Engineering Science and Technology, an International Journal 19 (2016) 1368–1380

Contents lists available at ScienceDirect

Engineering Science and Technology,

an International Journal
journal homepage: www.elsevier.com/locate/jestch

Full Length Article

Support vector machine based fault classification and location

of a long transmission line
Papia Ray a,⇑, Debani Prasad Mishra b
a
Department of Electrical Engineering, Veer Surendra Sai University of Technology, Burla, Odisha, India
b
Department of Electrical and Electronics Engineering, International Institute of Information Technology, Bhubaneswar, Odisha, India

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

Article history: This paper investigates support vector machine based fault type and distance estimation scheme in a long
Received 28 December 2015 transmission line. The planned technique uses post fault single cycle current waveform and
Revised 19 March 2016 pre-processing of the samples is done by wavelet packet transform. Energy and entropy are obtained
Accepted 4 April 2016
from the decomposed coefficients and feature matrix is prepared. Then the redundant features from
Available online 20 April 2016
the matrix are taken out by the forward feature selection method and normalized. Test and train data
are developed by taking into consideration variables of a simulation situation like fault type, resistance
Keywords:
path, inception angle, and distance. In this paper 10 different types of short circuit fault are analyzed.
Fault classification
Feature selection
The test data are examined by support vector machine whose parameters are optimized by particle
Fault location swarm optimization method. The anticipated method is checked on a 400 kV, 300 km long transmission
Support vector machine line with voltage source at both the ends. Two cases were examined with the proposed method. The first
Wavelet packet transform one is fault very near to both the source end (front and rear) and the second one is support vector
Transmission line machine with and without optimized parameter. Simulation result indicates that the anticipated method
for fault classification gives high accuracy (99.21%) and least fault distance estimation error (<0.21%) for all
discussed cases. In order to verify the accuracy of the proposed method, a comparison is carried out with
methods published by other researchers. Separate investigation is also carried out with the transmission line
placing thyristor controlled series capacitor in the middle and applying the same proposed method. It is
observed from the test results of the thyristor controlled series capacitor based transmission line model that
fault classification gives a high accuracy of 98.36% and absolute fault location error is >0.29%.
Ó 2016 Karabuk University. Publishing services by Elsevier B.V. This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction
Double line-to-ground fault (LLG).

Triple line fault (LLL).
Progress of a country is measured by per capita consumption of
electric energy. Protection engineers find it difficult to maintain After the occurrence of a fault, restoration of power supply is
uninterrupted electric power to the end users due to the presence possible only when the maintenance crew finishes the repair work.
of fault in a transmission line [1]. Generally the reason for the If failure of power supply is elongated then it leads to line outage,
transmission line fault is hard to discover, so it is very important economic losses and wastage of time and energy of maintenance
to build up a fault analyzer that can examine the type of the fault workers. So it is required to classify and locate the fault quickly
and estimate the fault distance quickly and accurately. Fault occur- and correctly, otherwise the whole transmission line has to be
rence takes place when conductors touch each other or the ground examined by the maintenance worker in order to find the exact
[2], and are classified in a three phase system as: fault position.
In the recent past, many researchers have investigated in a long

Single line-to-ground fault (SLG). transmission line following fault location and classification tech-

Line-to-line fault (LL). niques [3]

⇑ Corresponding author.
Impedance measurement based technique.

E-mail addresses: [email protected] (P. Ray), [email protected]
Traveling wave phenomenon based technique.
(D.P. Mishra).
Artificial Intelligence based technique.
Peer review under responsibility of Karabuk University.

http://dx.doi.org/10.1016/j.jestch.2016.04.001
2215-0986/Ó 2016 Karabuk University. Publishing services by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 P. Ray, D.P. Mishra / Engineering Science and Technology, an International Journal 19 (2016) 1368–1380 1369

Nomenclature

List of symbols w inertia weight in the case of particle swarm optimiza-

V volts tion
I current c1, c2 acceleration constant in case of particle swarm opti-
x(t) signal mization
x mean value of current signal wmin, wmax initial and final values of weighting coefficients in
t time (s) case of particle swarm optimization method
E energy Es voltage source at the sending end of the transmission
EN entropy line
c cost parameter of support vector machine Er voltage source at the receiving end of the transmission
g gamma parameter of support vector machine line
f fitness value of particle swarm optimization Rf fault resistance (in ohms)
y(k) discrete samples h fault inception angle (in degree)
e(w) evaluation function of forward feature selection e error
method Fs sampling frequency (in Hertz)

