Anda di halaman 1dari 9

Expert Systems with Applications 37 (2010) 40944102

Contents lists available at ScienceDirect

Expert Systems with Applications


journal homepage: www.elsevier.com/locate/eswa

A new algorithm for automatic classication of power quality events based on wavelet transform and SVM
Hseyin Eris ti a,*, Yakup Demir b
a b

Tunceli Vocational School, Tunceli University, Tunceli, Turkey Electrical and Electronics Engineering Department, Engineering Faculty, Firat University, Elazig, Turkey

a r t i c l e

i n f o

a b s t r a c t
This paper presents a new approach for automatic classication of power quality events, which is based on the wavelet transform and support vector machines. In the proposed approach, an effective single feature vector representing three phase event signals is extracted after signals are applied normalization and segmentation process. The kernel and penalty parameters of the support vector machine (SVM) are determined by cross-validation. The parameter set that gives the smallest misclassication error is retained. ATP/EMTP model for six types of power system events, namely phase-to-ground fault, phase-to-phase fault, three-phase fault, load switching, capacitor switching and transformer energizing, are constructed. Both the noisy and noiseless event signals are applied to the proposed algorithm. Obtained results indicate that the proposed automatic event classication algorithm is robust and has ability to distinguish different power quality event classes easily. 2009 Elsevier Ltd. All rights reserved.

Keywords: Automatic classication Power quality events Feature extraction Wavelet transform Support vector machines

1. Introduction Recently, power quality (PQ) has become one of the most important issues in modern power industry. The origin of events affected quality of power is mainly electromagnetic transients, which are common phenomena in every electric power systems. A number of causes of transients can be identied: lightning strokes, planned switching actions in the distribution or transmission system, self-clearing faults or faults cleared by current-limiting fuses, and the switching of end-user equipment. Transient phenomena are extremely critical since they can cause over voltages leading to insulation breakdown or ashover. These failures might trip any protection device initiating a short interruption to the supplied power. Excess current produced by transients may lead to complete damage to system equipment during the transient period. Moreover, if such disturbances are not mitigated, they can lead to failures or malfunctions of various sensitive loads in power systems and may be very costly. Therefore, there is a growing need to develop PQ monitoring techniques that can classify the potential sources of disturbances (El-Saadany, Abdel-Galil, & Salama, 2004; Gaing, 2004; Gargoom, Ertugrul, & Soong, 2007). The classication or recognition topic is an important issue for the development of the next generation of PQ monitoring equipment. Basically, it refers to the use of signal processing techniques to extract as few as possible and representative features from distur* Corresponding author. Tel.: +90 428 2131814; fax: +90 428 2121861. E-mail address: huseyineristi@gmail.com (H. Eris ti). 0957-4174/$ - see front matter 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2009.11.015

bance signals, followed by the use of a powerful and a simple technique to classify the detected disturbances. In the literature, the signal processing techniques are available for analyzing PQ disturbance. However, the most important of those are Fourier transform (FT), short-time Fourier transform (STFT) and wavelet transform (WT) for useful feature extraction from signals. Traditionally, the FT has been extensively used for analyzing the frequency contents of the signals. Besides, the FT is not an efcient analyzing tool for extracting the transient information of the nonstationary signals. Although the STFT divides the full-time interval into a number of small/equal-time intervals, can partly alleviate the problem, the STFT still has the limitation of a xed window width. It is difcult to detect the occurrence times for veryshort-duration and high-frequency signals. In order to avoid the disadvantages of both FT and STFT, the WT has been widely used for analyzing the PQ problems. The WT approach prepares a window that automatically adjusts to give proper resolutions of both the time and the frequency. In the approach, a larger resolution of time is provided to high-frequency components of a signal, and a larger resolution of frequency to low-frequency components. These features make the WT well suited for the analysis of the power system transients caused by various disturbances. A number of papers on automatic classication of PQ problems have been published during the last several years. These can be roughly divided into two groups as classication of disturbance waveforms and classication of events based on their origin. In rst group of the classication usually called disturbance classication, the detected disturbances are classied in a number of typical

