IEEJ TRANSACTIONS ON ELECTRICAL AND ELECTRONIC ENGINEERING

IEEJ Trans 2016; 11: 43–48

Published online in Wiley Online Library (wileyonlinelibrary.com). DOI:10.1002/tee.22187

Paper

Discrete Wavelet Transform and Support Vector Machine-Based Parallel

Transmission Line Faults Classification
Ahmed Sabera , Non-member
Ahmed Emam, Non-member
Rabah Amer, Non-member

This paper presents a scheme for classification of faults on double circuit parallel transmission lines using combination of
discrete wavelet transform and support vector machine (SVM). Only one cycle post fault of the phase currents was employed to
predict the fault type. Two features for each phase current were extracted using discrete wavelet transform. Thus, a total of 12
features were extracted for the six phase currents. The training data were collected, and SVM was employed to establish the fault
classification unit. After that, the fault classification unit was tested for different fault states. The power system simulation was
conducted using the MATLAB/Simulink program. The proposed technique took into account the mutual coupling between the
parallel transmission lines and the randomness of the faults on transmission line considering time of occurrence, fault location,
fault type, fault resistance, and loading conditions. The results show that the proposed technique can classify all the faults on the
parallel transmission lines correctly. © 2015 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.

Keywords: transmission line, support vector machine, wavelet transform

Received 6 February 2015; Revised 13 May 2015

1. Introduction of features is not restricted. Thus, the original data can be directly
used for training SVM without preprocessing them to extract the
Double circuit transmission lines are widely used for electrical required features. So, SVM is a good choice for fault classification
energy transmission to improve power transfer capability and problems.
power system reliability. However, the configuration of parallel Recently, wavelet transform was used for capturing the dynamic
transmission lines has difficulties in fault type classification characteristics of nonstationary signals using a short data window,
due to the effect of mutual coupling between the two circuits. and its potential benefits for analyzing transient fault signals in
In addition, the accuracy of the protection system is affected power systems have been recognized. The main advantage of the
by fault resistance, loading conditions, fault inception time, wavelet transform is that the band of analysis can be easily adjusted
fault type, and fault distance. Several techniques such as fuzzy and the waveforms obtained from wavelet transform can be
neural networks, artificial neural networks (ANNs), basis function displayed in both the time and the frequency domain. Application
neural networks, and backpropagation neural networks have been areas of the wavelet transform in power systems include power
presented with advance in computer technology to overcome quality, power system protection, power system transients, partial
these difficulties [1–3]. The techniques based on ANN integrated discharge, and transformer protection and condition monitoring.
with wavelet transform are very interesting for transmission line In Ref. [10], a survey of wavelet transform applications in power
protection. In Ref. [4] an overall survey of ANN application system protection is given.
in the area of power system protection such as transmission This paper proposes an improved technique for fault type
line protection, power transformer protection, detection of high predication of parallel transmission lines taking into account the
impedance faults, is given. The main disadvantage of the neural mutual coupling effect between the two circuits, fault resistance,
network technique is that it requires considerable amount of loading conditions, fault inception time, and fault distance. Only
training effort for good performance. the phase currents are used to predict the fault type. Therefore, the
In recent years, support vector machine (SVM) has been intro- size of feature matrix and training time are reduced. Consequently,
duced for solving pattern recognition problems such as face recog- the performance of the proposed method is increased by using this
nition, signal and image processing, and fault diagnosis [5–8]. new distinguished dataset.
One of the advantages of SVM is the formulation of its learn- The simulation results show that the proposed method is very
ing problem, leading to the quadratic optimization task. It greatly simple and effective in fault diagnosis, and is independent of
reduces the number of operations in the learning mode, and hence fault resistance, loading conditions, fault inception time, and fault
the SVM algorithm is usually much faster for large datasets [9]. distance.
SVM-based classifiers have better generalization characteristics
than ANN-based classifiers. Therefore, the performance of SVM- 2. Methodology
based classifiers is not based on the number of features, and this
property is very beneficial in fault diagnostics because the number 2.1. General The proposed SVM-based method consists
of two essential stages. In the first stage, one cycle of post
fault phase currents is analyzed with the aid of discrete wavelet
a Correspondence to: A. Saber. E-mail: [email protected] transform to obtain the frequency detail coefficients. Only two
Electrical Power and Machines Department, Faculty of Engineering, Cairo detail coefficients for each phase current are used to reduce the
University, Giza, Egypt dataset. Then, the absolute values for each detail coefficients are