Impedance measurement based technique mainly depends on fault classification and location in a high voltage power transmis-
fundamental frequency current and voltages. This method is sim- sion line is discussed by using wavelet packet transform and
ple and cheap, but gives erroneous results for huge value of the particle swarm optimization based support vector machine in
fault resistance [4]. Estimation of fault type and distance with combination with forward feature selection method. By using
impedance based technique in a transmission line is discussed in wavelet packet transform more number of features and better
[5–10]. In these schemes, single ended impedance measurement resolution is achieved. In scheme [29] one terminal current and
is used to estimate fault distance in a long transmission line. The voltage signal is analyzed and wavelet entropy criterion is applied
simulation results of these schemes show that due to large fault to reduce the size of feature vectors whereas in our proposed
resistance, estimation of fault type and distance error becomes method one terminal current signal is analyzed and wavelet
more. In a transmission line for estimation of fault type and energy and entropy is applied for pre-processing. Further in the
distance, the relationship between forward and backward waves proposed method, forward feature selection method is used to
travelling is the main theory behind travelling wave technique remove redundant features and to enhance the accuracy. It was
which has attracted widespread attention nowadays. These tech- observed from scheme [29] that fault classification accuracy was
niques estimate different type of fault and find the high impedance 99% and maximum fault location error was 0.74% whereas the pro-
fault in the transmission line almost accurately, but the sampling posed method in this paper says fault classification accuracy is
rate required is quite high (above 1 MHz) which is hard to imple- 99.21% and fault location error is >0.21%.
ment in practical field [11,12]. Travelling wave based distance Fault location of a transmission line using stationary wavelet
evaluation and fault classification in a long transmission line is transform in combination with determinant function feature
reported in [13–16]. In order to analyze the fault, these schemes (DFF), support vector machine (SVM) and support vector regres-
are based on correlation method to find the time difference sion (SVR) is discussed in [30]. The scheme in [30] uses single
between forward and backward wave. The methods discussed in end measurement and DFF to extract features. Also filtering is used
[13–16] gives less error for fault classification and distance evalu- in the scheme [30] to remove noise and decaying DC offset.
ation, however, they show the same pattern for fault near and at Simulation results of [30] show fault location error to be less,
the far end of the transmission line due to which it becomes quite however the instrumentation associated with it is quite complex.
difficult to identify and locate the fault. Fault detection, classification and location for transmission system
Nowadays, researchers are giving more emphasis on artificial with multigenerators applying discrete orthogonal stockwell trans-
intelligence based fault classification and distance estimation skills form is reported in [31]. In this scheme [31], synchronized current
such as neural network, fuzzy logic etc. because of its accuracy, self measurements from both ends of the transmission line is taken
adaptiveness and robustness to parameter variations. Fault classi- for fault analysis purpose. Also in the scheme [31] energy is
fication with fuzzy logic technique in a long transmission line is extracted as feature from the acquired signal and SVM is used as
reported in [17–20]. These schemes use wavelet transform (WT) fault locator. However the algorithm is quite complex and parame-
of the current signal to provide unseen fault data to the fuzzy logic ters of SVM are not optimized which leads to errors in fault analysis.
system for fault classification. In these schemes simple computa- This paper mainly focuses on two hybrid methodologies for
tional process is used, however the fault classification error estimation of fault type and distance in a long transmission line.
reported is quite large due to changes in simulation condition. The proposed method uses one cycle waveform of current which
Artificial neural network (ANN) is discussed in [21–28] for long is extracted from the sending terminal of the power system trans-
transmission line fault classification and distance evaluation. These mission line under study for fault classification and location.
schemes [21–28] use wavelet transform or wavelet packet trans- Thereafter current samples are pre-processed by wavelet packet
form (WPT) to extract distinctive features like energy and entropy transform and characteristics (also termed as features) like energy
from acquiring signals of voltage and current which are further and entropy are extracted from them. The best feature subset of
used in ANN for fault location and classification . The simulation the whole feature matrix is then selected by forward feature selec-
results show good accuracy, however the training time is quite tion technique during training. The data for training is generated
large due to which the task becomes quite complex and lethargic. by considering a variety of simulation condition like type of fault,
In the scheme [29] fault classification and location in a high voltage fault resistance, fault distance and fault inception angle. Further
power transmission line is proposed by using wavelet transform the feature set is scaled between [1, +1] which is then fed to
and support vector machine whereas in our proposed method, the support vector machine (SVM) for training the data and to
1370 P. Ray, D.P. Mishra / Engineering Science and Technology, an International Journal 19 (2016) 1368–1380

select the optimal features. The test data matrix is developed in an coefficient (a) through high and low pass filter. Also in case of
analogous way as training data matrix, but the operating condi- WPT, after the decomposition process for each level j 2 k, there
tions, taken is different in order to make the technique robust to are 2j numbers of node available [32]. Thus WPT gives more
parameter variation. Thereafter the test data set is validated in numbers of features, better frequency resolution, explores the
the trained SVM model for fault classification and location. Particle information content in the high frequency component and pro-
swarm optimization (PSO) technique is used to choose favorable vides a global view of the decomposed signal [32]. In the present
parameters of SVM. For fault classification, four SVC are considered work sampling frequency considered is 30 kHz, number of sample
where three of them are placed in the three phases and the fourth points is 600 per signal, so up to 4th level decomposition is per-
one is connected between phase and ground to detect ground formed which is based on Shannon’s entropy criterion of optimal
involvement. The sampling frequency taken for the whole process decomposition [33]. According to this criterion [33], at each level
is 30 kHz. The simulation results show that the proposed tech- entropy of the signal is calculated in order to find optimal
niques classifies and locates the fault in long transmission line fast decomposition. Signal is said to obtain optimal decomposition
and accurately as compared to approaches proposed by other when the entropy of parent level is higher than the total entropy
researchers. Separate investigation with the same proposed of decomposed level and a single piece of information is left to
method is also carried out with thyristor controlled series capaci- reconstruct the original signal.
tor (TCSC) placed at the middle of the transmission line. In order to minimize the dimension of huge data matrix,
The remaining sections of the paper are set as follows. Brief extraction of feature is used in pattern recognition where it con-
overviews of the wavelet packet transform and feature extraction verts the whole data matrix into feature matrix. In the present
process is focused on Section 2. In Section 3, the forward feature work two statistical features, i.e. energy and entropy are obtained
selection method is discussed. Section 4 reports an article on from the decomposed coefficient of WPT at each sub-band. In the
SVM and selection process of the parameters of SVM by the particle present work 16 decomposed WPT coefficients (32) are generated.
swarm optimization technique. In Section 5, the two proposed Energy is mathematically defined as [22,34,35]:
methods for fault location and classification are discussed in Z t2
details. Section 6 gives the simulation results for the estimation Eðt1 ; t 2 Þ ¼ ðjxðtÞjÞ2 dt ð1Þ
of fault type and distance. In Section 7, a comparison and discus- t1

sion with other researchers work is made. Section 8 reports fault where, signal is shown by x(t), signal energy from the time range
classification and location in a TCSC based transmission line with (t1, t2) by the symbol E. The value of transient energy signal is more
the same proposed method and Section 9 draws the conclusion. as compared to the value of undistorted or normal signal. Signal
information content is measured by entropy [34,35]. Additive infor-
2. Wavelet packet transform and feature extraction mation cost function [22] of x(t) signal is defined by entropy ‘EN’
such that E(0) = 0 and is defined as [22]:
Wavelet packet transform has popularly gained attention of X
ENðxÞ ¼ ENðxi Þ ð2Þ
researchers because in some cases critical data are placed in the i
high frequency component of the decomposed signal which needs
to be explored [22]. WPT is a generalization of discrete wavelet where, the decomposed coefficient of the signal x(t) is (xi). Entropy
transform (DWT) where the discrete time signal passes from a has large value of transient signal, whereas its value is small for
series of filters than DWT. The DWT does not give the best result normal signal. Here, a set of 32 features (2 statistical features  16
if small values of the signal are required since it is limited to WPT coefficients) is developed.
wavelet bases that increase in each step with a power of two. So,
in those cases another combination of bases is required which 3. Forward feature selection method
gives better result and thus WPT comes into the picture. In WPT
the signal (S) goes through a series of filters (low and high pass Feature selection is the process which chooses features which
filter) and simultaneously approximation coefficients (a) (low correctly correlates the target and removes redundant ones. This
frequency) and detail coefficients (d) (high frequency) are formed. method is important to implement as a large number of redundant
Thereafter the low (a) and high (d) coefficients are decomposed features are used as input in several cases of supervised learning
recursively up to level k to make the total decomposition structure tasks which increases the computational burden and gives
which is given in Fig. 1 whereas in case of DWT each step of the erroneous output which needs to be removed [36]. The feature
process is found by passing only the preceding low frequency selection method selects favorable features (those which are able

Fig. 1. Decomposition structure of WPT.