H. Eris ti, Y. Demir / Expert Systems with Applications 37 (2010) 40944102

4095

classes such as voltage sags, voltage swells, and interruptions, etc. In several studies (Gaing, 2004; He & Starzyk, 2006; Huang, Negnevitsky, & Nguyen, 2002; Uyar, Yildirim, & Gencoglu, 2008) based on WT, the feature vectors representing the disturbance signals are obtained by using different feature extraction techniques to WT coefcients. Resulting feature vectors are applied as inputs to articial intelligent techniques such as articial neural network (ANN), fuzzy systems and support vector machine (SVM). They are showed that typical PQ disturbances are correctly classied. In second group of the classication called event classication, the underlying causes of disturbances such as faults, capacitor switching, and transformer energizing are classied. In the literature papers (Axelberg, Irene, & Bollen, 2007; Bollen, Gu, Axelberg, & Styvaktakis, 2007; Hong & Wang, 2005; Santoso, Lamoree, Grady, Powers, & Bhatt, 2000; Silva, Souza, & Brito, 2006; Styvaktakis, Bollen, & Gu, 2002) about event classication, it can be generally seen that three feature vectors representing an event are obtained by applying several feature extracting techniques to three phase voltage signals of detected event. In these papers, feature vectors are created using total harmonic distortion values, harmonic magnitudes, root mean square (RMS) values, energy values of WT coefcients, and WT coefcients obtained from event signals. As the classier, SVM, ANN, self-organizing mapping neural network, and expert system are used. In this paper, an automatic event classication algorithm is proposed for identifying the PQ events. Firstly, the wavelet multi-resolution analysis (WMRA) technique is employed to extract the features of the three phase event signals. Single feature vector for each event is obtained from another feature extraction technique applied three phase signal features. Secondly, the SVM is used to classify event types. Finally, event signals generated by the ATP/ EMTP are applied to the proposed algorithm for both noiseless and noisy environments. The experimental results showed that the proposed method could analyze and classify the event signals efciently. The novelty presented in this paper can be summarized as follows: An effective single feature vector representing a three-phase event signal is offered for automatic PQ event classication. Moreover, proposed algorithm has the advantage of monitoring all phases of the three-phase signal simultaneously and the great potential because of practical implementation and high classication accuracy. The remaining part of the paper is organized as follows. In Section 2, it is given several brief denitions of WT and WMRA. A short review to SVM classiers is presented in Section 3. Section 4 contains the proposed automatic classication algorithm. Simulations and discussion for proposed algorithm are given in Section 5. Finally, conclusions are discussed in Section 6.

the wavelet and scaling functions. The wavelet and scaling functions are, respectively, given as,

wk

p X 2 g n/2k n;
n

1 2

p X hn/2k n; /k 2
n

where n is integers and represent the number of samples. While the low-pass ltering produces the approximations Aj, the high-pass ltering produces the details Dj of the decomposition. The relationship of the approximation coefcients and detail coefcients between two adjacent levels are given as,

Aj1 k Dj1 k

X
n X n

hn 2kAj n; g n 2kAj n;

3 4

where j is frequency band level. The WMRA is based on decomposition of the original signal into different signals at various levels of resolution. First, the original signal is passed through the two lters producing the detail D1 and approximate A1 coefcients for j = 1. After down-sampling by a factor of 2, the approximate coefcients A1 are passed through the same lters again to obtain the coefcients for j = 2. After another down-sampling, the approximate coefcients A2 are then ltered again to obtain the next level of coefcients. This ltering operation continues in this way. A given signal f (k) is expanded in terms of its orthogonal basis of scaling and wavelet functions. In essence, it is represented by one set of scaling coefcients, and one or several sets of wavelet coefcients,

f k

X
n

A1 n/k n

XX
n j 1

Dj n2j=2 w2j k n:

There are many wavelet functions named as mother wavelets. The choice of mother wavelet is important because different types of mother wavelets have different properties. Several popular wavelet functions are Haar, Morlet, Coiet, Symlet and Daubechies wavelets. On the left-hand side of Fig. 1, the frequency band levels and frequency divisions of levels are shown for a signal which has 10 kHz sampling rate. On the right-hand side of Fig. 1, the typical PQ phenomena are listed (Chen & Zhu, 2007). It can be seen that

5 kHz Level 1 2.5 kHz Level 2 1.25 kHz Level 3 625 Hz

Nyquist frequency for 10 kHz sampling rate High-frequeny transient characteristics