© 2015 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.
 A. SABER, A. EMAM, AND R. AMER

summed to minimize the size of training matrix. In the second For any function f (t), the discrete wavelet transform (DWT) is
stage, the training data are collected and SVM is employed written as [13,14]
to establish the fault classification unit. After that, the fault ∞
classification unit is tested for different fault cases. DWTψ f (m, n) = ∫ f (t)ψm,n (t)dt (4)
−∞

2.2. Feature extraction using discrete wavelet trans- In DWT, filters of different cut-off frequencies analyze the
form (DWT) Wavelet transform extracts the information in signal at different scales [15]. The signal is passed through a series
the time domain by decomposing the transient signal with a short of high-pass filters to analyze the high frequencies and a series of
window scale for high-frequency band and a long window scale low-pass filters to analyze the low frequencies. Hence the signal
for low-frequency band using scale and shift technique, contrary (S ) is decomposed into two types of components approximation
to Fourier transform [11]. So, it is effective for analyzing transient (A) and detail (D). Approximation is the high-scale, low-frequency
signals. The continuous wavelet transform (CWT) of transient sig- component of the signal, and detail is the low-scale, high-
nal f (t) is defined as [12]: frequency component of the signal. The decomposition process can
be iterated, with successive approximations that are decomposed
∞
CWTψ f (a, b) = Wf (a, b) = ∫ f (t)ψa,b (t)dt (1) in turn, so that one signal is broken down into many lower
−∞ resolution components. This is called the wavelet decomposition
where ψa,b (t) is the mother wavelet function, which is defined as tree. As decompositions are done on higher levels, lower frequency
  components are filtered out progressively. Figure 1 shows a nine-
1
−2 t −b level wavelet decomposition for one cycle post fault discrete
ψa,b (t) = |a| ψ (2)
a phase current that used in this paper, with 50 kHz sampling
frequency.
In this paper, instead of continuous scaling and shifting, the
mother wavelet is scaled and shifted discretely by selecting a = a0m
and b = nb0 a0m , where a0 and b0 are constants with a0 > 1, 2.3. Support vector machine Traditional neural net-
b0 > 1, m, n ∈ Z and Z is the set of positive integers. Then the work approaches have difficulties with generalization, and produce
mother wavelet function (ψ) is defined as models that can overfit the data. This is a result of the optimization
  algorithms used for parameter selection and the statistical mea-
−2
m
t − nb0 a0m
ψm,n (t) = a0 ψ (3) sures used for best model selection [16]. The formulation embodies
a0m the structural risk minimization (SRM) principle, which has been

Signal(S), 50 KHZ

A1 D1
0–12.5 KHZ 12.5–25 KHZ

A2 D2
0–6.25 KHZ 6.25–12.5 KHZ

A3 D3
0–3.125 KHZ 3.125–6.25 KHZ

A4 D4
0–1562 HZ 1562–3125 HZ

A5 D5
0–781 HZ 781–1562 HZ

A6 D6
0–390 HZ 390–781 HZ

A7 D7
0–195 HZ 195–390 HZ

A8 D8
0–97 HZ 97–195 HZ

A9 D9
0–48 HZ 48–97 HZ

Fig. 1. Wavelet decomposition tree

44 IEEJ Trans 11: 43–48 (2016)

 DWT AND SVM BASED PARALLEL TL FAULTS CLASSIFICATION

DWT for obtaining the frequency of detail component of the phase

currents.
Figures 4–6 show the current and voltage waveforms for double

6.7 m
5.1m line to earth fault in the first circuit seen at substation 1. Note that:
one cycle = 50 kHz/50 Hz = 1000 samples.
6.6 m
3.2. Fault classification (FC) unit The proposed tech-
35.1 m

nique uses only the phase currents to predict fault types. One cycle

24.9 m
post fault of the phase currents is considered to reduce the dataset.
5.4 m
157.7 m

