2282
4, OCTOBER 2013
A Fault Classification and Localization Method for
Three-Terminal Circuits Using Machine Learning
Hanif Livani, Student Member, IEEE, and Cansın Yaman Evrenosoğlu, Senior Member, IEEE
Abstract—This paper presents a traveling-wave-based method
for fault classification and localization for three-terminal power
transmission systems. In the proposed method, the discrete
wavelet transform is utilized to extract transient information
from the recorded voltages. Support-vector-machine classifiers
are then used to classify the fault type and faulty line/half in the
transmission networks. Bewley diagrams are observed for the
traveling-wave patterns and the wavelet coefficients of the aerial
mode voltage are used to locate the fault. Alternate Transients
Program software is used for transients simulations. The perfor-
mance of the method is tested for different fault inception angles,
different fault resistances, nonlinear high impedance faults, and
nontypical faults with satisfactory results.
Index Terms—Fault classification, fault location, support vector
machine (SVM), three-terminal network, traveling waves, wavelet
transformation.
I. INTRODUCTION
A. Motivation and Literature Review
A CCURATE and fast detection, classification, and loca-
tion of faults along power transmission and distribution
networks are vital to decrease operation costs, to improve
power system reliability, and to quickly restore the services
for minimum economic losses. In particular, three-terminal
circuits present challenges for fault location due to the lack of
synchronized measurements at three buses. The fault-location
methods for three-terminal circuits are either phasor or traveling
wave based. When compared to phasor-based methods, trav-
eling-wave fault locators are more accurate and more reliable
especially when metal–oxide varistor (MOV)-protected series
capacitors or mutually coupled line sections exist in the circuits.
The proliferation of high-frequency signal recorders without
using conventional current transformers (CTs) or capacitor
voltage transformers (CVTs) also provides a venue for the
implementation of traveling-wave-based methods. However,
the main challenge in traveling-wave-based fault location for
three-terminal circuits is due to the superimposed reflections of
the fault signal from the T-node and the fault point. This paper
uses support vector machines (SVMs) and discrete wavelet
Manuscript received September 26, 2012; revised January 30, 2013 and June
06, 2013; accepted July 06, 2013. Date of publication August 15, 2013; date of
current version September 19, 2013. This work was supported by the Bradley
Department of Electrical and Computer Engineering, Virginia Polytechnic In-
stitute and State University. Paper no. TPWRD-01013-2012.
The authors are with the Bradley Department of Electrical and Computer En-
gineering, Virginia Polytechnic Institute and State University, Blacksburg, VA
24061 USA (e-mail: hlivani@ieee.org; cyevrenosoglu@ieee.org).
Digital Object Identifier 10.1109/TPWRD.2013.2272936
0885-8977 © 2013 IEEE
Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on June 11,2021 at 10:44:34 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 28, NO.
transformation (DWT) to address this challenge and a new
traveling-wave-based fault classification and location method
is proposed for three-terminal systems. The following two
paragraphs provide an overview of the use of DWT and SVM
in power system fault classification and location followed by
the review of traveling-wave-based fault-location methods for
three-terminal networks.
DWT is used in [1]–[4] to extract the transient information
from the three-phase currents or voltages, and wavelet coeffi-
cients are used for faulty phase identification. In [5]–[7], neural
networks are used for fault-type classification. The steady-state
postfault three-phase voltages and currents are used as the input
to the classifiers. The fault classification performance is en-
hanced by the integration of fuzzy theory and neural networks
in [8] and [9]. The wavelet-based neural-network classifier for
power-quality (PQ) disturbances classification is implemented
and tested in [10]. The fuzzy-logic-based fault classification
method to identify the type of faults for digital distance protec-
tion system is developed in [11] and [12] using current phasors.
Support vector machine (SVM) is a statistical data classifica-
tion method, which can find the maximum marginal boundary
between different classes of a given data set and provides the
global optimal solution. This property is recognized as the main
advantage over artificial-neural-network (ANN)-based classifi-
cation methods. The use of SVM in fault-type classification ap-
pears in [13] and [14] where the steady-state postfault voltages
and currents are used as the input to the SVM fault-type clas-
sifiers. In [15], traveling wave-based fault location in transmis-
sion lines using SVM-neural network with voltage and current
transients is proposed. References [16] and [17] propose a fault
classification and location method for the power system based
on wavelet transform, SVM, and ANN. The wavelet-transfor-
mation coefficients of three-phase voltage and current transients
are used as the input to the SVM fault classifiers. The ANN is
then used for fault location according to the identified fault type.
A traveling wave-based hybrid wavelet-support vector machine
(SVM) technique for fault-zone identification in a series-com-
pensated transmission line is proposed in [18] using current and
in [19] using voltage transients. The proposed fault-zone detec-
tion methods depend on the fault type.
A recently proposed phasor-based fault-location method for
three-terminal systems is presented in [20] where the estimation
of a distance to fault and indication of a faulty section are per-
formed using three-phase current from all three terminals and,
in addition, three-phase voltage from the terminal at which a
fault locator is installed. The authors assume that synchronized
measurements are available. Travelling-wave-based fault loca-
tion for three-terminal transmission systems is first proposed in
LIVANI AND EVRENOSOĞLU: FAULT CLASSIFICATION AND LOCALIZATION METHOD FOR THREE-TERMINAL CIRCUITS 2283