 P. Ray, D.P. Mishra / Engineering Science and Technology, an International Journal 19 (2016) 1368–1380 1371

to predict the target properly) from the total matrix [36]. Feature Support vector classification (SVC) is used for classification pur-
selection is a robust greedy algorithm which avoids over fitting pose and support vector regression (SVR) for fault location.
[36]. Present work focuses on forward feature selection method In the present work, SVM parameters are explored by integrat-
where computation is performed in each step iteratively to choose ing software called LIBSVM [43] and optimal values are evaluated
favorable features (those showing highest scores) thus developing by (PSO) particle swarm optimization technique. The additional
a subset of inputs and removing the redundant features [37]. In values taken from LIBSVM are gamma parameter (g) and soft
this paper forward feature selection method is applied. Evaluation parameter or cost parameter (c). The tradeoff between forced, rigid
function for the present used feature selection method is leave one margin and train error is given by soft parameter or cost parameter
out (LOO) mean square error (MSE) of the k-nearest-neighbor (c) [44] and the radius and shape of the hyper plane is controlled
(KNN) estimator which provides an excellent estimate of the by gamma parameter (g) [44]. Also by increasing gamma
expected generalization error [38,39]. The weighted average of near- parameter, the number of support vector increases [44]. The most
est neighbor is defined as KNN estimator, where every neighbor’s favorable value of SVM parameter is determined by PSO, which is
weight is proportional to its proximity [39] and the definition of eval- as shown by flowchart in Fig. 2. The fitness value f for the PSO
uation function is the negative (halved) MSE of the weighted KNN is assumed as the residual mean square value (MSE) which is
estimator [39]. The locally optimal weight vector is searched by the represented mathematically as:
evaluation function which produces scores to the weight vector (w) vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N
over the features [39]. Further rank is given by the resulting weight u1 X
f ¼t
2
to each feature which makes a subset of the best features.
½yK ðkÞ  yðkÞ ð3Þ
N k¼1

4. Support vector machine where, y(k) is the actual discrete signal, the SVM output predictor is
yK(k) and the discrete samples is denoted by N. The most favorable
Statistical learning concept with an adaptive computational values of SVM parameters are selected by PSO during the training
learning method is defined as support vector machine (SVM). This process. In the present paper, nu-svr has been considered for fault
learning technique uses input vectors to map nonlinearly into a location and nu-svc for fault classification where an adaptable
feature space whose dimension is high [40]. To make the most of regularization parameter m (nu) is present to adjust the input data.
the capability of the fault classifier and locator, the optimal hyper Also, lower border on the fraction of support vectors and upper
plane is determined [40]. Training algorithm of an SVM fault border on the fraction of margin errors is done by the SVM
classifier for a given train data set which belongs to one of the parameter m (nu) [41]. Features lying near the boundary are called
two categories of the target variable [1, 1] builds a model which support vectors. The parameter m (nu) determines the loss function
is shown by space mapped features where the other features are (e) by adapting the error model. Brief algorithm of nu-SVR is
categorized by a transparent broad gap [40]. The two categories discussed in [45]. Nu-SVC uses a parameter nu (n) to control the
are separated by a gap which is called a hyper plane. The present number of support vectors and training errors. This parameter is
work uses a radial basis function (RBF) as kernel parameter which sometimes symbolizes as m 2 ð0; 1. This parameter nu is a lower
maximizes the gap between the two categories, thereby making bound on the fraction of support vectors and upper bound on the
the hyper plane optimal [40]. Further the test data set features fraction of training errors. Nu (n) is a parameter of Nu-SVC whose
are mapped into that same hyper plane and predicted by the default value is 0.5.
trained SVM model [40]. The merits of SVM are it does not
converge into local minima, prone to overfitting, sparse and gives 5. Proposed technique for fault classification and location
a global solution. Selection of proper SVM parameters is very
important for good generalization performance and high accuracy Hybrid SVM based scheme for estimation of short circuit fault
in fault location and classification of transmission line [41,42]. type and distance in a long transmission line is discussed in this

Establishing training data set

Initialization (setting swarm size , c1, c2, wmin, wmax),

value of position & velocity and Number of iterations

PSO Parameter selection of SVM Train SVM

NO Evaluate fitness
Parameter with best fitness achieved
of each data in
and stopping criterion met ?
the train set

YES
Obtain
optimized
parameter of
SVM

Fig. 2. Flow chart for selection process of favorable value of SVM parameter by PSO.
1372 P. Ray, D.P. Mishra / Engineering Science and Technology, an International Journal 19 (2016) 1368–1380

Pre- Best
Sending feature
processing Fault
end current selection SVC
of the Classification
measurem acquired and Data generation
ent of normalizati
current signal
transmissio on
by wavelet
n line
packet Fault
under SVR
transform Location
study

Fig. 3. Block diagram of SVM based fault classification and location in a transmission line.

Support vector machine

(SVC-p) in phase a, b,
SVC- ya
c
a
current
measurement
from sending Signal Best feature
decomposition and selection by forward SVC-
end of the yb
transmission line feature extraction feature selection b
under study by WPT method

SVC-
yc
c

SVC-g as yg
ground detector

Fig. 4. Detailed layout of support vector fault classifier.