Transient characteristics

2. Wavelet transform theory Due to its ability to extract time and frequency information of signal simultaneously, the WT is an attractive technique for analyzing PQ waveform. It is particularly attractive for studying disturbance or transient waveform, where it is necessary to examine different frequency components separately. WT can be continuous or discrete. Discrete WT (DWT) can be viewed as a subset of Continuous WT (CWT). In practical applications, the DWT is commonly used. The DWT is normally implemented by Mallats algorithm (Mallat, 1989). DWT uses the low-pass h(k) and the high-pass g(k) lters to divide the frequency-band of the input signal f(k) in respective low- and high-frequency components into octave bands (Daubechies, 1992). The low-pass lter h(k) is determined from the scaling function. The high-pass lter g(k) is determined from both

Level 4 312.5 Hz Level 5 156.25 Hz Level 6 78.12 Hz Level 7 38.06 Hz Level 8 19.03 Hz Approx. Level 0 Hz Main Frequency Harmonic characteristics

Fig. 1. Frequency division of DWT lters for 10 kHz sampling rate.

4096

H. Eris ti, Y. Demir / Expert Systems with Applications 37 (2010) 40944102


X X 1X yi yj ai aj K xi ; xj ai : 2 i1 j1 i1

the harmonic features appear at levels 56 because of having the bandwidth between 100 Hz and 250 Hz of dominant harmonics which are 2nd, 3rd, 4th and 5th. Oscillation frequency of a transient is typically between 300 and 1000 Hz. Therefore, transient features are viewed at levels 34. High-frequency transient features sight upper levels so high-frequency transients fall within a bandwidth between 2 kHz and 3 kHz. 3. Support vector machines for classication SVM is a powerful tool for solving pattern classication problems (Cortes & Vapnik, 1995; Vapnik, 1998). Given the training data (x1, y1), . . ., (x, y), x e RM, yi e {1, +1} for two class problem, SVM constructs the decision functions of form sgn((wTxi) + w0) by the maximum margin, where w is the normal vector of the separating hyperplane in the canonical form and wo is a bias term (Vapnik, 1998). The distances of the point closest to the hyperplanes of both 1 and +1 are calculated as 1/||w||. The separating margin is dened to be 2/||w||. In many practical cases in which data is corrupted by noise, the data may not be separable by a linear hyperplane. To allow to deviations from margin, the slack variables ni P 0 are introduced

Q a

K(x, xi) = xTxi in (8) is a linear kernel. One of the most common kernels is Gaussian kernel dened as

K x; xi expkx xi k2 =2r2

10

where r is the width parameter of Gaussian function. The decision surface of SVM is obtained by using only the training data xi with ai 0 lying closest to the decision boundary called as support vector,

g x wT uxi w0

X
ai 0

yi ai K x; xi w0 ;

11

where threshold can be obtained averaging over unbounded named support vectors with 0 < ai < C.

w0

1 X y wT uxi ; jU j i2U i

12

yi wT xi w0 P 1 ni ; i 1; . . . ; :

For the training data xi, if 0 < ni < 1, the data do not have the maximum margin but are still correctly classied. But if ni P 1, the data are misclassied by the optimal hyperplane. Thus the separation margin is increased by leaving intramargin the noise points occurring to near the boundaries or outlier points or both, so that generalization performance is improved. By accordingly, SVM constructs the constraint primal quadratic optimization problem that minimizes the training and generalization error by,

where U is the set of unbounded support vector indices. Multi-class SVM classiers can be obtained by combining two class SVMs. The most fundamental approaches are 1-against-r scheme (with r is the number of classes) and the 1-against-1 scheme. 1-against-r scheme users (r 1) two classiers: each machine is trained as classier for one class against all other classes. 1-against-1 scheme constructs a multi-class classier, i.e. (r(r 1)/2) two class machines are constructed. Each machine is trained as a classier for one class against other class. In order to classify test data, pair wise competition between all the machines is performed; each winner competes against another winner until a single winner remains. This nal winner determines the class of the test data (Abe, 2005; Vapnik, 1998). 4. Automatic power quality event classication algorithm