[21]. The cross-correlation between the forward and backward

travelling waves and a polarity change criterion is considered
to determine the faulty line prior to fault location. The trav-
eling-wave-based fault-location technique using postfault cur-
rents at three terminals is proposed in [22]. The use of DWT
for fault location in transmission systems is first proposed in
[23] followed by a single-ended travelling wave-based proce-
dure for three-terminal systems developed in [24]. The wavelet
transformation is applied to the modal voltages at all three ter-
minals to determine the faulty line. In [25], a wavelet-based Fig. 1. Mapping input data to the feature space and decision space [30].
fault location for the three-terminal transmission system is pro-
posed. Wavelet transform is applied to the synchronized tran-
sient voltages from all three buses to detect the arrival time of
the first traveling waves. The traveling arrival times are then
used for faulty line identification and fault location. In [26], the
combined impedance- and traveling wave-based fault-location
method is proposed for multiterminal transmission systems. The
impedance-based fault location is first utilized to identify the
faulty line and the faulty half of the line. The arrival times of
traveling waves are then detected using DWT which are used
for fault location. Reference [27] proposes the use of wavelet
energies for fault location in distribution networks where the
section of the fault is first estimated based on the observed char-
acteristic frequency, which is obtained by using wavelet ener- Fig. 2. The 2-D feature space with the optimal separating hyperplane [30].
gies scalogram. The initial estimate is then improved by using
the time differences between the local maxima of continuous
wavelet-transformation coefficients corresponding to the iden- This paper is organized as follows. In Section II, the fun-
tified characteristic frequency. damentals of SVM classifiers are briefly reviewed. Section III
presents the proposed hybrid DWT-SVM fault classification
B. Contributions and Paper Organization method. In Section IV, the proposed fault-location algorithm
based on DWT and SVM is presented. Fault classification and
The preliminary results of SVM and DWT utilization for location results are provided in Section V, and conclusions are
faulty section identification in a hybrid transmission system, an given in Section VI.
overhead line combined with an underground cable, are pre-
sented in [28]. This paper presents a fault classification and
II. REVIEW OF SUPPORT VECTOR MACHINES (SVMS)
localization method based on traveling waves for a three-ter-
minal (teed) transmission system using DWT and SVM. The SVMs were originally proposed by Vapnik for binary classi-
main contributions of this paper are as follows: 1) The pro- fication problems [29]. The objective of binary SVM classifica-
posed method uses SVM for faulty line identification based on tion is to find an optimal hyperplane to separate data sets with
the classified fault type, and the method does not require syn- two classes. The SVM data classification is performed in two
chronized measurements at all terminals. This way, the algo- steps. As shown in Fig. 1, the data are first mapped from input
rithm proposed in [24] is improved by using unsynchronized space onto feature space to obtain a linearly separable data set
transient voltage measurements from only two terminals and 2) using a non-linear transform
the proposed method uses SVM for faulty-half identification.
Since the existing traveling wave-based fault-location methods (1)
use the time delay between the arrival time of the initial trav-
elling waves in ground mode and aerial mode for faulty-half
identification, the SVM-based approach makes the algorithm in- Since the mapped data in higher dimensional feature space
sensitive to the possible errors resulting from calculation of the may be linearly separable, another function is then ap-
time delay, especially for faults close to the middle of the lines; plied to map the feature space onto the decision space
3) the proposed methodology uses a smaller set of inputs to the [30]:
SVM classifiers.
The normalized energies of the wavelet transform of voltage (2)
transients are used as the input to the SVM for classification.
The performance of the proposed classification algorithm is in- In the feature space, there is a separating hyperplane that cre-
sensitive to the fault inception angle (FIA) and the method is ates the maximum distance between two classes which is called
tested for different fault resistances, nonlinear high impedance the optimal separating hyperplane. Fig. 2 shows the optimal sep-
faults (NLHIFs), and nontypical faults. arating hyperplane in a 2-D feature space.

Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on June 11,2021 at 10:44:34 UTC from IEEE Xplore. Restrictions apply.
2284 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 28, NO. 4, OCTOBER 2013

