training
phase was set to 0.001 s. The training method for MLP NN is resilient back-propagation method. Total computations are performed
on a personal computer with 2 GB of RAM and 2.93 GHz Pentium
Dual-Core Processor.
In order to see how well the trained MLP NNs acts in simulated
fault scenarios, the NN was tested by 500 test patterns. The performance of trained MLP NN in evaluation of CCT is determined by
RMSE and MAPE between actual and estimated CCT. The RMSE
and MAPE for 500 total test patterns and sample fault scenarios
are shown in Table 5.
The results of Table 5 are proving the generalization accuracy of
trained MLP NN for all sample fault scenarios. Comparison
between the actual CCT and the estimated CCT for 30 out of 500
test patterns are shown in Fig. 17 for sample fault scenarios. It
can be observed that the trained MLP NNs estimates actual CCT
with an acceptable accuracy in sample fault scenarios.
When MLP NNs was trained by 67 inputs, it is used to perform
sensitivity analysis. The sensitivity of NNs output (CCT) respect to
the each input are calculated. Each time, one MLP NN input is chosen to train and perturbed by a xed percentage of its value (e.g.,
10%), where the other inputs remains constant. In the same time
inputs are fed to the MLP NNs and output (CCT) is calculated.
The mean absolute error (MAE) between outputs which are estimated by NN before and after each variation to the input is computed. These steps are repeated for all the NNs inputs and nally
MAE is computed.
After all the 67 inputs have been examined, those inputs whose
change leads to a big variation in the MAE are chosen as best candidates for input to the Type-2 fuzzy NN. The rest of inputs with
small MAE, does not considered for Type-2 fuzzy NN. As a matter
of fact, a big value for MAE means that the corresponding perturbed variable greatly inuences the system output. MAE is computed by the following equation:

Table 8
The RMSE and MAPE of MLP NN and Type-2 fuzzy NN for each fault scenario in testing phase.
Type-2 fuzzy neural network

RMSE
MAPE
Response time for 500 test patterns

MLP NN

Fault
scenario 1

Fault
scenario 2

Fault
scenario 3

Fault
scenario 4

Fault
scenario 1

Fault
scenario 2

Fault
scenario 3

Fault
scenario 4

0.0102
2.49%
0.1020 s

0.0061
1.69%

0.009
2.41%

0.0061
2.11%

0.0265
8.01%
0.1015 s

0.0208
6.44%

0.0208
7.13%

0.0248
9.09%

84

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187

Fig. 18. Estimated CCT by Type-2 fuzzy NN and MLP NN are compared with actual CCT for each fault scenario in the New England 39-bus test system.

MAE


N
0

1X
t cr p  t cr p


N p1
t cr p

21

where tcr p is estimated CCT before any variation to the input and
t 0cr p is estimated CCT after variation of input. N is the total number
of testing patterns and p represents pattern index. After conducting

sensitivity analysis for each fault scenario, 15 out of total 67 input


parameters that have higher impact on MAE are chosen. The 15
most important inputs and their MAE for each fault scenario are
shown in Table 6.
As illustrated in Table 6, most of the chosen input operating
conditions are voltage magnitudes and the active generated

Fig. 19. Distribution of error in percent between the actual CCT and the estimated CCT for all the sample fault scenarios using the MLP NN.

85

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187


Table 9
Different value of Ep, variances of Ep and means of Ep corresponding to the MLP NN and the Type-2 fuzzy NN for sample fault scenarios.
Fault scenario index

Maximum value of Ep
Percentage of worst case error in Ep distribution
Distribution percentage Ep between 1% and 1%
Variance of Ep
Mean of Ep

Type-2 fuzzy NN

MLP NN

12%, 11%
0.2%
30.4%
8.1099
0.1044%

12%, 11%
0.2%
40.6%
4.7271
0.1988%

9%, 8%
0.2%
27.6%
6.8101
0.0398%

10%, 11%
0.2%
34.4%
7.7759
0.0993%

29%, 30%
0.2%
8.2%
94.1163
0.7957%

24%, 23%
0.2%
9%
61.2167
1.7208%

28%, 27%
0.2%
7.4%
72.2927
1.169%

32%, 31%
0.6%
7.2%
127.4749
0.6521%

Fig. 20. Comparison of absolute error (AE) between actual CCT and the estimated CCT (by Type-2 fuzzy NN and MLP NN) for sample fault scenarios.

