See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/251998942

Transmission line fault detection and classification

Article · March 2011

DOI: 10.1109/ICETECT.2011.5760084

CITATIONS READS
28 11,343

3 authors:

Manohar Singh Bijaya Ketan Panigrahi

Central Power Research Institute Indian Institute of Technology Delhi
30 PUBLICATIONS   276 CITATIONS    528 PUBLICATIONS   6,910 CITATIONS   

SEE PROFILE SEE PROFILE

R.P. Maheshwari
Indian Institute of Technology Roorkee
105 PUBLICATIONS   2,021 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Voice and Gait Recognition View project

Computer Vision View project

All content following this page was uploaded by R.P. Maheshwari on 24 December 2013.

The user has requested enhancement of the downloaded file.

 PROCEEDINGS OF ICETECT 2011

Transmission Line Fault Detection and Classification

Manohar singh, Dr. B.K Panigrahi Dr. R. P. Maheshwari, Member IEEE
Department of Electrical Engineering, Indian Institute Department of Electrical Engineering, Indian Institute
of Technology, Delhi of Technology, Roorkee
Delhi, India-110016 Roorkee, India-247667
manoharsingh33@ gmail.com [email protected]

Abstract—Transmission line protection is an important issue in resistance, mutual-coupling from adjacent parallel lines, etc.
power system engineering because 85-87% of power system faults They do not have the ability to adapt dynamically to the system
are occurring in transmission lines. This paper presents a operating conditions, and to make correct decisions if the
technique to detect and classify the different shunt faults on a signals are uncertain.
transmission lines for quick and reliable operation of protection
schemes. Discrimination among different types of faults on the Fault detection and classification is a very challenging task.
transmission lines is achieved by application of evolutionary Different attempts have been made for fault classification
programming tools. including approaches based on traveling waves [1-2], adaptive
PSCAD/EMTDC software is used to simulate different Kalman filtering [3], fuzzy logic, neural networks[4], and the
operating and fault conditions on high voltage transmission line, fusion of different artificial intelligence techniques. Several
namely single phase to ground fault, line to line fault, double line researchers have proposed different techniques for fault
to ground and three phase short circuit. The discrete wavelet classification of transmission lines using different types of
transform (DWT) is applied for decomposition of fault transients, neural networks and their combination with different
because of its ability to extract information from the transient transforms, such as wavelet and hyperbolic-s [5]. Although the
signal, simultaneously both in time and frequency domain. The neural-network based approaches have been quite successful in
data sets which are obtained from the DWT are used for training determining the correct fault type, the main disadvantage of
and testing the SVM architecture. After extracting useful neural-network is that it requires a considerable amount of
features from the measured signals, a decision of fault or no fault
training effort for good performance, especially under a wide
on any phase or multiple phases of a transmission line is carried
out using three SVM classifiers. The ground detection task is
variation of operating conditions (such as system loading level,
carried out by a proposed ground index. Gaussian radial basis fault resistance, source impedance, etc.). Moreover, another
kernel function (RBF) has been used, and performances of disadvantage of neural-network-based algorithms is that the
classifiers have been evaluated based on fault classification training may not converge in some cases, as the starting point is
accuracy. In order to determine the optimal parametric settings chosen at random and can end up in a local minimum [6-7].
of an SVM classifier (such as the type of kernel function, its This paper presents a new approach for fault detection and
associated parameter, and the regularization parameter c), five-
classification on transmission lines using SVMs for training
fold cross-validation has been applied to the training set. It is
with an online wavelet-based preprocessing-stage. The
observed that an SVM with an RBF kernel provides better fault
classification accuracy than that of an SVM with polynomial
algorithm is based mainly on calculating the RMS value of
kernel. It has been found that the proposed scheme is very fast transient energy of pre-fault and post-fault signals of the three
and accurate and it proved to be a robust classifier for digital line currents and three line voltages. These preprocessed
distance protection. signals are trained with SVM. Fault classification accuracy is
Keywords—Discrete wavelet transforms, Support vector estimated for different loading and for source impedance
machine (SVMs), Feature extraction, Fault classification accuracy
and Ground index.