paper which is shown in Fig. 3. Post fault sending end current the index value becomes more than the threshold value and it is
waveform with one cycle is used to analyze the fault type and carried out in parallel with phase identification. The detailed struc-
distance in a long transmission line. The acquired samples of the ture of fault classifier is given in Fig. 4.
sending end current signal are divided into a large range of fre-
jIa þ Ib þ Ic j
quency sub-bands using WPT. Thereafter, from the decomposed Current index ¼ ð4Þ
coefficient, energy and entropy are extracted. The total feature
mean ðjIa j; jIb j; jIc jÞ
set consists of 32 features (16 WPT coefficients  2 features). Fur- where, Ia, Ib, Ic are the instantaneous values of current signal. In case
ther to scale feature set, normalization process is done between of faults classification, the performance criterion considered is
[1, +1] so that it can be compared appropriately. The data is then classification accuracy [47] which is defined as
generated for training and testing purpose considering a variety of
simulated conditions like the fault resistance, fault inception angle, Accurate fault classification
Classification accuracy ¼  100 ð5Þ
type of short circuit fault and fault distance. Now to develop the Number of samples tested
method insensitive to parameter variations, the simulation condi- The description of the discussed fault distance estimation
tion for generating train data matrix is made totally apart from the scheme is presented by the flowchart in Fig. 5. For estimating the
test data matrix. From the total feature set, some of the features fault distance, the performance criterion taken is absolute error
don’t predict the output properly. As a result the prediction accu- and the mean error. Mean error is defined in (6) and absolute error
racy reduces. So to improve the accuracy, redundant features are [46] is defined in (7).
removed from the total data set by applying forward feature selec-
tion technique during training. By using this feature selection je1 þ e2 þ e3 þ    en j
Mean error ¼  100 ð6Þ
method, the total feature set to be fed to the PSO based SVM is n
reduced, which in turn simplifies the process and makes it fast. where, e1, e2, en are the ‘n’ number of absolute errors.
The optimal feature set with the test data are then fed to the
trained SVM model for prediction purpose. The layout of the dis- jPr edicted fault location  Exact fault locationj
Absolute error ¼
cussed fault classification technique is presented in Fig. 4. In Total length of the line
Fig. 4, three SVC is placed in phase a, b and c to classify phase fault  100
and the fourth SVC is connected between phase and ground to
ð7Þ
detect ground fault. The output of SVC placed in phases a, b and
c are either ‘+1’ or ‘1’ denoting faulted phase. Often double line The transmission system under study in the present work is a
to ground (LLG) fault is mis-classified as line-to-line fault (LL) by 300 km long transmission line with 400 kV source at both ends
SVC [46]. So to overcome the difficulty a separate SVC is placed and 50 Hz system frequency which is as shown in Fig. 6. In Fig. 6
between phase and ground where a zero sequence current based the sending or relaying end voltage is denoted by Es and the receiv-
indicator proposed in [47] is applied as an index value as shown ing end voltage is denoted by Er. Two voltage sources placed at the
in (4) and a threshold value is set by trial and error. In this paper, front and rear ends of the transmission line are represented as
threshold value taken is 0.05. Ground detection is indicated when ideal one with its internal impedance. Appendix A gives the detail
 P. Ray, D.P. Mishra / Engineering Science and Technology, an International Journal 19 (2016) 1368–1380 1373

Faulty current signal

Signal decomposition and feature

extraction by WPT

Best feature selection by forward

feature selection method and
normalization

Train and test data generation Fig. 7b. Pre-fault and post-fault current signal for a-b fault.

SVR trained with the best feature set

Testing with the trained SVR model

Fig. 5. Flowchart of fault location technique.

400 kV voltage
source 300 km
Er
Es
Fault Fig. 7c. Pre-fault and post-fault current signal for ab-g fault.
Receiving end
Sending end/
Relaying end

Fig. 6. Long transmission line under study.

Fig. 7d. Pre-fault and post-fault current signal for abc fault.

from the decomposed WPT coefficients. Thus the total feature

Fig. 7a. Pre-fault and post-fault current signal for a-g fault. matrix has 32 features (16 WPT coefficients  2 features) which
are then normalized between [1, +1]. As Daubechies works well
in transient data, so it is taken for further analysis. Some of the
description of parameters of voltage sources and the transmission WPT coefficients of fourth level are shown in Fig. 8. Thereafter by
line. Distributed model of the transmission line is considered for using the forward feature selection method, 2 features, among 32
analysis. Single cycle prefault and postfault current signal acquired features as given in Table 1 are selected as best feature which
by the relaying end is reported in Fig. 7 for four types of fault (a-g, can predict the target properly. Further, these two features as
a-b, ab-g, abc). In the present work, a single cycle of post fault shown in Table 1 is taken for the entire task of fault classification
current is used for fault classification and position from the send- and location. The best feature plot and redundant feature plot
ing end of the transmission line at 30 kHz sampling frequency. are given in Figs. 9a–9c. It is noticed from Figs. 9a and 9b that
Sampling frequency considered is 30 kHz as it is noticed after a the best feature gives a well defined path and for each fault
series of analysis that the simulink model of the studied system distance it has some information, whereas from Fig. 9c which is a
responds better and produces more accurate results in 30 kHz than redundant feature plot, it is observed that the valuable information
any other value. Then the collected current signal is further decom- is lost as it shows an erratic path. During the training process
posed to 4-levels by WPT and further energy and entropy are taken optimal features are obtained.
1374 P. Ray, D.P. Mishra / Engineering Science and Technology, an International Journal 19 (2016) 1368–1380

Fig. 8. Some of the fourth level WPT coefficients.

Table 1
Best feature by WPT.

Signal type Feature (02) Best coefficient

Current Energy ADAD4
Entropy DDDD4

Fig. 9c. Non-optimal feature plot.

Table 2
Parameters to develop, train and test data set.

Data-set Fault resistance (Rf) (in X) Fault inception angle (h) (in
degree)
Train 0, 1, 5, 10, 20, 40, 50, 70, 100, 10°, 20°, 30°, 40°, 50°, 60°, 70°,
Fig. 9a. Optimal feature plot of coefficient ADAD4 of energy of current signal. data 150 80°
Test data 2, 9, 25, 45, 65, 85, 110, 140 5°, 11°, 17°, 24°, 45°, 65°, 90°

Further the trained SVM with the two optimal features are used
for testing purpose. The train and test data are developed by taking
into consideration a variety of simulation condition like the fault
inception angle (h) and fault resistance (Rf) as shown in Table 2
for each 1 km of 300 km long transmission line in case of ten differ-
ent categories of short circuit fault (a-g, b-g, c-g, a-b, b-c, c-a, ab-g,
bc-g, ca-g, abc). From Table 2 it is noticed that the parameters to
develop a train matrix is completely different from the test
parameter, to make the planned method insensitive to parameter
variations. Thus the total train data set consists of 240,000 data
samples (10 types of fault resistance  8 types of fault inception
Fig. 9b. Optimal feature plot of coefficient DDDD4 of entropy of current signal. angle  300 fault distances  10 short-circuit fault). The test data
 P. Ray, D.P. Mishra / Engineering Science and Technology, an International Journal 19 (2016) 1368–1380 1375

Table 3
Best value of SVM by PSO.