Daubechies-4 (db4) was selected as a mother wavelet because it

has the best similarity to fault transients [18–20].
In more detail, Fig. 1 shows the nine-level wavelet decomposi-
tion tree for one cycle post fault discrete phase current that is used
Fig. 2. 220 kV transmission line configuration in this paper, at 50 kHz sampling frequency. For the one-level
wavelet decomposition tree, there are approximate coefficient val-
ues (A1) containing lower frequency components (0–12.5) kHz
shown to be superior to traditional empirical risk minimization
and detail coefficient values (D1) containing high-frequency com-
(ERM) principle employed by conventional neural networks [16].
ponents (12.5–25) kHz. Also for the two-level wavelet decompo-
SVM classifiers have good generalization properties compared to
sition tree, there are approximate coefficient values (A2) contains
conventional classifiers because the structural misclassification risk
lower frequency components (0–6.25) kHz and two detail coef-
is minimized in training the SVM classifier, whereas the empiri-
ficient values (D1, 12.5–25 kHz; D2, 6.25–12.5) kHz. Thus for
cal risk is minimized in training the traditional classifiers [17]. In
a nine-level wavelet decomposition tree, there are approximate
Ref. [9] detailed information about SVM is provided.
coefficient values (A9) containing lower frequency components
(0–48) Hz and nine detail coefficient values from D1 to D9.
3. Results and Discussion Only the detail coefficient values corresponding to level 8 (sec-
ond and third harmonic) (97–195) Hz and level 9 (fundamental
3.1. System simulation The system studied consists of component) (48–97) Hz are extracted. After that, the summations
220 kV, 50 Hz double circuit transmission lines, with sampling
of the absolute values of D8 and D9 detail coefficient values
frequency 50 kHz, 80 km in length, connected to a source at each
are used as inputs to the SVM as the summations of the abso-
end. Short-circuit capacity of the equivalent Thevenin sources on
lute values of D8 and D9 detail coefficient values are higher
the two sides of the line is 2 GVA and X /R is 15. All components
than other levels and they satisfy the characteristic relationships
were modeled by the MATLAB Simulink and SimPowerSystem
for classification purposes. Thus, 2 features are employed for
toolbox (MathWorks, Natick, Massachusetts, USA). The transmis- each phase current, and for the parallel lines total inputs are
sion line was simulated by the distributed parameter line model 12 features.
taking into account the mutual coupling between the parallel trans-
mission lines. The actual transmission line configuration is shown The fault classification unit is divided into two sections:
in Fig. 2 and its parameters are shown in Tables I and II”.
The whole modeling system is shown in Fig. 3, where it is a) First classification unit (FC1): FC1 was used to predict single
regulated by two control blocks. The first control block is the line and double line to earth faults. FC1 will operate if the
‘fault control unit’, which is responsible for the occurrence of summations of the D8 or D9 detail coefficient values for the
the fault at different fault inception angle. The second block is phase currents (A, B, and C) are not equal zero.
‘fault analysis unit’, which allows continuous signal analysis using
The training and testing data for FC1 were carried out at
different fault types (Nf ), fault inception angle (Tf ), and fault
Table I. 220 kV transmission line parameters distance (Df ). Different fault resistances (Rf ) (1–100 ) and
power angles (δ) (5–25◦ ) are included in the training and
Voltage level (kV) 220
testing data of the fault classification unit. The size of feature
Span (m) 360
No of circuits 2 vector of FC1 is 852 rows and 12 columns.
No. of conductors per phase 2 The training data are 612 fault cases, and the other 240 fault
Diameter of conductor (mm) 27 cases are used for testing of the FC1 classifier. Every fault type
Sub-conductor spacing (cm) 30 is classified with number as shown in Table III.
No. of sky wires 1
Diameter of sky wire (mm) 13.4 b) Second classification unit (FC2): FC2 was used to predict
Sag (m) 9.5 double line and three line faults. FC2 will operate if the
summations of D8 or D9 detail coefficient values for the phase
currents (A, B, and C) are equal zero.
Table II. Double circuit transmission line parameters
The size of feature vector of FC2 is 578 rows and 12 columns.
Resistance (/km) Positive 0.0258 The training data are 408 fault cases, and the other 170 fault
Zero 0.152 cases are used for testing of FC2 classifier. The fault types and
Zero-mutual 0.126 class labels for FC2 are shown in Table IV. The least squares
Inductance (H/km) Positive 0.0010591 SVM Toolbox (LSSVM) was used for classification of FC1
Zero 0.0029834 and FC2 [21].
Zero-mutual 0.0018016
Capacitance (F/km) Positive 11.132e-09
Zero 7.6415e-09
Zero-mutual –2.2853e-09 3.3. Support vector machine classifier The free SVM
parameters are regularization parameter (γ ) and the squared