In Fig. 2, and are the weight vector and the bias term, trial-and-error procedure. For a more detailed review of SVM,
respectively. The separation margin (in meters) between the refer to [29].
classes is given as [30] In the following sections, the proposed fault-type classifica-
tion and fault-location methods based on SVM and DWT are
presented.
(3)
III. PROPOSED METHOD FOR FAULT CLASSIFICATION
In order to maximize , is minimized. Thus, the max- In this section, the traveling-wave-based fault-type classifica-
imum margin can be found by solving the following quadratic tion method in three-terminal transmission systems using DWT
optimization problem [30]: and SVM is presented.
Fault-type classification is performed using four binary sup-
(4) port vector machines . The SVMs are
subject to (5) trained to identify the faulty phase. is trained to detect
the fault at phase while is trained to detect the fault
where is the kernel function and a way of computing the inner at phase and is trained to detect the fault at phase .
product in the feature space directly as a func- is trained to detect whether the fault is grounded or not.
tion of the original input data. The solution to the problem pro- The output of each is either 1 or 1. For
vides the values of and such that the separation between example, if and are 1 while the other SVMs are
the classes is maximum. The SVMs are obtained by solving the 1, it is implied that the fault is grounded at phase .
following dual optimization problem [30]: First, the SVM classifiers need to be trained using different
fault scenarios in a given topology. Then, the performance of
(6) the SVM classifiers is evaluated using other fault scenarios. In
this paper, normalized wavelet energies of postfault three-phase
and ground-mode transient voltages are used as the input to the
subject to (7) binary SVM classifiers. The SVM classification is tested using
three different wavelets: Daubechies-4 (db-4), db-8, and Meyer.
The classification accuracy for three wavelets remains the same
where is the Lagrangian multiplier and is the number of and db-4 is utilized as the mother wavelet in this paper since
data sets. it is one of the widely adopted wavelets in the literature. The
Once the optimization problem is solved, the training points algorithm to obtain the input features of the SVMs is provided
with are the SVs, and then the and can be calcu- as follows.
lated as [30] 1) Clarke’s modal transformation for transposed lines is
applied to three-phase voltages to obtain aerial and
(8) ground-mode voltages. In the case of untransposed lines,
the modal transformation matrix obtained by Alternate
Transients Program (ATP) software can be used.
(9) 2) DWT is applied once to the recorded three-phase volt-
ages and ground-mode voltage for 40 ms
The optimal decision function can be expressed as follows to obtain the wavelet-transformation coefficients
[30]: in scale-2. The selection of scale-2 over scale-1 is due
to better performance for the investigated systems in this
study. are then squared to clearly identify the ar-
(10) rival instants and are denoted as .
3) The wavelet energies of voltages and
The most commonly used kernel functions in the literature: are calculated by summation of over one cycle
polynomial, sigmoidal, and Gaussian radial basis function after the fault is detected as
(RBF) are tested for training and evaluating the SVM classi-
fiers in this paper and the Gaussian RBF kernel is chosen based for and (12)
on its better performance. The Gaussian RBF kernel function is
where is the number of samples in one cycle.
(11) 4) The voltage wavelet energies are normalized as

where and are -dimension input vectors. , and for and

is the standard deviation of the Gaussian. The kernel func- (13)
tion parameter is tuned only once in order to achieve suffi- The training step for the classifiers is carried out to obtain the
cient accuracy for the fault-type classification or faulty line/half optimum decision functions. The input features are cal-
identification. The kernel function parameter is tuned by a culated by using transient voltages from two buses. Thus, the

Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on June 11,2021 at 10:44:34 UTC from IEEE Xplore. Restrictions apply.
LIVANI AND EVRENOSOĞLU: FAULT CLASSIFICATION AND LOCALIZATION METHOD FOR THREE-TERMINAL CIRCUITS
Fig. 3. Fault classification flowchart.
processed values are stored in an training matrix where
each column represents one feature and each row represents one
training sample. is the total number of different fault sce-
narios with different locations, types, FIAs, loadings, and fault
resistances. The flowchart of the algorithm is provided in Fig. 3.
SVM classifiers are trained using the training matrix corre-
sponding to an 8-D feature space similar to the 2-D one in Fig. 2.
Once the training process is completed and the optimal decision
function for the two-class separation are known for each phase,
the SVMs are ready to classify a new incoming data set.
IV. PROPOSED METHOD FOR FAULT LOCALIZATION
This section describes the proposed method for fault location
using SVM classifiers and DWT for a three-terminal transmis-
sion system in Fig. 4. Once the fault type is classified, the faulty
line is identified using two SVM classifiers. Then, the corre-
sponding SVM faulty-half identifier is used to identify the faulty
half. Aerial mode in scale-2 is finally observed based
on the Bewley diagram of the fault-initiated travelling waves to
identify the traveling-wave arrival instants which corresponds
to the peaks of . The arrival time differences between
consecutive waves are used in the fault-location procedure. The
fault location is developed based on the following assumptions:
• Only two terminal measurements are available.
• The measurements are not synchronized.
• The measurements are obtained through transient voltage
recorders instead of conventional CVTs.
The proposed fault-location procedure consists of two steps.
First, two different SVMs are used for faulty line identification.
In order to detect the faulty line, two binary SVMs (
1, 2) are trained. is trained to detect the fault in
line A-T and is trained to detect the fault in line B-T.
Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on June 11,2021 at 10:44:34 UTC from IEEE Xplore. Restrictions apply.
2285
Fig. 4. Three-terminal (teed) transmission system.
The algorithm identifies the fault in line C-T, if the outputs of
two SVMs are 1 at the same time.
Different SVMs are developed for different fault types.
of voltages at buses A and B are used as the inputs to both SVM
faulty line classifiers. Once the faulty line is identified, the faulty
half is determined by using another SVM which is trained by
using at bus A for a fault in line A-T, at bus B for a fault
in line B-T, and at buses A and B for a fault in line C-T.
For the faults in line A-T or B-T, the faulty half is first identi-
fied using the corresponding SVMs, and the fault-location equa-
tions developed in [23] are then used. In the case of a fault in
line C-T, if the fault is in the first half of the line ( in Fig. 4),
the distance of the fault to bus A is calculated by using:
(14)
where is the time difference between the first and the
second peaks of at bus A corresponding to the back-
ward traveling wave and the reflected backward traveling wave
from the fault point, respectively. is the traveling-
wave velocity on line C-T calculated at the frequency corre-
sponding to the middle value of scale-2 (i.e., 37.5 kHz). If the
fault is in the second half , the following equation is used
to determine the fault location with respect to bus A:
(15)
where is the time difference between the first and the
second peaks of at bus A corresponding to the back-
ward traveling wave and the reflected forward traveling wave
from bus C, respectively. The steps of the fault location in the
three-terminal circuit are summarized as follows.
1) Fault type is determined by using fault-classifier SVMs.
2) The corresponding line-identifier SVMs (for specific fault
type) are used to determine the faulty line.
3) The corresponding half-identifier SVM is used to determine
the faulty half in the line.
4) Fault location is performed using the Bewley diagram.
The simulation results of the fault classification and location
algorithms are presented in the next section.
, V. SIMULATION RESULTS
The performance of the proposed fault classification and lo-
cation method is evaluated on a 230-kV, 60-Hz three-terminal
2286 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 28, NO. 4, OCTOBER 2013

