Indonesian Journal of Electrical Engineering and Computer Science
Vol. 6, No. 2, May 2017, pp. 310 ~ 317
DOI: 10.11591/ijeecs.v6.i2.pp310-317
310
Classification of Power Quality Disturbances at
Transmission System Using Support Vector Machines
Shahrani Shahbudin, Zaki Firdaus Mohmad, Saiful Izwan Suliman*, Murizah Kassim,
Roslina Mohamad
Centre for Computer Engineering, Faculty of Electrical Engineering, Universiti Teknologi MARA,
Shah Alam, Selangor, Malaysia
*Corresponding author, e-mail:
[email protected]
Abstract
Power Quality has become one of the important issues in modern smart grid environment. Smart
grid generally utilizes computational intelligence method from the generation of electricity to electricity
distribution to the customers. This is done for the safety, reliability, tenacity and efficiency of the system.
The classification of power disturbances has become a major topic in maintaining power quality. These
disturbances occur due to faults, natural causes, load switching, energizing transformer, starting large
motor, as well as utilization of power electronic devices. The key issue is about maintaining the continuous
supply of electricity to the end-users without any problem. If a problem occurs, it might increase the
production cost significantly especially to large-scale industries. In this paper, S-transform is used to
extract distinctive features of real data from transmission system, and Support Vector Machine was utilized
to classify four types PQ disturbances namely, voltage sag, interruption, transient and normal voltage.
Results obtained indicate that performance of the One Against One classifier produces high accuracy
using k-fold cross validation and RBF kernel.
Keywords: power quality disturbances, S-Transform, k-fold cross validation, One-against-One SVM
Copyright © 2017 Institute of Advanced Engineering and Science. All rights reserved.
1. Introduction
The increasing demands of electricity nowadays has become a real challenge to the
power producer to maintain the quality of continuous power supply. Low quality of continuous
power supply will affect many aspects especially production cost to the industries.
Computational intelligence (CI) methods can be utilized as a tool to help the power quality
management to maintain and improve the electricity system. CI is mainly used for fault
detection, fault classification and section identification of fault in transmission lines for analysis
of disturbance that can lead to any fault [8]. Moreover, power quality disturbances classification
in transmission system has also become one of the most researched area in this area. The
significance of this study is to improve the power quality (PQ) monitoring system in minimizing
fault occurrence along the transmission lines.
Fault events in transmission line occur due to several factors such as bad weather,
human activities and accidents. Lightning is one of the main cause of voltage disturbance (e.g.
transient overvoltage). It can lead to line-to-ground fault event due to lack of insulation around
cable [1]. The occurrences of fault events have resulted to power quality disturbances in the
transmission lines. There are many types of disturbances that may affect power quality. These
include voltage sag, voltage swell, transient, harmonic, interruption and flicker. These
disturbances occur due to many factors such as faults, natural causes, load switching,
energizing transformer, starting large motor, and high utilization of power electronic devices [2].
The diagnosis of power quality disturbances is important for the improvement of monitoring
process in power transmission system particularly in the context of voltage stability. Voltage
stability is the ability of power system to remain at acceptable voltage at all busses under
normal operating system and after being affected by the disturbance [3]. Failure to maintain
voltage stability can lead to the occurrence of voltage collapse in transmission system.
Support Vector Machine (SVM) is one of the utilized technique in solving classification
problems. The goal of SVM is to determine a classifier that minimizes empirical risks namely
training set error and confidence interval. This confidence interval corresponds to the
Received January 7, 2017; Revised March 15, 2017; Accepted April 1, 2017
311
ISSN: 2502-4752
generalization or test set error [4, 5]. Support vector machine classifiers are based on the
statistical learning theory. The method is suitable for large size of training data. Compared to
other classification methods, no threshold is to be determined in SVM. An upper bound will be
utilized for the generalization performance (i.e. the performance for the test set). This upper
bound will be determined based on the statistical learning conducted [6, 7]. The applications of
support vector machine have been utilized in solving many problems occurring in power system
operation. These include power disturbances classification methods, fault classification and
identification faulty section occurred in transmission lines [8].