I. INTRODUCTION II. POWER SYSTEM NETWORK SIMULATION

Fault detection and classification on transmission lines are
A. System Studied
important task to safeguard electric power systems. A
fundamental part of a protective relay is a selector module The single-line diagram of a power system for a two-terminal
which classifies the type of fault that has occurred and also to is shown in Fig. 1. The generators are represented by
classify the “normal state”. Reliable phase selection of the equivalent potential source and equivalent source impedance.
faulted phase is thus vitally important in order to avoid either The transmission line has been represented using the Bergeron
tripping of the incorrect phase or unnecessary three-phase line model in PSCAD/EMTDC. The line parameters and
tripping. Moreover, a necessary requirement of phase selectors source impendence values are given in the Table I and Table
is high speed operation as the selection process must be II. To ensure proper fault classification, data used for training
completed in the immediate post-fault period before breaker and testing must be adequate in number. Thus, huge data
opens. Traditional phase selection schemes suffer from some generation for different faults and system conditions is
drawbacks due to complexity of the system model, lack of important for designing a fault classification scheme. The bulk
knowledge of its parameters, effect of remote-end infeed, fault data are generated using multiple run component of PSCAD

978-1-4244-7926-9/11/$26.00 ©2011 IEEE 15

software. With the help of this component we can use fault
resistance, fault inception angle, and fault duration as variable
parameter in PSCAD. The generated data is stored in output
file option in the project setting option of PSCAD [8].
Different fault on the transmission line are created by
using fault block in the PSCAD, fault timing in controlled
using the time fault logic block. The fault position on the
transmission line is adjusted manually by dividing the line into
two parts of required length then fault is applied at point of
division. In this way fault are simulated along the line in step
of 20% of 120 km long line. Variation of source impendence
is simulated by changing the impendence setting option of the
source block. The flow of power on the line is controlled by
changing the load angle between the two terminals of the line;
normally load angle is kept in range of 5 – 20 degree [9-10].
B. Typical Pyimary wave form
Data is collected for one cycle of fault, simulation time is
adjusted in such a way that 100 samples of fault data are taken
for one cycle of fault. Fig. 2 (from top to bottom) ) shows wave
form of current and voltage signals for a-g, a-b, ab-g, abc-g
faults and for no fault.

Fig.1. Transmission Line Model

Table I
Transmission line Parameters
Positive sequence resistance R1 Ω/KM 0.02056

Zero sequence resistance R1 Ω/KM 0.16270

Positive sequence inductance X1 Ω/KM 0.000096

zero sequence inductance X0 Ω/KM 0.003885

Positive sequence Capacitance C0, F/KM 2.50290e-05

Zero sequence Capacitance C0, F/KM 3.19500e-05

Table II
Source Parameters
Parameters Source 1 Source 2
Positive sequence 5∟85 10∟85
impendence

Zero sequence 10∟85 20∟85

impendence Fig.2. Simulation of shunt types of faults

III. WAVELET TRANSFORM

Wavelets are functions that satisfy the requirements of
both time and frequency localization. The necessary and