45 IEEJ Trans 11: 43–48 (2016)

 A. SABER, A. EMAM, AND R. AMER

Fault analysis unit

Ia1
Powergui
Discrete wavelet transform (DWT) The extracted features
Ib1
Discrete,
Ts = 2e-005 s.
Ic1

Ia2

Ib2

Ic2

CTs 220 kV – 80 km TL
+ -i
Substation 1 Substation 2
+ -i
A A
+ -i
B B

C + -i C

Short circuit level + -i Short circuit level

2GVA 2GVA
+ -i
X/R=15 X/R=15

R = 0.0258 /km

2
Breaker X = 0.3327 /km
B = 3.497e–6 /km

1
c
Fault Control
Out1 R

Fig. 3. The whole modeling system

× 104
1
Ia1
Fault instant Ib1
0.8
Ic1
0.6

0.4

0.2
Currents (A)

–0.2

–0.4

–0.6

–0.8

–1
0 500 1000 1500 2000 2500
Samples

Fig. 4. Current waveforms of the first circuit for phases A1–B 1–G fault

bandwidth (σ ) [21]. A common method to tune the free parameters which represents fault type class. The 10th row in Table V
is to use cross validation to select the best parameters from a shows double line to earth fault in the first circuit (A1–B 1–G)
preselected set. The best results were obtained by selecting γ = 40 at power angles (δ) = 9◦ , Rf = 76 , Tf = 4 ms, and Df =
and σ = 1.5 for FC1 and γ = 100 and σ = 3 for FC2. 11 km, and the sixth column equals 4, which represents the fault
type class.
3.4. Results of fault classification unit Some testing The first row in Table VI shows double line fault in the first
cases are shown in Tables V and VI for different fault types (Nf ), circuit (A1–B 1) at power angles (δ) = 7◦ , Rf = 24 , Tf = 4 ms,
power angles (δ), fault resistances (Rf ), fault inception times (Tf ), and Df = 80 km, and the fault type class equals 7. The last row in
and fault distances (Df ). Table VI shows three line fault in the first circuit (A1–B 1–C 1) at
For example, The first row in Table V shows single line power angles (δ) = 8◦ , Rf = 15 , Tf = 4 ms, Df = 46 km, and
to earth fault in the first circuit (A1–G) at power angles (δ) the fault type class equals 10.
= 9◦ , Rf = 4 , Tf = 5 ms, and Df = 56 km, and the sixth The same testing cases were applied for the second circuit, and
column is the SVM output that indicates fault type and equals 1, the same results were achieved.

46 IEEJ Trans 11: 43–48 (2016)

 DWT AND SVM BASED PARALLEL TL FAULTS CLASSIFICATION

800
Fault instant Ia2
600 Ib2
Ic2

400

200
Currents (A)

–200

–400

–600

–800
0 500 1000 1500 2000 2500
Samples

Fig. 5. Current waveforms of the second circuit for phases A1–B 1–G fault

× 105
3
Va
Fault instant Vb
Vc
2

1
Voltages (V)

–1

–2

–3
0 500 1000 1500 2000 2500
Samples

Fig. 6. Voltage waveforms for phases A1–B 1–G fault

Table III. The output data of first classification unit Table IV. The output data of second classification unit