Various classification techniques have been successfully applied in PQ disturbances
diagnosis such as wavelet transform and SVM [2], [5], [9-11], S-Transform and SVM [12],
wavelet Transform and optimized ANN [13], curvelet Transform and SVM [14], Quarter-Sphere
Support Vector Machine (QSSVM) [15], Hybrid neural network [1] and ensemble technique [16].
Support vector machine is a well-known supervised classification technique that can classify
any type of objects or signals accurately. This method has been proven to be a good tool in
solving many problems in transmission lines and machinery faults [2]. It is an effective
classification tool for fault classification in power system.
There are two strategies for multi-class SVM which are “one-against-one” and “oneagainst-all”. However, based on the studies conducted by many researchers found in the
scientific literature, the “one-against-one” strategy takes less computational time during the
training process. It performs well for problems with very large number of classes [11], [17, 18].
Therefore, “one-against-one”- based multi class SVM (OAO-SVM) approach was chosen as the
classification method in this study. The objective of this study is to classify the disturbances for
real data from transmission system in Malaysia using “one-against-one” multi-class support
vector machine technique.
2. Research Method
The proposed method in this study is divided into three stages. It starts with the preprocessing stage, followed by feature extraction stage and finally the implementation of
multiclass SVM in the classification phase. The processes involved in this study utilizing SVM is
shown in Figure 1.
Pre-processing
Feature extraction
S-transform
Multi-class SVM
classifier
GUI
Development
Figure 1. Flow of the proposed method
IJEECS Vol. 6, No. 2, May 2017 : 310 – 317
ISSN: 2502-4752
IJEECS
312
2.1. Pre-Processing
A pre-processing of the signals is required to normalize raw data. This is because
signals are collected from various voltage levels in the distribution system [19]. The collected
data are 3-phase voltages (UR, UY, UB), 3-phase currents (IR, IY, IB) and neutral line current
(IN). The samples data collected are shown in Table I. However, only values for 3-phase
voltages were utilized during this process. Meanwhile, four types of disturbances were classified
in this study which are voltage sag, interruption, transient and normal voltage.
Table 1. Sample of the Real Data of Transmission System for One Feeder
(t)
UR
UY
UB
IR
IY
IB
IN
1
-15008
17632
-2528
-80
96
0
0
2
-15632
17152
-1408
-80
80
16
16
3
-16224
16576
-288
-80
80
16
16
4
-16768
15952
864
-96
80
32
16
5
-17280
15264
2016
-96
80
32
16
6
-17792
14496
3200
-96
80
32
16
7
-18224
13680
4400
-96
64
48
0
8
-18560
12800
5568
-96
64
48
16
9
-18816
11840
6768
-96
64
48
16
10
-18960
10832
7952
-96
64
64
16
2.2. Feature Extraction using S-transform
After the pre-processing process, S-transform process was applied in the feature
extraction stage. S-transform is one of the feature extraction methods based on time-frequency
analysis. It extracts the pre-processing data into the most salient features that represent the
power quality phenomenon [20]. The following was applied for the s-transform process:
(1)
S-transform will generate a matrix known as S matrix. Each row of the s-matrix
represents specific frequency, and each column corresponds to a sampling point of the original
signal. Every element from this matrix is a complex value. According to Wenda et al. [21],
statistical techniques are used to the amplitude of contour matrix of S-transform by using
maximum amplitude and frequency amplitude plot. The resulting features are formulated as
follows.
Amplitude factor, F1 as the first feature with the given equation:
(2)
where,
std1
std2
norm1
norm2
: maximum value from standard deviation of distortedsignal.
: minimum value from standard deviation of distortedsignal.
: maximum value of normal signal.
: minimum value of normal signal.
The second feature is given as :
(3)
The third feature is given as :
))
(4)
Classification of Power Quality Disturbances at Transmission System … (Saiful Izwan Suliman)
313
ISSN: 2502-4752
where,
ds = absolute value of S-transform from distorted signal
The forth feature is an absolute value of S-transform from highest amplitude of
frequency (fm). The equation is given as follows,
(5)
The fifth feature is the mean of the square root STD from amplitude of st-matrix,
(6)
All of the five features were used as the inputs to the multi-class OAO-SVM.