SVM parameters For fault classification For fault distance estimation

For fault detection of the ground For phase fault detection
Kernel type Radial basis function Radial basis function Radial basis function
Gamma (g) 0.4 0.52 0.63
Cost (c) Not used Not used 12.4
Nu (nu) 0.5 0.45 0.15

Table 4 6. Simulation results

Comparison of mother wavelet test result.
Initial testing has been carried out for fault classification and
Different order For fault classification For fault distance
of Daubechies estimation location to find the best order of Daubechies (dB) mother wavelet
mother wavelet Fault classification Maximum Mean
with the generated data sets in order to proceed the fault analysis
accuracy (%) absolute error (%) task with the best one. An evaluation of the performance criterion
error (%) for fault classification and location is made with four types of Dau-
dB1 92.7 1.02 0.34 bechies mother wavelet which is given in Table 4. It is observed
dB2 95.2 0.65 0.25 from Table 4 that dB4 shows better fault classification accuracy
dB3 94.3 0.88 0.30 (99.21%) and least fault location error (>0.21%) as compared to
dB4 99.21 0.20 0.10
other mother wavelets. So dB4 is selected as mother wavelet for
estimation of fault type and distance. Also, Daubechies wavelet
has maximum number of vanishing moment. The wavelet’s ability
to represent information in a signal is limited by vanishing
matrix consists of 168,000 data samples (8 types of fault resis- moments. Vanishing moments for each wavelet is equal to half
tance  7 types of fault inception angle  300 fault distances  10 the number of coefficients. dB1 has one vanishing moment and
types of fault). The test data set is taken as 70% of the train data set. encodes polynomial of one coefficient, dB2 has two vanishing
The favorable values of SVM parameter are found by PSO, which is moments and encodes linear and constant signal components,
given in Table 3. The parameters of PSO used in this paper are dB3 has three vanishing moments and encodes quadratic, linear
reported in Appendix A. The test result of fault distance error is and constant signal components. Due to the nature of the transient
shown in the present work by box-plot. Box-plot is a graphical complex data and vanishing moments, results of dB3 is not linear.
representation to show the error. In box-plot, the upper quartile In Table 4, bold portions indicate that daubechies (dB4) gives
indicates maximum error, the lower quartile shows minimum highest fault classification accuracy and minimum fault location error.
error, the area within the box indicates maximum error lying It is observed from Table 5 that the average of classification
within the range of the box and the middle band shows mean or accuracy is quite high (99.21%). The total test results of fault dis-
average error. tance estimation by the present discussed scheme is shown by

Table 5
Test results of fault classification.

Fault type No. of test data samples No. of test samples classified correctly No. of test samples misclassified Classification accuracy (%)
LG (a-g, b-g, c-g) 50,400 49,855 545 98.91
LL (a-b, b-c, c-a) 50,400 50,100 300 99.40
LLG (ab-g, bc-g, ca-g) 50,400 49,970 430 99.14
LLL (abc) 16,800 16,750 50 99.70
Total 168,000 166,675 1325 99.21

Fig. 10. Box plot of test results for ten types of fault.
1376 P. Ray, D.P. Mishra / Engineering Science and Technology, an International Journal 19 (2016) 1368–1380

Table 6
Test results of fault location method.

Fault No. of Minimum Maximum Mean fault Range of the

type samples absolute absolute distance box (error
error (%) error (%) error (%) range)
a-g 16,800 0.00052 0.18 0.10 0.013–0.148
b-g 16,800 0.0027 0.17 0.08 0.022–0.17
c-g 16,800 0.002 0.19 0.10 0.02–0.18
a-b 16,800 0.00021 0.15 0.08 0.014–0.13
b-c 16,800 0.02 0.15 0.07 0.028–0.11
c-a 16,800 0.007 0.14 0.07 0.02–0.12
ab-g 16,800 0.001 0.20 0.10 0.006–0.19
bc-g 16,800 0.006 0.20 0.10 0.041–0.18
ca-g 16,800 0.0012 0.19 0.10 0.013–0.17
abc 16,800 0.0048 0.12 0.04 0.02–0.10

Fig. 11b. Actual versus predicted distance plot for AB fault.

Fig. 11a. Actual versus predicted distance plot for AG fault.

box-plot in Fig. 10 and further analyzed in Table 6. From Fig. 10,

the observations noticed are presented in Table 6. It is observed
from Table 6 that maximum absolute fault distance error is
>0.21% and average error is also >0.11% and most of the error lies
in the range 0.02–0.10%. The training time of the proposed fault
Fig. 11c. Actual versus predicted distance plot for ABG fault.
classification and location method is 0.15 s. For a specific case of
fault resistance (Rf = 45 X), fault inception angle (h = 65°), six
selected fault distance and four types of short circuit fault (a-g located very near to both the ends of the transmission line. It can
a-b, ab-g, abc), Fig. 11 shows plots of the predicted output (SVM be observed from Table 9 that all the errors are >0.21% and error
output) versus actual fault location. It is noticed from Fig. 11 that very near to the source end is more. In order to show the superior-
>0.21% of maximum absolute fault distance error is there. ity of SVM as fault classifier, it is compared with fault classifier’s
Separate investigation is also carried out to find a suitable like artificial neural network (ANN), probabilistic neural network
sampling frequency (fs). In this regard, a series of high sampling (PNN) and adaptive neural fuzzy inference system (ANFIS) in
frequencies (<30 kHz) and low sampling frequencies are consid- Table 10. The algorithm and other parameters as discussed in
ered which is given in Table 7. It is shown from Table 7 that the Section 5 remains the same for ANN, PNN and ANFIS for fault
data acquired in this paper give the best result for fault classifica- classification. The parameters of ANN, PNN and ANFIS used in this
tion and location with 30 kHz sampling frequencies. So, it is paper are given in Appendix. It can be observed from Table 10 that
considered for further fault analysis. Also in order to show the highest fault classification accuracy is provided by the proposed
importance of PSO optimization technique based SVM, a compar- fault classifier i.e. SVC (99.21%).
ison has been done in Table 8 with and without PSO for fault In Table 8, bold portions shows that with optimized parameter
classification and fault location. It can be observed from Table 8 of SVM with PSO, fault classification accuracy is more and fault
that with optimized parameter of SVM with PSO in the proposed location error is less as compared to without optimized parameter
algorithm, fault classification accuracy (99.21%) and fault location of SVM.
error (>0.20%) is much better than without optimized parameter
of SVM. The test results of fault location with and without opti-
mized parameter of SVM in the proposed algorithm are shown in 7. Background
Figs. 10 and 12. For a particular fault inception angle (h = 24°)
and fault resistance (Rf = 65 X), Table 9 shows some selected fault In order to prove that the present discussed scheme for fault
distances for four types of fault (a-g, a-b, ab-g, abc) which are classification and distance estimation gives better accuracy as
 P. Ray, D.P. Mishra / Engineering Science and Technology, an International Journal 19 (2016) 1368–1380 1377

Table 9
Fault location test results for distances very near to source end of transmission line.