transmission system. Line lengths are assumed to be TABLE I

200 mi, 180 mi, and 170 mi. Transient sim- FAULTY-LINE IDENTIFICATION ACCURACY
FOR DIFFERENT FAULT TYPES
ulations are carried out using ATP software. Overhead lines
are modeled as frequency dependent using the data provided
in [24]. The sampling time interval of voltage measurement is
5 s (sampling frequency, 200 kHz). The traveling-wave
velocities calculated using ATP software at 37.5 kHz in aerial
and ground mode are and , respec-
tively. The MATLAB Wavelet Toolbox and SVM Toolbox [31]
are used to implement the proposed method.
Gaussian RBF is considered for training and testing the SVM
classifiers. The kernel function parameter is tuned to achieve
sufficient accuracy for each SVM. The tuning is carried out by
changing the values of from 0.1 to 10 in steps of 0.1. The
tuned parameters for different classifiers are provided accord-
ingly. The results for transposed lines are provided; however,
the proposed method is also tested on untransposed lines with
satisfactory results which are not provided due to space lim-
that the final accuracy of the fault-type classification is 90.3%,
itations. Gaussian noise is added to the measured voltages in
which is obtained by multiplying the accuracies of four SVMs.
all simulations in order to take the impacts of random error
Once the fault type is classified, the faulty line is identified.
into account. The noise is selected to have a zero mean and a
For each fault type, different trained SVMs are utilized to iden-
standard deviation equal to 1% of the sampled measure-
tify the faulty line. Two SVMs are used to identify the faulty
ments. Different fault scenarios under various system conditions
line. The accuracies of faulty line identification for different
( —loading, , , and location of the fault) are simulated
fault types are given in Table I. The tuned parameters for the
to evaluate the performance of the proposed method. The sim-
two SVM faulty-line identifiers are provided for each fault type
ulations are carried out for the following cases:
as and , corresponding to line A-T and B-T, respectively.
1) , 7 and 10 , , 10 and 13 ;
Once the faulty line is identified, then the faulty half of the
2) , 45 , 80 and 120 ;
line is determined using the SVM corresponding to the line.
3) , 1 and 100 ;
Faulty half identifier SVM 1.3) achieves 98% accuracy
4) of the lines length.
for faults in line C-T, while 99.1% accuracy is achieved for the
The simulations are carried out for the aforementioned listed
SVM 1.1) for faults in line A-T and 98.2% accuracy for
conditions and 13 824 different cases are generated. The data
the SVM 1.5) for faults in line B-T. The possible conflicts
sets corresponding to all fault conditions at load angles
of two line-identifier SVMs for A-T and B-T are handled by
and are used for training the SVM classifiers. The
comparing the value of the decision functions obtained by (10).
classifiers are then tested using the data sets corresponding to
For example, if both SVMs for A-T and B-T give 1 as the
all fault conditions at load angles , 10 and ,
output, the one which has the largest value for decision function,
13 . Subsequently, other fault scenarios with different fault con-
is superior to the other one, and the corresponding line is
ditions ( and ) corresponding to the intermediate fault
identified as the faulty one.
locations in the training set (e.g., the fault occurring at ,
Next, the results of two scenarios are provided. The first sce-
, , , ) are also tested.
nario assumes that a phase -to-ground fault occurs in line C-T
The accuracy of the fault classification algorithm is calculated
at 235 mi from bus A (i.e., 35 mi from the T-point), The fault
by using
conditions are , , , and
. The fault type is classified accurately since (cor-
of accurately classified fault types responding to phase ) and (corresponding to ground)
(16)
of test cases outputs are 1 while the other SVMs are 1. The two faulty
line identifier SVMs (for phase to ground fault) are 1 and,
According to the fault classification results, 96.3% accuracy thus, the faulty line is identified as C-T. The SVM for faulty half
is obtained for classifying the faults including phase using identification gives the output as 1, implying that the fault is
while 96.2% accuracy is achieved for clas- in the first half. Once the faulty half is identified, DWT is ap-
sifying the faults including phase using plied to the aerial mode (mode 1) voltage at scale-2 at bus A.
and 97.5% accuracy is achieved for the faults including phase Fig. 5 shows the voltage . The time difference between
using . The grounded faults are identified the first and the second traveling waves peaks in ,
with 100% accuracy using 1.5). The classification is observed as 380 s. The fault location is then calculated by
accuracy deteriorates during extreme cases when a high-resis- using (14)
tance fault (i.e., ) occurs at small fault
inception angles (i.e., ). This rare phenomenon results in
small wavelet coefficients subsequently resulting in low signal
energies making it difficult for SVMs to classify the faults. Note

Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on June 11,2021 at 10:44:34 UTC from IEEE Xplore. Restrictions apply.
LIVANI AND EVRENOSOĞLU: FAULT CLASSIFICATION AND LOCALIZATION METHOD FOR THREE-TERMINAL CIRCUITS 2287

TABLE II
FAULT-LOCATION ERROR FOR DIFFERENT CASES

Fig. 5. Voltage in aerial mode in scale-2 for a single-phase-to-ground

fault in line C-T at 235 mi from bus A.

TABLE III
FAULT-TYPE CLASSIFICATION OUTPUT FOR DIFFERENT

Fig. 6. Voltage in aerial mode in scale-2 for a single-phase-to-ground

fault in line C-T 347 mi from bus A.

The second scenario assumes that an -to-ground fault Different cases of single-phase-to-ground faults with 0.5-
occurs in line C-T 347 mi from bus A with the conditions as fault resistance are considered. The results are presented in
, , , and . The Table II.
SVM classifiers accurately determine the type of the fault, since For the faults in line C-T, the fault locations are calculated
, , and outputs are 1 while the using the aerial mode voltage at bus A and the dis-
is 1. The two faulty line identifier SVMs (for – -to-ground tances are represented with respect to T-point. The calculated
fault) are 1 and, thus, the faulty line is identified as C-T. The fault locations show good correlation with the actual fault
SVM for faulty half identification gives the output as 1, im- locations.
plying that the fault is in the second half. Voltage in
aerial mode at scale-2 at bus A is shown in Fig. 6. is observed A. Effect of Fault Inception Angle
as 250 s. The fault location is then calculated using (15) The impact of the fault inception angle on the proposed
method is evaluated in this section. Although different fault
inception angles affect the severity of the fault-initiated trav-
eling waves, the proposed method uses the normalized wavelet
As a common practice, the absolute error is calculated as a energies as the input to the SVM classifiers. The results
percentage of the total section length in order to evaluate the for four different faults with respect to the changes in are
performance of the proposed fault-location method [25] shown in Tables III and IV. As expected, the results show that
the fault classification and faulty line identification methods
Error (17) give accurate results for a wide range of fault inception angles
Total Section Length
varying between 5 and 350 . The method has satisfactory
where is the actual fault distance and is the calcu- performance for locating the faults occurring under different
lated fault distance. fault inception angles as demonstrated in Table IV.

Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on June 11,2021 at 10:44:34 UTC from IEEE Xplore. Restrictions apply.
2288 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 28, NO. 4, OCTOBER 2013

TABLE IV TABLE V
FAULTY LINE IDENTIFICATION AND FAULT LOCATION FOR DIFFERENT FAULT-TYPE CLASSIFICATION OUTPUT WITH DIFFERENT RESISTANCE

TABLE VI
FAULTY LINE IDENTIFICATION AND FAULT LOCATION FOR
DIFFERENT FAULT RESISTANCE
The sensitivities of faulty line/half identification procedures
to small fault inception angles and high-resistance faults are fur-
ther investigated. Our investigations show that the misclassifi-
cation zone around the middle point of the line CT is 3 mi for
extreme cases such as high resistance (i.e., ) faults or
faults occurring at small fault inception angles (i.e., ).
The misclassification zone around the tap point is 2 mi for such
cases.