2.3. Multi-Class SVM Classifier
SVM was originally developed for binary classification which can only classify two
classes of data. Training data set must be in form of (x1, y1),…, (xl , yl) where xi RN is a feature
vector and yi {-1,1} is a label of class. The largest margins that separate these data are defined
as decision surface. However, this decision surface is not generated by the input space, but is
decided in the high-dimensional feature space. Kernel function was used because it is very
useful in nonlinear data. Equations (7) - (9) show how binary classification took place [22].
(7)
where,
α
y
K
: Lagrange multiplier
: Class label
: The similarity between pattern xi and xj from the stored training set.
: Two different patterns of inputs
Subject to the constrains,
(8)
After finding the optimal value, αi , the decision boundary needs to be constructed in the
form of :
(9)
where,
x
xi
b
: the class from the sign of f(x)
: support vector correspond to αi ≠0
: threshold of the decision boundary from origin
In making SVM computationally very efficient, the numbers of support vectors are
considerably lower than the number of training samples. The regularization parameter C,
controls the trade-offs between the margin and the size of the slack variables [6].
Multi-class SVM uses combination of binary SVMs for k number of classes. OAO-SVM
is involved in the construction of binary SVM classifiers for all pairs of classes. In overall, there
are [k (k-1)/2] pairs, which implies that there is only one binary SVM for one pair. The
classification of unknown input pattern is determined according to maximum voting of all SVMs
as shown in Figure 2.
In the classification stage, a model validation technique was utilized for assessing how
the results of a statistical analysis will generalize to an independent data set. There are various
techniques in dividing data set into training and testing set, such as Hold-out, k-fold cross
validation (k-fold CV), LeaveOut, and re-substitution. However, only two techniques were
implemented and tested in this work, which are Holdout and k-fold CV.
IJEECS Vol. 6, No. 2, May 2017 : 310 – 317
ISSN: 2502-4752
IJEECS
314
Hold-out validation technique splits the data without overlapping the training and testing
data. In k-fold CV, the training set is split into k smaller sets. A model is trained using the folds
as training data. The resulting model is then validated and tested on the remaining part of the
data. The training and testing processes were repeated for each different combination. In this
paper, the value of k was set from two until ten folds.
Class1 vs. Class2
Class1 vs. Class3
Class1 vs. Class4
Input Vectors
Max.
Voted win
Class2 vs. Class3
Class2 vs. Class4
Class3 vs. Class4
Figure 2. OAO-SVM classifier
Polynomial (Poly) and Radial Basic Function (RBF) are non-linear kernel function
commonly used in SVM classification. Equation (10) and (11) show the basic formula of Poly
and RBF.
Poly:
(10)
RBF:
(11)
3. Results and Analysis
3.1. S-transform Feature Extraction
Figure 3 depicts four plots for S-transform results for each disturbance. The first plot
shows the per unit voltage for the actual data and the contour of the s-matrix area shown in the
second plot. Final plot represents the maximum amplitude of the frequency for each sampling
points. The sample outputs of S-transform are tabulated in Table 2.
Table 2. Sample of Features Vectors Data Used in OAO-SVM
Types
Sag
Interruption
Transient
Normal
F1
1.0654
1.0402
1.0881
1.0728
F2
0.0335
0.0337
0.0537
0.0364
F3
0.0011
0.0007
0.0013
0.0012
F4
0.4999
0.4216
0.0621
0.5420
F5
0.0364
0.0182
0.5100
0.0182
Classification of Power Quality Disturbances at Transmission System … (Saiful Izwan Suliman)
315
ISSN: 2502-4752
(a) Normal voltage
(b) Transient
(c) Voltage Interruption
(d) Voltage Sag
Figure 1. S-transform results for each disturbance (a) Normal Voltage, (b) Transient,
(c) Voltage interruption, (d) Voltage sag
3.2. OAO-SVM Classifier
Several parameters need to be set up before the training process. In this analysis, for kfold CV, the value of k is set between 2 to 10, and the best average accuracy was recorded in
Table 3 and the results were compared with Holdout technique as shown in Table 4. Holdout
technique used 50% for both training and testing. In addition, two types of kernels were used for
both techniques, which are Polynomial (Poly) and Radial Basis Function (RBF). Table 5
represents the confusion matrix generated after classification stage using OAO-SVM using kfold CV with k=9.