Actual fault location (km) Absolute error (%)

AG fault AB fault ABG fault ABC fault
2 0.19 0.15 0.20 0.12
4 0.18 0.14 0.18 0.11
6 0.17 0.11 0.17 0.10
8 0.15 0.10 0.16 0.09
294 0.17 0.13 0.18 0.09
296 0.18 0.14 0.19 0.10
298 0.19 0.14 0.20 0.11

Table 10
Comparison of different fault classifiers.

Fault classifier Classification accuracy (%)

ANN 96
PNN 97
ANFIS 89
Proposed one (SVC) 99.21
Fig. 11d. Actual versus predicted distance plot for ABC fault.

Table 11
Comparison of different methods.
Table 7
Test results with different sampling frequencies. Schemes Fault classification Fault location
error (%)
Sampling frequency (kHz) Fault classification Fault location No. of test Classification
Classification accuracy (%) error (%) samples accuracy (%)

0.1 93.5 <0.9 Method in [28] – – <0.30

0.3 90.7 <1.0 Method in [48] 28,800 99.11 <0.45
50 92.4 <1.5 Method in [49] 200 97.2 –
100 80.5 <2.7 Method in [50] – – <0.90
Proposed method 99.21 >0.21 Method in [51] – – <1.0
Proposed method 168,000 99.21 >0.21

Table 8
Test result with and without PSO. Sending end
150 km
For fault For fault distance
classification estimation ES TCSC Er
150 km
Fault Maximum Mean
classification absolute error Fault
Voltage Voltage
accuracy (%) error (%) (%)
Source-1 Source-2
With optimized parameter of SVM 99.21 0.20 0.10
with PSO Fig. 13a. TCSC based transmission line under study.
Without optimized parameter of 95.01 0.32 0.22
SVM
Capacitor (C)

Thyristor pairs Triggered air

gap
MOV

Inductor (L)

Fig. 13b. TCSC details.

compared to other researcher’s work, a comparison has been

made in Table 11. It can be seen from Table 11 that the present
discussed scheme gives better accuracy for fault classification
(99.21%) and minimum fault location error (>0.21%) as compared
to methods used by other researchers. Also the training time of
the proposed method is quite small (0.15 s). So, the proposed
method is suggested for estimation of fault type and distance in
Fig. 12. Box plot of test results for ten types of fault without PSO. a long transmission line.
1378 P. Ray, D.P. Mishra / Engineering Science and Technology, an International Journal 19 (2016) 1368–1380

Table 12 8. Fault classification and location with TCSC based

Best feature in case of TCSC based transmission system by FFS.
transmission line
Signal type Feature (07) Best coefficient
Current Energy AAAA4, ADAD4, AADA4 Fault classification and location of a 300 km transmission line
Entropy ADDA4, AADD4, DADA4, DDDA4 with 400 kV source at both ends and a thyristor controlled series
capacitor (TCSC) placed at the middle is discussed in this section.
The model under study is as shown in Fig. 13a and its detailed
parameters are given in Appendix A. In Fig. 13a, Es and Er indicates
front and rear end ideal voltage sources and the TCSC consists of an
antiparallel connection of thyristors in each phase and a series
combination of a reactor in parallel with a capacitor. A metal oxide
varistor (MOV) with a parallel air gap arrangement protects the
capacitor from overvoltage which is shown in Fig. 13b. The protec-
tion level of the MOV was adjusted to 721.81 kV based on minimal
voltage across the capacitor which is 2.5 times when a standard
current of 2 kA is flowing through it [35]. Post fault single cycle
sending end current signal is acquired to classify and locate the
ten types of short circuit fault (a-g, b-g, c-g, a-b, b-c, c-a, ab-g,
Fig. 14a. Optimal feature plot for TCSC based transmission line. bc-g, ca-g, abc) in the transmission line under study with a sam-
pling frequency of 30 kHz. The same technique for fault classifica-
tion and location as proposed in Section 5 is implemented in this
section. The other parameters of WPT, SVM and FFS method are
the same as mentioned in Section 5. Here, seven optimal features
are selected by FFS from a total of 32 feature which is shown in
Table 12. One of the optimal and non-optimal feature plot is shown
in Figs. 14a and 14b respectively in order to verify that optimal fea-
ture gives a distinct path from which information can be extracted
whereas non-optimal feature is random and unpredictable in nat-
ure. The parameters to generate the train and test data set are the
same as mentioned in Table 2 of Section 5. The test results of fault
classification are given in Table 13 and it can be observed that fault
classification accuracy for all test cases is 98.36%. The test results of
Fig. 14b. Non-optimal feature plot for TCSC based transmission line.

Table 13
Test results of fault classification for TCSC based transmission line.

Fault type No. of test data samples No. of test samples classified correctly No. of test samples misclassified Classification accuracy (%)
LG (a-g, b-g, c-g) 50,400 49,392 1008 98.00
LL (a-b, b-c, c-a) 50,400 49,745 655 98.70
LLG (ab-g, bc-g, ca-g) 50,400 49,443 957 98.10
LLL (abc) 16,800 16,673 127 99.24
Total 168,000 165,253 2747 98.36

Fig. 15. Test results of fault location of TCSC based transmission line.

Table 14
Test results of fault location for TCSC based transmission line.

Type of fault No. of test data samples Minimum absolute error (%) Maximum absolute error (%) Mean error (%) Range of the box (%)
LG (a-g, b-g, c-g) 50,400 0.00048 0.24 0.11 0.05–0.15
LL (ab, bc, ca) 50,400 0.006 0.10 0.02 0.01–0.05
LLG (ab-g, bc-g, ca-g) 50,400 0.0005 0.27 0.15 0.10–0.25
LLL (abc) 16,800 0.006 0.12 0.02 0.01–0.07
 P. Ray, D.P. Mishra / Engineering Science and Technology, an International Journal 19 (2016) 1368–1380 1379