B. Effect of Fault Resistance

Simulations are carried out for a wide range of fault resis- arc voltage, and is the arc time constant. The stationary arc
tances to evaluate the performance of the proposed method. voltage is estimated as [32]
In high impedance faults (HIFs), the phases and ground mode
peaks are smaller than low impedance faults which re-
sult in smaller wavelet energies. However, as mentioned before, (21)
the use of for fault-type classification and faulty line/sec-
tion identification makes the method robust to different fault re- where is the constant voltage parameter per arc
sistances. The performance of the algorithm for fault classifica- length, is the resistive component per arc length,
tion and faulty line identification is tested for a range of fault and is the arc length. The following parameters are used
resistance values from 0.01 to 90 . to simulate the fault [32]; 9.6, 1.6, 350, and
The fault-type classification and faulty line identification re- 1 ms.
sults for three fault resistances are shown in Tables V and VI for A single-phase-to-ground NLHIF is assumed to be located
three fault locations. As expected, the calculated fault distances 10 mi in line C-T from the T-point (i.e., 210 mi from bus A).
and the corresponding errors do not change with fault resistance. The method classifies the fault type according to the outputs of
The results show that the performance of the proposed method four SVM fault-type classifiers. The faulty line SVMs identifies
is insensitive to fault resistance. the faulty line as line C-T according to the outputs of SVM
classifiers. The SVM for the half-side identifier in line C-T gives
C. Effect of Nonlinear High Impedance Fault the output as 1. Thus, the fault section is determined in the
first half of the line. Voltage at bus A in aerial mode
A dynamic time-varying resistance model is used to evaluate
at scale-2 is finally used to calculate the fault location by using
the performance of the proposed method with a nonlinear high
(14)
impedance fault (NLHIF) [32]

(18)

(19) NLHIF has an impact on the magnitude of the traveling waves

and it delays the arrival of traveling waves due to its impact on
(20) reflection and refraction coefficients. However, the time differ-
ence between the first and the second traveling-wave arrivals
where is the arc resistance which varies with time, are not affected. The delay for the simulated fault is observed to
is the time-varying arc conductance, is the arc current, be 320 s with respect to a resistive fault. Thus, the calculated
is the stationary arc conductance, is the stationary fault location has good correlation with the actual fault location.

Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on June 11,2021 at 10:44:34 UTC from IEEE Xplore. Restrictions apply.
LIVANI AND EVRENOSOĞLU: FAULT CLASSIFICATION AND LOCALIZATION METHOD FOR THREE-TERMINAL CIRCUITS
D. Effect of Nontypical Fault
Finally, the performance of the proposed fault-location
method is evaluated for a nontypical short-circuit fault. The
0.1- resistance is connected through a 1- inductance to
the ground. A single-phase-to-ground fault is located 140 mi
in line C-T from the T-point (i.e., 340 mi from bus A). The
method first classifies the fault type as a phase- -to-ground
fault according to the output of the SVMs fault-type classifiers.
The faulty line identification is then performed according to the
classified fault. The SVM for the half-side identifier in line C-T
gives the output 1. Thus, the fault section is estimated in the
second half of the line. The voltage at bus A in aerial
mode at scale-2 is used, and is observed as 330 s. Thus,
the fault location is calculated using (15)
Even though the magnitudes of the traveling waves decrease
as in the case of high-impedance faults, the time difference be-
tween the first two arriving waves remains unchanged. It can be
concluded that the proposed method is insensitive to nontypical
faults.
E. Discussion
In this section, the performance of the proposed method
for faulty half identification is compared with an existing
traveling wave-based method. The existing wavelet-based
traveling-wave methods use the time delay between the first
traveling-wave arrival in the ground and aerial mode ,
, for faulty half identification. The fault is identified in the
first half if the is less than the precalculated time delay for
a fault located in the middle .
For comparison, a phase – -to-ground fault is considered in
line A-T located 95 mi from bus A. Fig. 7 shows the voltage
at bus A in aerial and ground mode at scale-2. The time
delay for a fault in the middle of line A-T (i.e., 100 mi) calcu-
lated using the traveling-wave velocities in aerial and ground
mode is s. In this example, is calculated
as 20 s, identifying the faulty half accurately; however, very
high precision is needed for comparison of and . It is
also a challenging task to accurately calculate since Fig. 7
demonstrates the proximity of the peak instants.
The proposed method in this paper calculates the normalized
wavelet energies and uses them as the input to the fault-
type classifier SVMs, resulting in 1 output for , ,
and while the output of is 1. Thus, the faulty
line identifier SVMs corresponding to fault -to-ground are
used. The SVMs outputs for faulty line identification give 1
for line A-T and 1 for line B-T, and the half-identifier SVM for
line A-T corresponding to the classified fault type is further uti-
lized to determine the faulty half. Finally, the faulty half is iden-
tified as the first-half of line A-T since the output of the SVM
corresponding to this line is 1, indicating the first half. The
proposed method overcomes the precision problem pertinent to
existing traveling-wave methods using the wavelet transform.
Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on June 11,2021 at 10:44:34 UTC from IEEE Xplore. Restrictions apply.
2289
Fig. 7. Voltage at bus A in aerial and ground mode in scale-2 for
phase – -ground fault 95 mi away from bus A in line A-T.
VI. CONCLUSION
In this paper, a traveling-wave-based fault classification and
localization method for three-terminal transmission systems is
proposed. DWT is applied to the three-phase and ground-mode
voltages at two terminals to obtain the normalized wavelet ener-
gies. These are later used as the input to the SVM classifiers for
fault classification and faulty line/half identification. Once the
fault is classified, faulty line and faulty half are identified, and
the wavelet coefficients of aerial-mode voltages are used to lo-
cate the fault. The performance of the proposed method is tested
for various fault scenarios, including different fault resistances,
fault inception angles, fault locations, loading levels, nonlinear
high impedance, and nontypical faults with satisfactory results.
REFERENCES
[1] A. M. Gaouda, M. M. A. Salama, M. R. Sultan, and A. Y. Chikhani,
“Power quality detection and classification using wavelet-multiresolu-
tion signal decomposition,” IEEE Trans. Power Del., vol. 14, no. 4, pp.
1469–1476, Oct. 1999.
[2] O. A. S. Youssef, “New algorithm to phase selection based on wavelet
transforms,” IEEE Trans. Power Del., vol. 17, no. 4, pp. 908–914, Oct.
2002.
[3] X. Dong, W. Kong, and T. Cui, “Fault classification and faulted-phase
selection based on the initial current traveling wave,” IEEE Trans.
Power Del., vol. 24, no. 2, pp. 552–559, Apr. 2009.
[4] Z. He, L. Fu, S. Lin, and Z. Bo, “Fault detection and classification
in EHV transmission line based on wavelet singular entropy,” IEEE
Trans. Power Del., vol. 25, no. 4, pp. 2156–2163, Oct. 2010.
[5] T. Dalstein and B. Kulicke, “Neural network approach to fault classi-
fication for high speed protective relaying,” IEEE Trans. Power Del.,
vol. 10, no. 2, pp. 1002–1011, Apr. 1995.
[6] W. M. Lin, C. D. Yang, J. H. Lin, and M. T. Tsay, “A fault classification
method by RBF neural network with OLS learning procedure,” IEEE
Trans. Power Del., vol. 16, no. 4, pp. 473–477, Oct. 2001.
[7] R. K. Aggarwal, Q. Y. Xuan, R. W. Dunn, A. T. Johns, and A. Bennett,
“A novel fault classification technique for double-circuit lines based
on a combined unsupervised/supervised neural network,” IEEE Trans.
Power Del., vol. 14, no. 4, pp. 1250–1256, Oct. 1999.
[8] H. Wang and W. W. L. Keerthipala, “Fuzzy-neuro approach to fault
classification for transmission line protection,” IEEE Trans. Power
Del., vol. 13, no. 4, pp. 1093–1104, Oct. 1998.
[9] S. Vasilic and M. Kezunovic, “Fuzzy ART neural network algorithm
for classifying the power system faults,” IEEE Trans. Power Del., vol.
20, no. 2, pt. 2, pp. 1306–1314, Apr. 2005.
2290
[10] Z. L. Gaing, “Wavelet-based neural network for power disturbance
recognition and classification,” IEEE Trans. Power Del., vol. 19, no.