16
sufficient condition for wavelets is that it must be oscillatory, data. When formulated under SVMs the optimal hyper-plane
must decay quick1y to zero and must have an average value of is the decision boundary that attains the maximum margin of
zero. In addition, for the discrete Wavelet transform separation between the two classes. Optimal hyper-plane
considered here, the Wavelets are orthogonal to each other under these conditions can be uniquely determined by solving
[11]. Wavelets can provide multiple resolutions in both time a constrained optimization problem whose solution has an
and frequency. The signals can be accurately reproduced with expansion
the Wavelet analysis using relatively small number of w = ∑iai xi
components. The analyzing wavelets are called the “mother
wavelets” and it’s dilated and translated versions are called the which is the weighted sum of the training patterns that lie on
“daughter wavelets”. It has a digitally implementable the margin [24]. The training vectors which lie on the margin
counterpart called the discrete Wavelet transform (DWT). The are referred to as the support vectors. Formally, given a set of
generated waveforms are analyzed with Wavelet multi-
features vectors xi and corresponding class labels yi we form
resolution analysis (MRA) to extract sub-band information
from the simulated transients. Daubechies Four (db-4) wavelet the index pair {( xi , yi )} where xi ∈ R and yi ∈ {+1,−1} .
is used in this work for the analysis as it closely matches the
signal to be processed which is of utmost importance in We denote those input features xi with label +1 as belonging
wavelet applications. Wavelet co-efficient of the signal are to Class 1 and those feature vectors with label -1 as belonging
obtained by the decomposition of a discrete fault current and to Class 2. In the linearly separable case we can find a hyper-
voltage signals using Mallat’s algorithm [12]. plane which is able to separate the training data. This
Due to the unique feature of providing multiple separating hyper-plane will take on the following form,
resolutions in both time and frequency by wavelets, the sub- f ( x) = wx + b
band information can be extracted from the original signal.
When applied to faults, this sub-band information of faulted Where wi ∈ R and is the normal vector to the separating
power system is seen to provide useful signatures of faults, so hyper-plane and b ∈ R is the scalar bias. Hence, if x belongs
that fault classification can be done elegantly. By randomly to Class 1(+1 label) then f ( x ) ≥ 0 and conversely if x
shifting the point of fault on the transmission line, a number of
simulations are carried out employing PSCAD/ EMTDC. The belongs to Class 2 (-1 label) we have f ( x ) ≤ 0 .
generated time domain signal for each case is analyzed using Consequently, we have the following constraints,
Wavelet transform. From the different decomposed levels [13- yi ( wi xi + b) ≥ 0 for i= 1…n
17]. Only first level output is considered for the analysis. As
for a weight vector w and bias b we define the margin
the absolute values of the summation of first level output for
separation ρ as the separation between the hyper-plane and
all the inception angles considered in the analysis are found
higher as compared to that of other level outputs. It indicates the closest feature vector. In particular an SVM finds the
that the total area under the characteristics of first level outputs separating hyper-plane which maximizes the margin of
is more than that of other level outputs. Another reason for the separation ρ . The idea behind maximizing the margin of
first level output to be selected as the parameter for fault separation is rooted in concepts of structural risk
Classification is due to discrimination in transient energy minimization.
content of faulty and healthy signals is higher as compare to Roughly, when the margin of separation is maximized the
other levels. Moreover, as the sampling time considered in the generalization error on unseen test data tends to its minimum
analysis is 0.0002 s which corresponds to a sampling value [21-25].
frequency of 5 kHz and the total number of wavelet levels The SVM classifier is also defined as
considered is 6, therefore, 6th level wavelet output corresponds f ( x ) = wΦ ( x ) + b (1)
to a frequency band of 70 – 160 Hz. Level 1 corresponds to
frequency band of 1.25 – 5 kHz, which are rich in higher order where Φ ( x ) is a mapping function to map the input pattern
harmonics, predominant in case of faults. The RMS value of x into higher dimensional space . This classifier f ( x) is
transient energy of current and voltage signals is utilized as a linear in terms of the transformed data, but non-linear in terms
feature for detection and discrimination among the faults. of the original data
However, the fault is to be discriminated from other types of Following non-linear transformation, for the parameters of
transient disturbances before using the classification the decision function f ( x ) the minimization of following cost
algorithm.
function is
IV. SUPPORT VECTOR MACHINE 1 2
j ( w, b) = w + c ∑i ξ i
Support vector machines (SVMs) were first introduced 2
by Vapnik [19-20] in the late 1960’s as a part of his Subjected to constraints
development of statistical learning theory. Support vector
machines are a type of hyper-plane classifier which attempts
yi ( w.φ ( xi ) + b ≥ 1 − ξ For i = 1…n (2)
to find an optimal separating hyper-plane for the input training The solution to minimize the above cost function subject to the
constraints in Eq. (2) is given from following dual:

17
Maximize and voltage are exported to support vector machine model for
1 training. The output of each SVM indicates weather a particular
Ld = ∑ α i − ∑ α iα j yi y jφ ( xi )φ ( x j ) phase is involved in fault or not [34].
i 2 i, j
Ground fault detection is carried through detection of
Zero-Sequence component. Phases which do not involve the
1
Ld = ∑ α i − ∑ α i α j y i y j k ( xi , x j ) (3)
ground has almost zero, zero-sequence component, where as
2 i, j for phases involving ground have observable amount of zero-
i
sequence component.
Where k ( xi , x j ) known as kernel, is a non-linear
function and is defined as
k ( xi , x j ) = φ ( xi , x j ) (4)
Subjected to constraint
0 ≤ αi ≤ c
∑α y
i
i i =0 (5)

Solution is given as

Ns
w = ∑ α i yiφ ( xi )
i

where N s is the number of support vectors.