The average accuracies for both techniques were recorded for ten readings. Based on
the data in Table V, the best k value of 9 was obtained when k-fold technique was utilized.
Accuracy rate of 100% was achieved using RBF kernel as compared to 99.623% for Poly
kernel. Meanwhile, in Holdout cross validation technique, Poly kernel performs better with
average accuracy of 99.582% as compared to 98.954% by RBF kernel. Therefore, it shows that
the combination of k-fold cross validation (k-fold CV) technique and RBF kernel produces the
superb accuracy rate.
Table 3. Accuracy Rate of K-Fold When K Varies from 2 Until 10
Value of k
2
3
4
5
6
7
8
9
10
IJEECS Vol. 6, No. 2, May 2017 : 310 – 317
Accuracy (%)
99.16
99.37
99.16
98.96
99.20
98.55
98.31
100.0
89.58
ISSN: 2502-4752
IJEECS
316
Table 4. Output Sample Of OAO-SVM Using K-Fold CV (K=9) With RBF Kernel
Type of disturbances
Sag
Interruption
Transient
Normal
Accuracy
Sag
5
0
0
0
Interrupt
0
11
0
0
Transient
Normal
0
0
0
0
14
0
0
23
100 %
Total(%)
100
100
100
100
Table 5. Comparison Results of the OAO-SVM with Two Cross Validations Techniques
Cross validation techniques
Kernel
Average accuracy (%)
K-fold
(K=9)
Poly
99.623
Holdout
RBF
100
Poly
99.582
RBF
98.954
3.3. Graphical User Interface (GUI)
The graphical user interface (GUI) for the proposed method was developed as
illustrated in Figure 4. It consists of two parts, which are the feature extraction stage on the left
side and classification stage on the right side. This model enables the user to conduct
experiments as well as analyseparts easily based on the specific requirements for the power
quality disturbances diagnosis system.
Figure 4. Developed Graphical User Interface
4. Conclusion
Continuous supply of electricity is very essential as almost all our daily activities depend
on electricity power. Big industries will be the most affected if electricity supply is interrupted as
machines and heavy equipment operate on electricity power. Disturbance in electricity supply
will have direct effect on the operational cost of a factory, thus will incur extra cost and reduce
the profit. Therefore, it is very important to have a mechanism to minimize the severity of supply
interruption in the event of a disturbance or voltage collapse. This study investigates the
application of support vector machine (SVM) for the classification of power quality disturbance
by using s-transform and One-Against-One SVM (OAO-SVM) classifier. Based on the presented
results and analysis conducted, the proposed method has managed to classify the disturbances
with 100% accuracy rate when combination of k-fold cross validation (k=9) and Radial Basis
Function (RBF) kernel was utilized. Therefore, OAO-SVM shows great potentials for future
power quality disturbances classification method.
Classification of Power Quality Disturbances at Transmission System … (Saiful Izwan Suliman)
317
ISSN: 2502-4752
Acknowledgment
The authors would like to express the gratitude to the Ministry of Higher Education,
Malaysia and UniversitiTeknologi MARA, Selangor, Malaysia for the financial support given for
this project (Fundamental Research Grant Scheme - FRGS) [File No : 600-RMI/FRGS 5/3
(109/2014)].
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
Huang NT, et al. Power quality disturbances classification by ensemble and hybrid Neural
networks. Power System Technology (POWERCON), 2010 International Conference on. IEEE. 2010.
Erişti H, Demir Y. Automatic classification of power quality events and disturbances using wavelet
transform and support vector machines. IET generation, transmission & distribution. 2012; 6(10):
968-976.
Ibe OG, Onyema AI. Concepts of reactive power control and voltage stability methods in power
system network. IOSR Journal of Computer Engineering. 2013; 11(2): 15-25.