4, pp. 1560–1568, Oct. 2004.
[11] B. Das and J. V. Reddy, “Fuzzy-logic-based fault classification scheme
for digital distance protection,” IEEE Trans. Power Del., vol. 20, no.
2, pt. 1, pp. 609–616, Apr. 2005.
[12] A. K. Pradhan, A. Routray, and B. Biswal, “Higher order statistics-
fuzzy integrated scheme for fault classification of a series-compen-
sated transmission line,” IEEE Trans. Power Del., vol. 19, no. 2, pp.
891–893, Apr. 2004.
[13] P. G. V. Axelberg, Y. H. G. Irene, and M. H. J. Bollen, “Support vector
machine for classification of voltage disturbances,” IEEE Trans. Power
Del., vol. 22, no. 3, pp. 1297–1303, Jul. 2007.
[14] P. K. Dash, S. R. Samantaray, and G. Panda, “Fault classification and
section identification of an advanced series compensated transmission
line using support vector machine,” IEEE Trans. Power Del., vol. 22,
no. 1, pp. 67–73, Jan. 2007.
[15] R. Salat and S. Osowski, “Accurate fault location in the power trans-
mission line using support vector machine approach,” IEEE Trans.
Power Syst., vol. 29, no. 2, pp. 979–986, May 2004.
[16] J. A. Jiang, C. L. Chuang, Y. C. Wang, C. H. Hung, J. Y. Wang, C.
H. Lee, and Y. T. Hsiao, “Hybrid framework for fault detection, clas-
sification, and location part I: Concept, structure, and methodology,”
IEEE Trans. Power Del., vol. 26, no. 3, pp. 1988–1997, Jul. 2011.
[17] J. A. Jiang, C. L. Chuang, Y. C. Wang, C. H. Hung, J. Y. Wang, C.
H. Lee, and Y. T. Hsiao, “Hybrid framework for fault detection, clas-
sification, and location part II: Implementation and test results,” IEEE
Trans. Power Del., vol. 26, no. 3, pp. 1999–2008, Jul. 2011.
[18] U. B. Parikh, B. Das, and R. P. Maheshwari, “Combined wavelet-SVM
technique for fault zone detection in a series compensated transmission
line,” IEEE Trans. Power Del., vol. 23, no. 4, pp. 1789–1794, Oct.
2008.
[19] A. A. Yusuffa, C. Fei, A. A. Jimoha, and J. L. Munda, “Fault location
in a series compensated transmission line based on wavelet packet de-
composition and support vector regression,” Elect. Power Syst. Res.,
vol. 81, no. 7, pp. 1258–1265, Apr. 2011.
[20] J. Izykowski, E. Rosolowski, M. M. Saha, and M. Fulczyk, “A Fault-
location method for application with current differential relays of three-
terminal lines,” IEEE Trans. Power Del., vol. 22, no. 4, pp. 2099–2107,
Oct. 2007.
[21] S. Rajendra and P. G. McLaren, “Travelling-wave techniques applied
to the protection of teed circuits: Multi-phase/multi-circuit system,”
IEEE Trans. Power App. Syst., vol. PAS-104, no. 12, pp. 3351–3557,
Dec. 1985.
[22] A. O. Ibe and B. J. Cory, “A travelling wave-based fault locator for two
and three terminal networks,” IEEE Trans. Power Del., vol. PWRD-1,
no. 2, pp. 283–288, Apr. 1986.
[23] F. H. Magnago and A. Abur, “Fault location using wavelets,” IEEE
Trans. Power Del., vol. 13, no. 4, pp. 1475–1480, Oct. 1998.
[24] C. Y. Evrenosoglu and A. Abur, “Travelling wave based fault location
for teed circuits,” IEEE Trans. Power Del., vol. 20, no. 2, pt. 1, pp.
1115–1121, Apr. 2005.
Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on June 11,2021 at 10:44:34 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 28, NO. 4, OCTOBER 2013
[25] M. da Silva, D. V. Coury, M. Oleskovicz, and E. C. Segatto, “Com-
bined solution for fault location in three-terminal lines based on
wavelet transforms,” IET Gen. Transm. Distrib., vol. 4, no. 1, pp.
94–103, 2010.
[26] E. E. Ngu and K. Ramar, “Combined impedance and traveling wave
based fault location method for multi-terminal transmission lines,”
Elect.Power Syst. Res., vol. 33, no. 10, pp. 1767–1775, Dec. 2011.
[27] A. Borghetti, M. Bosetti, C. A. Nucci, M. Paolone, and A. Abur, “Inte-
grated use of time-frequency wavelet decompositions for fault location
in distribution networks: Theory and experimental validation,” IEEE
Trans. Power Del., vol. 25, no. 4, pp. 3139–3146, Oct. 2010.
[28] H. Livani and C. Y. Evrenosoglu, “A hybrid fault location method
for overhead lines combined with underground cables using DWT and
SVM,” presented at the IEEE/Power Energy Soc. Gen. Meeting, San
Diego, CA, Jul. 2012.
[29] V. Vapnik, Statistical Learning Theory. New York: Wiley, 1998.
[30] N. Cristianini and J. Shawe-Taylor, An Introduction to Support Vector
Machines and Other Kernel-based Learning Methods. Cambridge,
U.K.: Cambridge Univ. Press, 2000.
[31] SVM and Kernel Methods MATLAB Toolbox. Sep. 2012. [Online].
Available: http://asi.insa-rouen.fr/enseignants/~arakoto/toolbox/index.
html
[32] M. Kizilcay and T. Pniok, “Digital simulation of fault arcs in power
systems,” Eur. Trans. Elect. Power J., vol. 1, no. 1, pp. 55–60, Jan.
1991.
Hanif Livani (S’09) received the M.S. degree in electrical engineering from the
University of Nevada, Reno, NV, USA, in 2011 and is currently pursuing the
Ph.D. degree in electrical and computer engineering at the Virginia Polytechnic
Institute and State University, Blacksburg, VA, USA.
His research interests are fault-location applications in smart grid, statistical
methods and signal processing in power systems, power system static and
dynamic state estimation, and power system operation with renewable energy
sources.
Cansın Yaman Evrenosoğlu (S’00–M’06–SM’11) received the B.S. and M.S.
degrees in electrical engineering from İstanbul Teknik Űniversitesi, Istanbul,
Turkey, in 1998 and 2001, respectively, and the Ph.D. degree in electrical and
computer engineering from Texas A&M University, College Station, TX, USA,
in 2006.
He was a Senior Consultant with Nexant Inc. Software and Information
Systems Group, Chandler, AZ, USA, for two years before he joined the Elec-
trical and Biomedical Engineering Department, University of Nevada, Reno,
NV, USA, in 2008. Since 2011, he has been with the Bradley Department of
Electrical and Computer Engineering, Virginia Polytechnic Institute and State
University, Blacksburg, VA, USA. His research interests are power system
static and dynamic state estimation, fault location, application of optimization
methods in power systems, power system data visualization, and reliable
integration of intermittent sources to the electrical power grid.