The kernel function in an SVM plays the central role of
implicitly mapping the input vector (through an inner product)
into a high-dimensional feature space. In the present study,
radial basis functions (RBFs)) have been used. They are
defined as follows Fig 3. Fault detection and classification algorithm

k ( x, y ) = e
(−γ x− y ) 2

Where γ =
1
(6) A. Data Generation
2σ 2 For training and testing of SVM, the fault simulation has
been carried out on transmission system as shown in Fig.1.
V. PROPOSED FAULT CLASSIFICATION SCHENE PSCAD software is used to simulate the faults on the
transmission line. Table III shows the different conditions that
A 400kv, 50 Hz power system shown in Fig.1, consisting of
are simulated on transmission line for data generation used for
two sources representing two areas connected by 128 km long
training the SVMs and Table IV shows condition
transmission line is used for simulation studies. Faults take
corresponding to the data simulated for testing the trained
place at different position of transmission line. All the data is
models.
collected from sending end of the system. Simulations of faults
which occur at 0.6 sec are shown in Fig. 2. . Fault signals Table III
(voltages and currents) are collected for a cycle at 50 Hz. These Line and system parameters used for generation of training patterns
signals are taken in discrete form, 100 samples of current and Fault type a-g, b-g, c-g, a-b, b-c, c-a, a-b-g, b-c-g, c-a-g, a-
b-c and a-b-c-g
100 samples of voltage are taken for all three phases for all 11 Fault location (KM) 25.6, 51.2, 76.8 and 102.4 (128 KM line length)
types of faults [25]. Fault resistance (Ω) 0 , 25 , 50 , 75 , 100 , 125 and 150
The proposed scheme makes use of SVM classifier for Fault inception angle 0, 45, 90, 135, 180, 225, 270, 315 and 360
(0’)
detection and classification of fault. Fig. 3 shows the flow chart Load angle variation 10% and 20% ( source 1)
of proposed algorithm. The fault classifier has three classifiers (0’)
for each phase. The extracted features are fed to these Table IV
classifiers and their output indicates the type of fault. The Line and system parameters used for generation of testing patterns
voltages and currents signal are sampled at a frequency of 5 Fault type a-g, b-g, c-g, a-b, b-c, c-a, a-b-g, b-c-g, c-a-g, a-b-c
kHz. The RMS value of transient energy of voltages and and a- b-c-g
Fault location (KM) 19.2 and 57.6
currents are extracted using db4 as mother wavelet. The Fault resistance ( ) 5 Ω, 30 Ω, 55 Ω, 80 Ω, 105 Ω, 135 Ω and155Ω
decomposition is performed at level 1, which contains rich. Fault inception angle
harmonics of order 25-50 Hz. From the wavelet decomposition (0’) 0, 60, 120, 180, 240, 300 AND 360
of simulated signals (voltage and current), it is clearly
Load angle variation 5% and 15% ( source 1)
detectable that RMS values of transient energy of current (0’)
signals for faulty phases is much more as compare to healthy
phases. Then these RMS value of transient energy of current

18
B. Selection of mother wavelet for feature Extractions
The Daubechies family is one of the most suitable wavelet
families in analyzing power-system transients although there
are no definite criteria for the selection of wavelets; the best
choice is a wavelet that most strikingly exhibits the
phenomena to be studied. It has been found that db4, with its
characteristic as the shortest wavelet in its family, is an
excellent choice to detect abrupt change in RMS value of
transient energy of the pre-fault and post-fault signals. Even
though other families of wavelet-transform also give change in
transient energy of the pre-fault and post-fault signals, but this
change is more in case of db4 mother wavelet. Hence db4 is Fig.5. Wavelet coefficients of phase “a” for a-g
used as mother wavelet. The Fig.4.shows the variation of
wavelet coefficients decomposed at level 1 using db4 as
mother wavelet. On the top of Fig. 4 comparison of phase ‘a”
fault current for single line to ground fault vs. normal load
current is shown, middle corresponds to phase ‘b’ current and
bottom to phase ‘c’ current for the same type of fault.