powers of those generators that are in proximity of the fault


scenario in test system.
Comparison between Type-2 fuzzy NN and MLP NN for CCI prediction
In order to compare the performance of proposed Type-2 fuzzy
NN with other relevant methods, we implement a MLP NN like the
approach which is used in papers [6,12,13]. In order to better convergence of MLP NN, we used 20 inputs which are including the
Type-2 fuzzy neural inputs and extra ve inputs (next ranks in
the sensitivity analysis). Five additional inputs are as follows:
PG9, PG3, PG10, VG6 and PG8. Also the same number of train
(1500) and test (500) patterns as we applied to the Type-2 fuzzy
NN are used for the MLP NN.
The MLP NN has 20 neurons in input layer and one hidden layer
that includes 12 neurons. A tangent sigmoid transfer function is
applied to the hidden layer neurons. Also a linear transfer function
is applied to the output layer neurons. MSE threshold in the training phase of NN was set to 0.035 s. The MLP NN is trained by resilient back-propagation method. It took about 16 s with average of

1842 epochs to train the MLP NN. The trained MLP NN was tested
by using 500 test patterns. The response time for 500 test patterns
and the RMSE and MAPE values for 500 total test patterns are computed for each scenario. The result of experiments for comparison
between the MLP NN and Type-2 fuzzy NN are illustrated in Tables
7 and 8.
It can be seen in Tables 7 and 8, that in the training phase, MLP
NN consumes more training time and takes more epochs to converge in compare with the Type-2 fuzzy NN. Also in testing phase
of the system RMSE and MAPE of MLP NN is higher than the Type-2
fuzzy NN results, that indicates higher rate of estimation error. The
estimated CCT results that obtained by Type-2 fuzzy NN and MLP
NN are compared with actual CCT for rst 30 test patterns out of
500 total test patterns are shown in Fig. 18. The results are presented for each fault scenario separately. Also, the distribution of
errors in percent are shown in Fig. 19 for sample fault scenarios.
As it can be seen in Fig. 19, the distribution of Ep follows a
Gaussian distribution. For rst fault scenario about 8.2% of total
test patterns have Ep between 1% and 1%. Also, only 0.2% of
500 total test patterns have Ep between 29% and 30% which is

86

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187

worse case error. In the second fault scenario about 9% of total test
patterns have Ep between 1% and 1%. Moreover, only 0.2% of total
500 test patterns have Ep between 24% and 23% which is also a
worse case error. In third fault scenario about 7.4% of the total test
patterns have Ep between 1% and 1%. Where, only 0.2% of the 500
total test patterns have Ep between 28% and 27% which is
worse case error. Finally in the fourth fault scenario about 7.2%
of total test patterns have Ep between 1% and 1%. Moreover, only
0.6% of total 500 test patterns have Ep between 32% and 31%
which is worse case error.
As shown in Fig. 19 we can conclude that for all sample scenarios, small |EP|s (less than 5%) are about half of samples (about 50%)
and large |EP|s are more frequent than Type-2 fuzzy NN results
(Fig. 16). On the other hand if a Gaussian probability distribution
function is tted to the EP distribution, it will be have large variance and negative mean value near to zero. So in 50% of conditions
the MLP NN estimation error is less than 5%. Different values of Ep,
variances of Ep and the means of Ep corresponding to two different
methods are compared in Table 9 for sample fault scenarios.
Also absolute error (AE) between the actual CCT and the estimated CCT by Type-2 fuzzy NN and MLP NN for rst 30 out of
500 test patterns are shown in Fig. 20. The results are presented
for each fault scenario separately. It can be concluded that the
Type-2 fuzzy NN could estimate CCT with higher degree of accuracy than MLP NN for sample fault scenarios. Also the proposed
method requires less computational cost in compare to the MLP
NN method.
Conclusion
In this paper we propose a new direct method for online TSA
problem in power systems. A hybrid Type-2 fuzzy NN system is
designed to estimate the CCT of sample contingency (fault scenarios). Also, MLP NN based sensitivity analysis is used to reduce the
number of inputs to Type-2 fuzzy NN about four times. That results
to less complex and faster Type-2 fuzzy NN system with negligible
decrease in estimation accuracy. By using Type-2 fuzzy sets as
Type-2 fuzzier, we can handle the uncertainties which are associated to measurements of operating conditions and device parameters in a power system and complexity of power networks
effectively. Type-2 fuzzy layer converts uncertain and noisy inputs
to more dependable and reliable linguistic variables which are
used as inputs to the MLP NN layer. Moreover, heavy computational burden is avoided in online applications. The outputs of
Type-2 fuzzier are injected to a single layer perceptron NN to estimate CCT. We applied resilient back-propagation method for fast
ofine training of Type-2 fuzzy NN system.
New England 10-machine 39-bus standard test power system
was applied as an example to demonstrate the efciency of proposed method. Simulation results show that the proposed Type-2
fuzzy NN could estimate the CCT for sample fault scenarios with
reasonable accuracy at different system operating conditions in
compare to widely used MLP NN method. The proposed Type-2
fuzzy NN reduces RMSE and MAPE of CCT estimation about four
times in compare to MLP NN method. In addition, the proposed
method has very simple structure and consumes low computational power for training and in online systems; therefore the solution is feasible and can be employed for fast assessment of the
transient stability in a power system control center.
References
[1] Kundur P. Power system stability and control. McGraw-Hill; 1994.
[2] Sauer PW, Pai MA. Power system dynamics and stability. New Jersey: PrenticeHall; 1998.
[3] IEEE/CIGRE Joint Task Force. Denition and classication of power system
stability. IEEE Trans Power Syst 2004: 1(2).