Circuit Type Label Circuit Type Label

Circuit 1 L1–G A1–G 1 Circuit 1 L1–L1 A1–B 1 7
B 1–G 2 A1–C 1 8
C 1–G 3 B 1–C 1 9
L1–L1–G A1–B 1–G 4 L1–L1–L1 A1–B 1–C 1 or A1–B 1–C 1–G 10
A1–C 1–G 5 Circuit 2 L2–L2 A2–B 2 17
B 1–C 1–G 6 A2–C 2 18
Circuit 2 L2–G A2–G 11 B 2–C 2 19
B 2–G 12 L2–L2–L2 A2–B 2–C 2 or A2–B 2–C 2–G 20
C 2–G 13
L2–L2–G A2–B 2–G 14
A2–C 2–G 15
to predict the fault type. Then, the summations of the absolute
B 2–C 2–G 16
values for each detail coefficients have been applied to reduce
the learning data required for SVM. Different fault inception
time, fault types, and fault locations have been included in the
4. Conclusion
training and testing of the fault classification unit. The proposed
The proposed technique based on DWT and SVM has showed technique has taken into account different fault resistances (1–100
successful performance for fault type classification. DWT has ) and different power angle (5–25◦ ), and can be extended to
been employed to extract the frequency detail components of the other power system protection problems such as transmission line
phase currents. Only two detail coefficients have been employed fault location.