Fig.6.Transient energy of current and voltage for phase “a” for a-g fault

D. Scaling for feature extraction

The input patterns (training and test patterns) are
normalized to [+1,-1] before inputting to the SVM module
[40]. The main advantage is to avoid attributes in greater
numeric ranges dominate those in smaller numeric ranges.
Another advantage is to avoid numerical difficulties during the
calculation. Because kernel values usually depend on the inner
Fig.4. Discrete Samples of fault currents products of feature vectors, large attribute values might cause
numerical problems
C. Feature extraction
The proposed fault classification technique is based on E. Performance Evaluation
RMS value of transient energy of the detail coefficient of the 1. Parameter Selection and Training
respective phases of voltage and current signals. Considering a
sampling frequency of 100 samples/cycle of a 50 Hz Once the training samples are obtained, the next step is to determine
fundamental frequency signal, the measured signals (three the optimal parametric settings of the SVM. In this process, the type
voltages and three currents) are decomposed into of kernel function, its associated parameter, and the regularization
approximation (A1) and detail coefficients (D1) using a db4 parameter C must be decided. To optimize these parameters, five-
mother wavelet. Decomposition is performed up-to 10 levels; fold cross-validation [44] has been applied to the training set. This
the details coefficients extracted by wavelet transform at level procedure consists of the following steps. First, randomly divide all of
1 decomposition have more value of transient energy under the available samples in the training set into five equal-sized subsets.
fault duration as compare to pre-fault signals. Hence fault Second, for each model-parameter setting, train the SVM classifier
classification can be done more effectively using transient five times; during each time, one of the subsets is held out in turn
energy of these details coefficients at level-1 of voltage and while the rest of the subsets are used to train the SVM. This trained-
current signals. Fig. 5, Shows the complete decomposition of SVM classifier is then tested using the held-out subset, and its
fault signal up-to level 6 for phase “a” current and voltage accuracy values are recorded. Finally, the model with the highest
signal when a-g fault occur at 50% of line length. Fig. 6 gives accuracy is adopted. Various parameters for the SVM (such as
value of transient energy of samples corresponding to different regularization parameter C, sigma of RBF, etc.) are varied as C from
simulated signals. The shaded potion shows that transient 1 to 10000, and sigma of RBF from 2000 to 0.02, to choose the best
energy of fault phase ‘a’ current on left side and voltage on parameters for SVM. For C = 10 and σ = 20 gives the maximum
right side when phase “a-g’ fault takes place[35-36]. accuracy.

19
 Table V
Selection of parameters for RBF kernel function
C Accuracy variation with various value of sigma (ߪ)

1 2000 200 20 2 0.2 0.02

10 96.7532 96.7532 98.7013 87.9870 88.0345 90.2597
Fig.7 Zero sequence current component for a-g and a-b faults
100 96.7532 96.7532 98.7013 97.0779 87.9897 89.6104
1000 96.7532 96.7532 98.7013 96.4286 96.1039 87.9870 3. Fault Classification Performance
The variation of fault classification accuracy for different
10000 96.7532 96.7532 98.7013 96.4286 96.1039 96.1039
types of fault at different fault location is tabulated in Table
VI. Variation of source (S2) impendence at receiving end as
2. Ground Detection in Ground faults shown in Table III, is also taken into account. For model
Usually, it is not possible to identify the involvement of generalization the line parameter of another line is also
ground using voltage and current signals. Therefore, the taken into consideration. Training and Testing samples are
ground fault involvement detection task is carried out using collected for all the possible conditions that occur on the
zero-sequence current ( I 0 ). For detecting the involvement transmission line. Testing samples are different from the
training samples. These training samples are trained through
of ground fault, a ground detection index (GI ) is proposed.
SVM structure, and are tested against the different testing
Where I pha , I phb , I phc are phase currents? samples.
The classification accuracy is between 90-100%. When
1
GI = I 0 = ( I pha + I phb + I ph c ) source variation and parameter variation are taken into
3 account the classification accuracy falls marginally. But this
can be improved by training the SVM for mixed training
When the value of GI index is more than the threshold (α )
data of variable source impendence and line parameter on a
of 0.025 for fault resistance of 5 Ώ, it indicates the another line. Results are shown in Table VII for different
involvement of fault with ground. The variation of zero- training and testing samples. Following are notation used in
sequence current for different types of faults is calculated. It Table VII.
is seen from Fig.7 that for a fault involving ground (L-G); K Training data at 50% and 20% of line length
the value of GI is much above the threshold value. Whereas, L Training data at 50%, 20% and 40%of line length
for a fault not involving ground (L-L), the value of GI is M Training data at 50%, 20%, 40% and 60%of line
much below the threshold value [35]. length
N Training data at 50%, 20%, 40%, 60% and 80%of
line length
O Training data at 50%, 20%, 40%, 60% and 80% of
line length and source variation at receiving end as shown in
Table III
Table VI
Detailed fault classification result for different training samples and testing samples for different fault location
Line length Training Testing TP PhA TN TP TN TP TN Accuracy Ph Accuracy Ph Accuracy Ph C
in ( %) samples samples PhA Ph B Ph B Ph C PhC A (%) B (%) (%)