[4] Sauer PW, Tomsovic KL, Vittal V. Dynamic security assessment. In: Grigsby LG,
editor. Power system stability and control. New York: CRC Press; 2007.
[5] A-Wahab NI, Mohamed A, Hussain A. Fast transient stability assessment of
large power system using probabilistic neural network with feature reduction
techniques. Expert Syst Appl 2011;38(9):111129.
[6] Karami A. Estimation of the critical clearing time using MLP and RBF neural
networks. Eur Trans Electr Power 2010;20(2):20617.
[7] Pai MA. Energy function analysis for power system stability. Kluwer Academic;
1989.
[8] Sobajic DJ, Pao YH. Articial neural-net based dynamic security assessment for
electric power systems. IEEE Trans Power Syst 1989;4(1):2206.
[9] Pao YH, Sobajic DJ. Combined use of unsupervised and supervised learning for
dynamic security assessment. IEEE Trans Power Syst 1992;7(2):87884.
[10] Aboytes F, Ramirez R. Transient stability assessment in longitudinal power
systems using articial neural networks. IEEE Trans Power Syst
1996;11(4):200310.
[11] Bahbah AG, Girgis AA. New method for generators angles and angular
velocities prediction for transient stability assessment of multimachine power
systems using recurrent articial neural network. IEEE Trans Power Syst
2004;19(2):101522.
[12] Karami A, Esmaili SZ. Transient stability assessment of power systems
described with detailed models using neural network. Int J Electr Power
Energy Syst 2013;45:27992.
[13] Karami A. Power system transient stability margin estimation using neural
networks. Int J Electr Power Energy Syst 2011;33:98391.
[14] Assis TML, Nohara AA, Valentini TM. Power system dynamic security
assessment through a neuro-fuzzy scheme. In: 15th International conference
on intelligent system applications to power systems, 2009. p. 16.
[15] Haykin S. Neural networks: a comprehensive foundation. 2nd ed. NJ: PrenticeHall; 1999.
[16] Refaee JA, Mohandes M, Maghrabi H. Radial basis function networks for
contingency analysis of bulk power systems. IEEE Trans Power Syst
1999;14(2):7728.
[17] Sawhney H, Jeyasurya B. A feed-forward articial neural network with
enhanced feature selection for power system transient stability assessment.
Electr Power Syst Res 2006;76:104754.
[18] Lin YJ. Explaining critical clearing time with rules extracted from a multilayer
perceptron articial neural network. Int J Electr Power Energy Syst
2010;32:8738.
[19] Haidar AMA, Mustafa MW, Ibrahim FAF, Ahmed IA. Transient stability
evaluation of electrical power system using generalized regression neural
networks. Appl Soft Comput 2011;11:355870.
[20] Ferreira WP, Silveira MCG, Lotufo ADP, Minussi CR. Transient stability analysis
of electric energy systems via a fuzzy ARTARTMAP neural network. Electr
Power Syst Res 2006;76:46675.
[21] Marchiori SC, Silveira MCG, Lotufo ADP, Minussi CR, Lopes MLM. Neural
network-based on adaptive resonance theory with continuous training for
multi-conguration transient stability analysis of electric power systems. Appl
Soft Comput 2011;11:70615.
[22] Augutis J, Zutautaite I, Radziukynas V, Krikstolaitis R, Kadisa S. Application of
Bayesian method for electrical power system transient stability assessment.
Int J Electr Power Energy Syst 2012;42:46572.
[23] Robert Fullr. Introduction to neuro-fuzzy systems, advances in soft
computing series. Berlin/Heildelberg: Springer-Verlag; 2000 [chapter 3].
[24] Zadeh LA. The concept of linguistic variable and its application to approximate
reasoning-1. Inf Sci 1975;8:199249.
[25] Abbadi A, Nezli L, Boukhetala D. A nonlinear voltage controller based on
interval type 2 fuzzy logic control system for multimachine power systems. Int
J Electr Power Energy Syst 2013;45:45667.
[26] Sudha KR, Vijaya Santhi R. Robust decentralized load frequency control of
interconnected power system with generation rate constraint using Type2fuzzy approach. Int J Electr Power Energy Syst 2011;33:699707.
[27] Sudha KR, Vijaya Santhi R. Load frequency control of an interconnected reheat
thermal system using Type-2fuzzy system including SMES units. Int J Electr
Power Energy Syst 2012;43:138392.
[28] Afzalian A, Linkens DA. Training of neurofuzzy power system stabilisers using
genetic algorithms. Int J Electr Power Energy Syst 2000;22:93102.
[29] Radaideh SM, Nejdawi IM, Mushtaha MH. Design of power system stabilizers
using two level fuzzy and adaptive neuro-fuzzy inference systems. Int J Electr
Power Energy Syst 2012;35:4756.
[30] Jang JSR, Sun CT, Mizutani E. In: Neuro-fuzzy and soft computing. New
Jersey: Prentice-Hall; 1997 [ch 17].
[31] Juang CF. A TSK-type recurrent fuzzy network for dynamic systems processing
by neural network and genetic algorithm. IEEE Trans Fuzzy Syst
2002;10(2):15570.
[32] Efe MO, Kaynak O. On stabilization of gradient-based training strategies
for computationally intelligent systems. IEEE Trans Fuzzy Syst 2000;8(5):
56475.
[33] Abiyev RH, Kaynak O. Fuzzy wavelet neural networks for identication and
control of dynamic plantsa novel structure and a comparative study. IEEE
Trans Ind Electron 2008;55(8):313340.
[34] Abiyev RH, Kaynak O, Alshanableh T, Mamedov F. A Type-2neuro-fuzzy system
based on clustering and gradient techniques applied to system identication
and channel equalization. Appl Soft Comput 2011;11:1396406.
[35] Pandit M, Srivastava L, Sharma J. Voltage contingency ranking using fuzzied
multilayer perceptron. Electr Power Syst Res 2001;59:6573.