fault location for a particular case of fault inception angle (h = 45°) A.3. Particle swarm optimization parameters
and fault resistance (Rf = 85 X) is shown as box-plot in Fig. 15. The
analysis of Fig. 15 is given in Table 14 from which it can be c1 = 4, c2 = 4, particle size = 50, No. of iteration = 1000.
observed that maximum absolute error is >0.28% and mean error wmin = 0.5, wmax = 0.9.
is >0.15% which is acceptable.
In Table 14, bold portions shows that LG fault has minimum A.4. Parameters of ANN
absolute error among all other faults (LL, LLG and LLL) and LLG
fault has maximum absolute error and maximum mean error as Parameters of ANN is given in Table A.1.
compared to their faults (LG, LL, LLL).
A.5. Parameters of PNN and ANFIS
9. Conclusion
Kernel function used in PNN: Radial basis function.
Support vector machine based estimation of fault type and
distance scheme in a long transmission line is proposed. For ten Spread factor ðrÞ ¼ 0:025
types of short circuit fault event, the proposed technique gives ANFIS generates a Sugeno-type fuzzy inference system (FIS)
quick, correct and robust fault classification and location assess- using subtractive clustering technique with a radius of 0.5.
ment of the collected one cycle post fault current signal. The
uniqueness of the proposed technique is that it uses transient
References
data to analyze the fault, a large number of features are collected
by wavelet packet transform, forward feature selection method is [1] R. Das, D. Novosel, Review of fault location techniques for transmission and
applied to remove redundant features thereby enhancing the sub-transmission lines, in: Proc. 54th Annual Georgia Tech Protective Relaying
Conf., 2000.
prediction accuracy, optimized value of support vector machine
[2] M.E. Hawary, Electrical Power Systems, IEEE press, 1995. pp. 469–536.
is used, variety of simulation conditions are considered to develop [3] M.M. Saha, J. Izykowski, E. Rosolowski, Fault Location on Power Networks,
train and test data matrix and the simulation condition to develop Springer Pub., 2010.
[4] Y. Tang, H.F. Wang, R.K. Aggarwal, A.T. Johns, Fault Indicators in transmission
test data matrix is made totally apart from the train one to make
and distribution systems, in: Proc. Int. Conf. Electric Utility Deregulation and
the proposed technique insensitive to parameter variations. The Restructuring and Power Technologies 2000, City University, London, 2000, pp.
simulation result shows maximum fault classification accuracy 238–243. 4–7 April.
(99.21%) and minimal fault position error (>0.21%) as compared [5] D. Novosel, B. Bachmann, D.G. Hart, Y. Hu, M.M. Saha, Algorithms for locating
faults on series compensated lines using neural network and deterministic
to schemes proposed by other researchers. Also the training time methods, IEEE Trans. Power Delivery 11 (4) (1996) 1728–1736.
is small (0.15 s) with SVM. The proposed method is also tested [6] T. Adu, A new transmission line fault locating system, IEEE Trans. Power
with long transmission line having TCSC placed in the middle. Delivery 16 (4) (2001) 498–503.
[7] T. Takagi, Y. Yamakoshi, M. Yamaura, R. Kondou, T. Matsushima, Development
Simulation results of TCSC based model give >0.29% fault location of a new type fault locator using the one-terminal voltage and current data,
error and fault classification accuracy of 98.36%. IEEE Trans. Power Appl. Syst. PAS-101 (8) (1982) 2892–2898.
[8] L. Eriksson, M.M. Saha, G.D. Rockefeller, An accurate fault locator with
compensation for apparent reactance in the fault resistance resulting from
Appendix A. remote-end infeed, IEEE Trans. Power Appl. Syst. PAS-104 (2) (1985) 424–436.
[9] S. Guobing, S. Jiale, G. Yaozhong, An accurate fault location algorithm for
A.1. Parameters of the system under study parallel transmission lines using one terminal data, Int. J. Electr. Power Energy
Syst. 31 (2–3) (2009) 124–129.
[10] X. Lin, H. Weng, B. Wang, A generalized method to improve the location
(i) Receiving and sending end voltage source parameter [35]: accuracy of the single-ended sampled data and lumped parameter model
Positive sequence impedance (Z1): 1.31 + j 16.0 X. based fault locators, Int. J. Electr. Power Energy Syst. 31 (5) (2009) 201–205.
[11] L. Eriksson, M.M. Saha, G.D. Rockefeller, An accurate fault locator with
Zero sequence impedance (Z0): 2.22 + j 27.6 X. compensation for apparent reactance in the fault resistance resulting from
Frequency of the system: 50 Hz. remote-end feed, IEEE Trans. Power Appl. Syst. PAS 104 (2) (1985) 1424–1436.
(ii) Parameter of long transmission line [35]: [12] X. Dong, W. Kong, T. Cui, Fault classification and faulted phase selection based
on the initial current travelling wave, IEEE Trans. Power Delivery 24 (2) (2009)
Length: 300 km, voltage: 400 kV. 552–559.
Impedance of positive sequence = 8.15 + j 94.5 X. [13] F. Chen, G. Qian, F. Wang, Study on traveling wave differential protection for
Impedance of zero sequence = 92.5 + j 308 X. series compensated line, J. Int. Council Electr. Eng. 1 (3) (2011) 359–366.
[14] E.H. Shehab-Eldin, P.G. McLaren, Travelling wave distance protection-problem
Positive sequence capacitance = 14 nF/km. areas and solutions, IEEE Trans. Power Delivery 3 (3) (1998) 894–902.
Zero sequence capacitance = 7.5 nF/km. [15] E.E. Ngu, K. Ramar, A combined impedance and travelling wave based fault
location method for multi-terminal transmission lines, Int. J. Electr. Power
Energy Syst. 33 (10) (2011) 1767–1775.
A.2. Details of TCSC parameter [35] [16] K. Liao, Z. He, X. Li, A fault location method based on traveling wave natural
frequency used on ±800 kV UHVDC transmission lines, in: Proc. IEEE Electrical
and Control Engineering (ICECE), 2011, pp. 5652–5655. 16–18 September.
L = 61.9 mH, C = 21.977 lF. [17] M.J. Reddy, D.K. Mohanta, A comparative study of artificial neural network
(ANN) and fuzzy information system (FIS) approach for digital relaying of
transmission line faults, AIML J. 6 (4) (2006) 1–7.
[18] M.J. Reddy, D.K. Mohanta, A wavelet-fuzzy combined approach for
Table A.1 classification and location of transmission line faults, Int. J. Electr. Power
Parameter details of ANN. Energy Syst. 29 (9) (2007) 669–678.
[19] A.K. Pradhan, A. Routray, S. Pati, D.K. Pradhan, Wavelet fuzzy combined
Network type Feed-forward back propagation network approach for fault classification of a series compensated transmission line,
IEEE Trans. Power Delivery 19 (4) (2004) 1612–1618.
Training function Levenberg–Marquardt [20] O.A.S. Youssef, Combined fuzzy logic wavelet based fault classification
Size of first hidden layer 40 technique for power system relaying, IEEE Trans. Power Delivery 19 (2)
Size of second hidden layer 04 (2004) 582–589.
Size of input layer Depends on the size of the optimal feature set [21] T. Dalstein, B. Kulicke, Neural network approach to fault classification for high
Size of output layer 01 speed protective relaying, IEEE Trans. Power Delivery 10 (2) (1995) 1002–
Train parameter goal 10e9 1011.
Performance function MSE(mean squared error) [22] S. Ekici, S. Yildirim, M. Poyraz, Energy and entropy based feature extraction for
No. of epochs 5000 locating fault on transmission lines by using neural network and wavelet
packet decomposition, Expert Syst. Appl. 34 (4) (2008) 2937–2944.
1380 P. Ray, D.P. Mishra / Engineering Science and Technology, an International Journal 19 (2016) 1368–1380