50 1848 308 304 04 304 04 307 01 98.7013 98.7013 99.6753

20 1848 1155 1137 18 937 218 1155 00 98.4416 81.1255 100.0000

20 6853 1155 1108 47 1155 00 1155 00 95.9307 100.0000 100.0000

40 1848 1155 1155 00 1119 36 1155 00 100.0000 96.8831 100.0000

40 6853 1155 1146 09 1155 00 1155 00 99.2208 100.0000 100.0000

60 1848 1155 1130 25 924 231 1144 11 97.8355 80.0000 99.0476

60 6853 1155 1115 40 1155 1155 1150 05 96.5368 100.0000 99.5671

80 1848 1155 1106 49 891 264 1115 40 95.7576 77.1429 96.5368

80 6853 1155 1107 48 1155 00 1150 05 95.8442 100.0000 99.5671

20
 Table VII
Fault classification result for different training samples and testing samples for different fault location, including source variation
Line Training Testing TP Ph TN Ph TP Ph TN Ph TP Ph TN Ph Accuracy Ph Accuracy Ph Accuracy Ph
length samples samples A A B B C C A (%) B (%) C (%)

K 2998 201 153 48 186 15 172 29 76.1194 92.5373 85.5721

L 4158 509 496 14 505 04 508 01 97.4460 99.2141 99.8035

M 5313 710 641 69 694 16 703 07 90.2817 97.7465 99.0141

N 6468 911 791 120 893 18 905 06 86.8276 98.0241 99.3414

O 6853 1112 1077 35 1094 18 1106 06 96.8525 98.8309 99.4040

P 8008 1313 1078 253 1202 111 1118 195 82.1021 91.5461 85.1485

[6] H. Wang and W. W. L. Keerthipala, “Fuzzy-neuro approach to

VI. CONCLUSION fault classification for transmission line protection,” IEEE
Transactions on Power Delivery, Vol. 13, No. 4, pp. 1093–1104,
In this paper, an accurate technique of identification October 1998.
and classification of faults on transmission line has been [7] P. K. Dash, S. R. Samantaray and G. Panda, “Fault classification
proposed. The method utilizes the samples of current and and section identification of an advanced series-compensated
transmission line using support vector machine,” IEEE
voltage extracted from the fault point. Wavelet transforms is Transactions on Power Delivery, Vol. 22, No. 1, pp. 67–73,
used to extract the transient energy feature from these January 2007.
sampled signals. This feature vector then acts as input to the [8] PSCAD/EMTDC Version 3.08, Winnipag, MB, Canada:
SVM for training. The proposed scheme was found to be Manitoba HVDC Research Centre.
[9] PSCAD/EMTDC Power System Simulation Manual, Winnipeg,
immune of fault resistance, fault inception angle, change in Mb, Canada S. Haykin, Neural Networks- A Comprehensive
load, lightning disturbance, auto-reclosing action, and even Foundation, Second Edition Englewood Cliffs, NJ: Prentice-
CT saturation as the transient energy content in these action Hall ,1999
is very low due to their very short existence time The [10] Michel Misiti, Yves Misiti, Georges Oppenheim and Jean-
Michel Poggi, “Wavelet Tool box 4 matlab user’s Guide,”