A. Sharian, S. Sharian / Electrical Power and Energy Systems 64 (2015) 7187


[36] Abdelaziz AY, Abbas AA, Naiem AF, Elsharkawy MA. Transient stability
assessment using an adaptive fuzzy classication technique. Electr Power
Compon Syst 2006;34:92740.
[37] Wadduwagea DP, Wub ChQ, Annakkage UD. Power system transient stability
analysis via the concept of Lyapunov Exponents. Electr Power Syst Res
2013;104:18392.
[38] You D, Ye Lei, Wu J, Huang R. Transient stability assessment of power system
using support vector machine with generator combinatorial trajectories
inputs. Int J Electr Power Energy Syst 2013;44(1):31825.
[39] Lin Y. Comparison of CART- and MLP-based power system transient stability
preventive control. Int J Electr Power Energy Syst 2013;45(1):12936.
[40] Fang J, Yao W, Wen J, Cheng Sh, Tang Y, Cheng Z. Probabilistic assessment of
power system transient stability incorporating SMES. Phys C: Supercond
2013;484:27681.
[41] Dou Ch, Yang J, Li X, Gui T, Bi Y. Decentralized coordinated control for large
power system based on transient stability assessment. Int J Electr Power
Energy Syst 2013;46:15362.

87

[42] Mahmuda MA, Hossain MJ, Pota HR. Transient stability enhancement of
multimachine power systems using nonlinear observer-based excitation
controller. Int J Electr Power Energy Syst 2014;58:5763.
[43] Chakrabarti S, Jeyasurya B. Generation rescheduling using ANN-based
computation of parameter sensitivities of the voltage stability margin. Eng
Appl Artif Intell 2008;21:11649.
[44] Fouad AA, Vittal V. Power system transient stability analysis using the
transient energy function method. Prentice-Hall; 1992.
[45] Karnik NN, Mendel JM, Liang Q. Type-2fuzzy logic systems. IEEE Trans Fuzzy
Syst 1999;7(6):64358.
[46] Riedmiller M, Braun H. A direct adaptive method for faster backpropagation
learning: the RPROP algorithm. In: Proceedings of the international conference
on neural networks, San Francisco, 1993.
[47] MATPOWER, Version 3.0.0. Power systems engineering research center, school
of electrical engineering, Cornell University, Ithaca, 2005.