[23] S. Ekici, S. Yildirim, M. Poyraz, A transmission line fault locator based on Elman [38] L. Devroye, The uniform convergence of nearest neighbor regression function
recurrent networks, Appl. Soft Comput. 9 (1) (2009) 341–347. estimators and their application in optimization, IEEE Trans. Inf. Theory 24 (2)
[24] T. Bouthiba, Fault location in EHV transmission lines using artificial neural (1978) 142–151.
networks, Int. J. Appl. Math. Comput. Sci. 14 (1) (2004) 69–78. [39] A. Navot, L. Shpigelman, N. Tishby, E. Vaadia, Nearest neighbor based feature
[25] A.K. Pradhan, P.K. Dash, G. Panda, A fast and accurate distance relaying scheme selection for regression and its application to neural activity, in: Proc. Int. Conf.
using an efficient radial basis function neural network, Electr. Power Syst. Res. Advances in Neural Information Processing Systems NIPS, 18, 2006, pp. 995–
60 (1) (2001) 1–8. 1002.
[26] H.K. Zadeh, M.R. Aghaebrahimi, A novel approach to fault classification and [40] U.B. Parikh, B. Das, R. Maheshwari, Fault classification technique for series
fault location for medium voltage cables based on artificial neural network, compensated transmission line using support vector machine, Int. J. Electr.
World Acad. Sci. Eng. Technol. 18 (2008) 1100–1103. Power Energy Syst. 32 (6) (2010) 629–636.
[27] S.R. Samantaray, P.K. Dash, G. Panda, Fault classification and location using HS- [41] en.wikipedia.org/wiki/Support_vector_machine.
transform and radial basis function neural network, Electr. Power Syst. Res. 76 [42] K. Schittkowski, Optimal parameter selection in support vector machines, J.
(9–10) (2006) 897–905. Ind. Manage. Optim. 1 (4) (2005) 465–476.
[28] J. Sadeh, N. Hadjsaid, A.M. Ranjbar, R. Feuillet, Accurate fault location [43] LIBSVM-A library for support vector machines (online). Available: <http://
algorithm for series compensated transmission lines, IEEE Trans. Power www.csie.ntu.edu.tw/~cjlin/libsvm>.
Delivery 15 (3) (2000) 1027–1033. [44] <http://scikit-learn.org/dev/auto_examples/svm/plot_rbf_parameters.html>.
[29] S. Ekici, Support vector machines for classification and locating faults on [45] I. Steinwart, On the optimal parameter choice for m-support vector machines,
transmission lines, Appl. Soft Comput. 12 (6) (2012) 1650–1658. IEEE Trans. Pattern Anal. Mach. Intell. 25 (10) (2003) 1274–1284.
[30] A.A. Yusuff, A.A. Jimoh, J.L. Munda, Fault location in transmission lines based [46] P.K. Dash, A.K. Pradhan, G. Panda, A novel fuzzy neural network based distance
on stationary wavelet transform, determinant function feature and support relaying scheme, IEEE Trans. Power Delivery 15 (3) (2000) 902–907.
vector regression, Electr. Power Syst. Res. 110 (2014) 73–83. [47] M. Akke, J.T. Thorp, Some improvements in the three-phase differential
[31] M.J. Reddy, P. Gopakumar, D.K. Mohanta, A novel transmission line protection equation algorithm for fast transmission line protection, IEEE Trans. Power
using DOST and SVM, Eng. Sci. Technol. Int. J. 19 (2016) 1027–1039. Delivery 13 (1) (1998) 66–72.
[32] B.K. Panigrahi, V.R. Pandi, Optimal feature selection for classification of power [48] V. Malathi, N.S. Marimuthu, S. Baskar, K. Ramar, Application of extreme
quality disturbances using wavelet packet-based fuzzy k-nearest neighbor learning machine for series compensated transmission line protection, Eng.
algorithm, IET Gener. Transm. Distrib. 3 (3) (2009) 296–306. Appl. Artif. Intell. 24 (5) (2011) 880–887.
[33] <www.mathworks.in/help/wavelet/ug/wavelet-packets.html>. [49] P.K. Dash, S.R. Samantaray, G. Panda, Fault classification and section
[34] P. Ray, D. Mishra, Application of extreme learning machine for underground identification of an advanced series compensated transmission line using
cable fault location, Int. Trans. Electr. Energy Syst. (2014), http://dx.doi.org/ support vector machine, IEEE Trans. Power Delivery 22 (2007) 67–73.
10.1002/etep.2032.25(7). Wiley. [50] S. Sahoo, P. Ray, B.K. Panigrahi, N. Senroy, A computational intelligence
[35] P. Ray, B.K. Panigrahi, N. Senroy, Hybrid methodology for fault distance approach for fault location in transmission lines, in: Proc. IEEE Conf. Power
estimation in series compensated transmission line, IET Gener. Transm. Electronics, Drives and Energy Systems (PEDES-2010), 2010, pp. 1–6.
Distrib. 7 (5) (2013) 431–439. December 21–23.
[36] P. Luukka, Feature selection using fuzzy entropy measures with similarity [51] M.M. Saha, J. Izykowski, E. Rosolowski, B. Kasztenny, A new accurate fault
classifier, Expert Syst. Appl. 38 (4) (2011) 4600–4607. locating algorithm for series compensated lines, IEEE Trans. Power Delivery 14
[37] O. Maron, A. Moore, The racing algorithm: model selection for lazy learners, (3) (1999) 789–797.
Artif. Intell. Rev. 11 (1997) 193–225.