proposed scheme is also free from many difficulties that are March 2008.
faced by traditional neural networks approaches, such as [11] G. Strang and T. Nguyen, “Wavelet and Filter Banks”
generalization Cambridge, MA: Wellesley-Cambridge, 1996
The proposed scheme can be extended for locating the [12] Daubechies, “Ten Lectures on wavelet,” Philadelphia, PA: Soc.
Industrial Application Math., 1992.
exact location of fault and classification of fault on a series [13] C. H. Kim and R. Aggarwal, “Wavelet transform in power
compensated lines, but a few more features may be required system,” IET Power Engineering Journal, Vol.15, pp. 193-200,
for the training of SVM. August 2001.
[14] C. S. Burrus, R. A. Gopinath, and H. Guo, “Introduction to
Wavelets and Wavelet Transform: A primer. Englewood Cliff,”
REFERANCE NJ: Prentice-Hall.1998.
[1] D. W. P. Thomas and C.Christopulo, “Ultra-high speed [15] T. M. Lai, L. A. Snider and E. Lo, “wavelet transform based
protection of series compensated lines,” IEEE Transactions on relay algorithm for the detection of stochastic high impedance
Power Delivery, Vol. 7, pp. 139–145, January 1992. faults,” ELSEVIER Transaction on Electric Power System
[2] Jamil, M. Thomas, M. S. Moinuddin and P. Kumar, “Fuzzy Research, pp. 626-633.
approach to fault classification for transmission line protection,” [16] T. M. Lai, L. A. Snider, E. Lo, and D. Suntanto, “High-
Proc. IEEE Tencon 99 Conf., Vol. 2, pp. 1046–1050, September impedance fault detection using discrete wavelet transform and
1999. frequency range and rms conversion,” IEEE Transactions on
[3] A. A. Girgis, A. A. Sallam and A. Karim, El-Din, “An adoptive Power Delivery, Vol.20, No.1, pp. 397-407, January2005.
protection scheme for advanced series compensated (ASC) [17] Chunju Fan, K. K. Li, W. L. Chan and Weiyong Yu, “Study of
transmission lines,” IEEE Transaction on Power Delivery, Vol. protection scheme for transmission line based on wavelet
13, No. 2, pp.414–420, April 1998. transient energy,” Electrical Power & Energy System, Vol.28,
[4] B. Das and J. V. Reddy, “Fuzzy-logic-based fault classification pp. 459-470, 2006.
scheme for digital distance protection,” IEEE Transactions on [18] Fernando H. Magnago and Ali Abur, “Fault location using
Power Delivery, Vol. 20, pp. 609–616, April 2005. wavelet,” IEEE Transactions on Power Delivery, Vol. 13, No. 4,
[5] R. N. Mahanty and P. B. Dutta Gupta, “A fuzzy logic based October 1998.
fault classification approach using current samples only,” Int. J. [19] Vladimir N. Vapnik, “An overview of statistical learning
Elect. Power Syst. Res., Vol. 77, pp. 501–507, 2007. theory,” IEEE Transactions on Neural Networks, Vol.10, No.5,
1999