Lin WM, Wu CH, Lin CH, Cheng FS. Classification of multiple power quality disturbances using
support vector machine and one-versus-one approach. Power System Technology. IEEE. Power Con
2006. International Conference on 2006: 1-8.
Tong W, Song X, Lin J, Zhao Z. Detection and classification of power quality disturbances based on
wavelet packet decomposition and support vector machines. In Signal Processing, IEEE 2006 8th
International Conference on 2006; 4.
Shawe-Taylor J, Cristianini N. Kernel methods for pattern analysis. Cambridge university press.
2004; 28.
Vapnik V. The nature of statistical learning theory. Springer science & business media. 2013; 29.
García VV, Gualdron CA, Plata GO. Obtaining patterns for classification of power quality
disturbances using biorthogonal wavelets, RMS value and support vector machines. Electrical Power
Quality and Utilisation. IEEE. EPQU 2007. 9th International Conference on 2007: 1-6.
Gao P, Wu W. Power quality disturbances classification using wavelet and support vector machines.
Intelligent Systems Design and Applications. Sixth International Conference on 2006; 1: 201-206.
Latha RS, Babu CS, Prasad KD. Detection & analysis of power quality disturbances using wavelet
transforms and SVM. International Research Journal of Signal Processing. 2011; 2(02): 58-69.
Yang BS, Han T, Hwang WW. Fault diagnosis of rotating machinery based on multi-class support
vector machines. Journal of Mechanical Science and Technology. 2005; 19(3): 846-59.
ÇÖTELİ R. A combined protective scheme for fault classification and identification of faulty section in
series compensated transmission lines. Turkish Journal of Electrical Engineering & Computer
Sciences. 2013; 21(Sup. 1): 1842-56.
Ijaz M, Shafiullah M, Abido MA. Classification of power quality disturbances using Wavelet Transform
and Optimized ANN. Intelligent System Application to Power Systems (ISAP), IEEE. 2015 18th
International Conference on 2015: 1-6.
Risnidar CB, Daut I, Yussof MI, Warman E, Hasan S, Nisja I. Investigation of Power Factor on
Harmonic Effect due to Types of Voltage Source. International Journal on Advanced Science,
Engineering and Information Technology. 2011; 1(5): 476-80.
Shahid N, Aleem SA, Naqvi IH, Zaffar N. Support vector machine based fault detection &
classification in smart grids. Globecom Workshops (GC Wkshps), 2012 IEEE 2012: 1526-1531.
Karasu S, Başkan S. Classification of power quality disturbances by using ensemble technique.
Signal Processing and Communication Application Conference (SIU). IEEE. 2016: 529-532.
Tripathi P, Sharma A, Pillai GN, Gupta I. Accurate fault classification and section identification
scheme in TCSC compensated transmission line using SVM. World Academy of Science,
Engineering and Technology. 2011; 60: 1599-605.
Thukaram D, Agrawal R. Discrimination of Faulted Transmission Lines Using Multi Class Support
Vector Machines. 16th National Power Systems Conference. 2010: 497-502.
Amirullah A, Penangsang O, Soeprijanto A. Power Quality Analysis of Integration Photovoltaic
Generator to Three Phase Grid under Variable Solar Irradiance Level. TELKOMNIKA
(Telecommunication Computing Electronics and Control). 2016; 14(1): 29-38.
Zhang X, Han G, Chen L. De-noising of Power Quality Disturbance Detection Based on Ensemble
Empirical Mode Decomposition Threshold Algorithm. TELKOMNIKA (Telecommunication Computing
Electronics and Control). 2013; 11(4): 637-644.
Wenda A, Hussain A, Samad SA, Mohamed R, Hannan MA. Web-based on mobile phone for
automatic classification of power quality disturbance using the S-transform and support vector
machines. TENCON 2010-2010 IEEE Region 10 Conference 2010: 1432-1437.
Duan KB, Rajapakse JC, Nguyen MN. One-versus-one and one-versus-all multiclass SVM-RFE for
gene selection in cancer classification. European Conference on Evolutionary Computation, Machine
Learning and Data Mining in Bioinformatics. Springer Berlin Heidelberg. 2007: 47-56.
IJEECS Vol. 6, No. 2, May 2017 : 310 – 317