47 IEEJ Trans 11: 43–48 (2016)

 A. SABER, A. EMAM, AND R. AMER

Table V. Testing data for single line and double line to earth (8) Vapnik V. The support vector method of function estimation. In
fault in the first circuit Nonlinear Modeling: Advanced Black-Box Techniques. Suykens J,
Vandewalle J (eds), Kluwer Academic Publishers; AT&T Labs-
Fault Research, USA; 1998; 55–86.
classification (9) Ravikumar B, Thukaram D, Khincha HP. Fault diagnosis in power
Nf δ Rf () Tf (ms) Df (km) output transmission systems using support vector machines. International
Journal of Emerging Electric Power Systems 2007; 8(4):1–32.
A1–G 9 4 5 56 1 (10) Rosa M, Nelson V. An overview of wavelet transforms application
A1–G 25 60 8 60 1 in power systems. 14th PSCC, Sevilla; 2002; 24–28.
A1–G 9 71 3 17 1 (11) Misiti M, Misiti Y, Oppenheim G, Poggi JM. Wavelet Toolbox Users
B 1–G 23 37 6 14 2 Guide. The MathWorks, Natick, Massachusetts, USA, Inc: 2001.
B 1–G 18 23 2 34 2 (12) Tag Eldin EM. Characterization of power quality disturbances based
B 1–G 15 17 9 53 2 on wavelet transforms. International Journal of Energy Technology
C 1–G 15 64 2 16 3 and Policy 2006; 4:74–84.
C 1–G 11 81 1 10 3 (13) Fikri M, El-Sayed MAH. New algorithm for distance protection of
C 1–G 23 22 3 37 3 high voltage transmission lines. IEE Proceedings C Generation,
A1–B 1–G 9 76 4 11 4 Transmission and Distribution 1988; 135:436–440.
A1–B 1–G 27 21 6 48 4 (14) Sadeh J, Ranjbar AM, Hadjsaid N, Feuillet R, Tran-Quoc T. New
A1–B 1–G 22 2 4 7 4
method for fault location in power transmission lines using one ter-
A1–C 1–G 10 97 5 14 5
minal data. International Conference on Electric Power Engineering,
A1–C 1–G 18 59 5 17 5
1999. Power Tech Budapest 99, August 29 1999–September 2 1999,
A1–C 1–G 11 62 5 56 5
B 1–C 1–G 5 61 8 19 6 Budapest, Hungary, 1999; p. 254.
B 1–C 1–G 24 78 3 31 6 (15) Yusuff AA, Fei C, Jimoha AA, Munda JL. Fault location in a series
B 1–C 1–G 18 53 9 52 6 compensated transmission line based on wavelet packet decomposi-
tion and support vector regression. Electric Power Systems Research
2011; 81(7):1258–1265.
Table VI. Testing data for double line and three line fault in the (16) Gunn SR. Support Vector Machines for Classification and Regression
Technical Report, University of Southampton, 1998.
first circuit
(17) Dash K, Samantaray SR, Ganapati P. Fault classification and section
Fault identification of an advanced series-compensated transmission line
using support vector machine. IEEE Transactions on Power Delivery
classifica-
2007; 22:67–73.
tion (18) Truchetet F, Laligant O. Wavelets in industrial applications SPIE
Nf δ Rf () Tf (ms) Df (km) output Proceedings 2004; 5607:1–14.
A1–B 1 7 24 4 80 7 (19) Ekici S, Yildirim S, Poyraz M. Energy and entropy-based feature
A1–B 1 23 27 8 44 7 extraction for locating fault on transmission lines by using neural
A1–B 1 19 93 8 56 7 network and wavelet packet decomposition. Expert Systems with
A1–C 1 12 94 8 68 8 Applications 2008; 34(4):2937–2944.
A1–C 1 22 61 3 63 8 (20) Ekici S, Yildirim S. Fault location estimation on transmission lines
A1–C 1 6 48 7 40 8 using wavelet transform and artificial neural network. Proceedings
B 1–C 1 12 85 3 42 9 of the 2006 International Conference on Artificial Intelligence, ICAI
B 1–C 1 11 81 2 9 9 2006, Las Vegas, Nevada, USA, vol. 1, 2006; 181–184.
B 1–C 1 25 68 3 51 9 (21) Suykens JAK, Van Gestel T, De Brabanter J, De Moor B, Vande-
A1–B 1–C 1 14 7 8 52 10 walle J. Least Squares Support Vector Machines. World Scientific:
A1–B 1–C 1 10 77 2 80 10 Singapore; 2002. Available at: http://www.esat.kuleuven.be/sista/
A1–B 1–C 1 20 5 5 65 10 lssvmlab/. Accessed July 2012.
A1–B 1–C 1 21 63 5 74 10
A1–B 1–C 1 13 55 3 37 10 Ahmed Saber (Non-member) received the M.Sc. degree in Elec-
A1–B 1–C 1 8 15 4 46 10 trical power and Machines Engineering from the
Faculty of Engineering, Cairo University (CU),
Egypt, in 2013. He is presently pursuing the
References Ph.D. degree at the Faculty of Engineering, CU,
and works as a Teaching Assistant with the Elec-
(1) Khorashadi-Zadeh H. Artificial neural network approach to fault trical Power and Machines Department, CU.
classification for double circuit transmission lines. Transmission and Ahmed Emam (Non-member) received the B.Sc., M.Sc. and the
Distribution Conference and Exposition: Latin America, IEEE/PES, Ph.D. degrees in Electrical Engineering from
8–11 November 2004, 2004; 859–862.
Cairo University (CU), Egypt, in 1998, 2002,
(2) Thukaram D, Khincha HP, Vijaynarasimha HP. Artificial neural net-
work and support vector machine approach for locating faults in radial
and 2010, respectively. Currently, he is with
distribution systems. IEEE Transactions on Power Delivery 2005; the Faculty of Engineering, CU. His research
20:710–721. interests include electric and magnetic fields,
(3) Razi K, Hagh NT, Abrabian G. High accurate fault classification of transient overvoltages, and digital protection.
power transmission line using fuzzy logic. International Power Engi- Rabah Amer (Non-member) received the Ph.D. degree in high
neering Conference (IPEC 2007), 3–6 December 2007, Singapore,
Voltage Engineering from the Faculty of Engi-
2007; 42–46.
neering, Cairo University (CU), in 1983. He is
(4) Kezunoic M. A survey of neural net application to fault analysis.
Engineering Intelligent Systems 1997; 5(4):185–192. presently an Emeritus Professor with the Faculty
(5) Vapnik VN. The Nature of Statistical Learning Theory. Springer of Engineering, CU. He has worked on earth-
Verlag: New York; 1995. ing, power system protection, lightning protec-
(6) Vapnik VN. The Statistical Learning Theory. Wiley, New York; 1998. tion systems, and the testing and commissioning
(7) Salat R, Osowski S. Accurate fault location in the power transmission of electrical equipment. Also, he is a consultant for many industrial
line using support vector machine approach. IEEE Transactions on companies.
Power Systems 2004; 19:879–886.

48 IEEJ Trans 11: 43–48 (2016)