21
[20] C. Cortesand and V. Vapnik, “Support vector network,” [39] Thomas Dalstein and Bernd Kulicke, “Neural Network
Machine Learning, Vol. 20, pp. 273-297, 1995. Approach to Fault Classification for High Speed Protective
[21] C. J. C. Burges, “A tutorial on support vector machine for Relaying,” IEEE Transactions on Power Delivery, Vol.10, No.
pattern recognition,” Data Mining and knowledge Discovery, 2, pp. 1002-1011, April 1995.
Vol.2, No.2, pp. 121-167, 1998. [40] C. H. Kim and R. Aggarwal, “Wavelet transform in power
[22] Haykin Simon “Neural networks: A comprehensive system,” IET Power Engineering Journal, Vol.15, pp. 193-200,
foundation”, 2nd Edition, Prentice-Hall, Inc., 1999. August 2001.
[23] Chih-Wei Hsu and Chih-Jen Lin, A, “Comparison of methods [41] B. Ravikumar, H. P. Khincha, D. Thukaram, “An intelligent
for multiclass support vector machines,” IEEE Transactions on approach using support vector machines for monitoring and
Neural Networks, Vol.13, No.2, March 2002. identification of faults on transmission systems,” IEEE Power
[24] Chih-Wei Hsu and Chih-Jen Lin, “LIBSVM: A library for India Conference, pp. 184-190, 05, June, 2006.
support vector machines,” 2001 [42] P. K. Dash, A. K. Pardhan, and G. Panda, “Application of
[25] Joe-Air Jiang, ching-Shah chen, and Chih-Wen Liu, “A new minimal radial basis function neural networks to distance
protection scheme for fault detection, direction discrimination, protection,” IEEE Transactions on Power Delivery, Vol.16, No.
classification, and location in transmission lines,” IEEE 1, pp. 68-74, January 2001.
Transactions on Power Delivery, Vol.18, No.1, pp. 1253-1261, [43] R. N. Mahanty, P. B. and Dutta, “Application of radial basis
January 2003. function neural networks to fault classification and location in
[26] U. Parikh, R. P. Maheshwari, B. Das and B. Bhalja, “A new transmission lines,” IEE Proc. Gener. Transm. Distrib., Vol.151,
fault classification technique for protection of series No. 2, March 2004.
compensated transmission lines,” International Conference on [44] Jeo-Air Jiang, Ching-shan Chen, and chin-Wen Liu, “A new
Power System Protection, CPRI, Bangalore, India, 20–21 protection technique for fault detection, direction
February 2007. discrimination, classification, and location in transmission
[27] R. K. Aggarwal, Q. Y. Xuan, A. T. Johns, R. W. Dunn and lines,” IEEE Transactions on Power Delivery, Vol.18, No. 1,
A.Bennett, “A novel fault classification technique for double- pp. 34-42, January 2003.
circuit lines based on a combined unsupervised/supervised [45] A. S. Omar Youssef, “New algorithm to phase selection based
neural network,” IEEE Transactions on Power Delivery, Vol. on wavelet transform,” IEEE Transactions on Power Delivery,
14, No. 4, pp. 1250–1256, October 1999. Vol.17, No.4, pp. 908-914, October 2002.
[28] Bhavesh, Bhalja and R. P. Maheshwari, “Trends in adaptive
distance protection of multi – terminal and double-circuit lines,”
Int. J. Elect. Power Components Syst., Vol. 34, pp. 603–617,
June 2006.
[29] Bhavesh, Bhalja and R. P. Maheshwari, “Protection Of Double
Circuit Line Using Wavelet Transform,” IE Journal-EL, Vol.
87, pp. 67-70 September 2006.
[30] Bhavesh, Bhalja and R. P. Maheshwari, “Philosophy of
protection for multi-terminal and double-circuit lines,” Int.
Conf. on Computer Applications in Electrical Engineering
Recent Advances, Roorkee, India, Sept. 28-October 1, 2005, pp.
582-588.
[31] Yi, Hu, Damir, Novosel, Murari Mohan Saha, and Volker
Leitloff, “An Adaptive Scheme For Parallel Line Distance
Protection,” IEEE Transactions on Power Delivery, Vol. 17, No.
1, pp. 105-110 January 2002.
[32] Xiangning Lin, Hanil Weng, and Bin Wang, “Identification of
cross country fault of power transformer for fast unblocking of
differential protection,” IEEE Transactions on Delivery, Vol.
24, No. 3, pp. 1079-1086 July 2009.
[33] Bhavesh, Bhalja and R. P. Maheshwari, “New differential
protection scheme for tapped transmission line,” IET Gener.
Transm. Distr., Vol. 2, pp. 271–279, 2008
[34] Bhavesh, Bhalja and R. P. Maheshwari, “Wavelet-based fault
classification scheme for a transmission line using a support
vector machine,” Electric Power Components and Systems,
Vol.36, Pp. 1017-1030, 2008.
[35] V. S. Kale, S. R. Bhide, R. P. Bedekar and G. V. K. Mohan,
“Detection and classification of faults on parallel transmission
lines using wavelet transform and neural network,” International
Journal of Electric and Electronics Engineering Vol. 1, pp.364-
368, 2008.
[36] P. S. Bhowmik, P. Purkait, K. Bhattacharya, “A novel wavelet
transform aided neural networks based transmission line fault
analysis method,” Electrical Power & Energy System, Vol.31,
No. 5, pp. 213-219, 2009.
[37] Robert Salat and Stanislaw Osowski, “Accurate fault location in
the power transmission line using support vector machine
approach,” IEEE Transaction on Power Sysyem, Vol.19, No.2,
May 2004.
[38] S.M. De Souza, A.P. Alves da Suva, and A.C.S. Lima,
“Attribute selection for fault location recognition in
transmission lines,” IEEE Power Engineering Society General
Meeting, pp. 6, October, 2006.

22

View publication stats