min Uw; n
w;n

X 1 T ni w wC 2 i1

7
The proposed algorithm for PQ event classication is divided into the three stages: pre-processing, feature extraction, and classication shown as Fig. 2. In this algorithm, the segmentation and normalization processes are applied to PQ event signals. A distinctive feature vector for per event is extracted by using the WMRA. PQ events are classied by using feature vectors obtained from feature extraction stage. Then, proposed algorithm is tested for both the noisy and noiseless PQ event signals. 4.1. The pre-processing stage In the pre-processing stage, segmentation and normalization are fullled. A PQ event is characterized by the event segments. An event segment is located between the two adjacent transition segments (Bollen & Gu, 2006). In the proposed algorithm, the PQ event signals are segmented as three cycles in the form of one cycle of the pre-event segment and two cycles of the event segment. In the normalization step, the event voltage waveform is converted to a relative scale (in pu, where pu stands for per unit) by dividing the input signal by the nominal RMS voltage. 4.2. The feature extraction stage The PQ event samples are too large for the classier input. The feature extraction being a special form of dimensionality reduction in the classication problems must be applied the event signals. Therefore, event signals are transformed into a reduced representation set. In this stage, the goal is to extract the distinctive features from event signals. The wavelet coefcients, detail and approximation,

s:t:

yi wT xi w0 P 1 ni ;

ni P 0; i 1; . . . ; ;

7a

where C is a parameter which controls the penalty incurred by each misclassied point in the training set. Generally, larger C values generate SVM models with smaller margin and better training accuracy as relatively smaller C values produce larger margin and better generalization accuracy. To solve the primal problem in (7), Lagrange function is rstly formulated (Bertsekas, 1999). Then its derivatives with respect to the primal variables w, n and wo are calculated and KKT conditions are satised (Bertsekas, 1999). Finally the obtained dual optimization problem is solved,

max Q a
a

X X 1X y y ai aj xT ai i xj 2 i1 j1 i j i1

s:t:

X i1

ai yi 0 0 6 ai 6 C ; i 1; . . . ; ;

8a

where ai is nonnegative Lagrange multipliers. SVM maps the inputs x into some higher dimensional space by means of a nonlinear feature mapping u(x) for solving the classication problem separated by only highly complex decision boundaries in the input space. Thus the problem changes into linearly separable case at the feature space. If only scalar product xT i xj in (8) is replaced by the kernel function K xi ; xj uT xi uxj assumed to be symmetric and positive denite (Vapnik, 1998), the dual problem subject to constraints in (8a) is rewritten as

H. Eris ti, Y. Demir / Expert Systems with Applications 37 (2010) 40944102

4097

Pre-processing Stage Get the voltage samples of three phases


Voltage Samples of three phases

Feature Extraction Stage Compute the DWT of the voltage samples


DWT coefficients of three phases

Classification Stage
For training For testing
Feature vector of event

Determine the SVM classifier parameters


Feature vectors of events

Segment the voltage samples


Voltage Samples of three phases

Compute the feature vectors of three phases


Feature vectors of three phases

Train the SVM classifier


Net

Normalize the all voltage samples

Compute the single feature vector for event

Classify the event

Fig. 2. Flowchart of the proposed automatic event classication algorithm.

obtained by applying the WMRA to event signals have effective feature information. The detail coefcients or approximation coefcients are impractical to apply as inputs to directly the classier because these coefcients have large data size. In order to reduce the feature dimension, the feature extraction methods are generally implemented to these coefcients at each decomposition level. In this paper, the detail and approximation coefcients are obtained by applying 8-level WMRA to the event signals. The energy at each decomposition level for per phase is calculated using the following equations:

Then, the feature vector (E) obtained from the feature extraction stage is applied to the SVM input.

5. Application of the proposed algorithm 5.1. PQ events In this paper, in order to evaluate the proposed algorithm, seven classes of PQ events are considered as the classication problem. These events are divided into two main groups: fault-induced events and switching events. Besides, normal class (C1) named undisturbed sinusoid is selected for non-event case. For fault-induced events, three major event types of faults which are phaseto-ground fault (C2), phase-to-phase fault (C3) and three-phase

Eij

N X n1

jDjn j2 ; jAn j2

j 1; . . . ; i 1; 2; 3

13 14

Eij1

N X n1

where j is decomposition level and N is number of the coefcients of detail or approximation at each decomposition level. i represents phase-1, -2 and -3. In this way, a nine-dimensional feature vector is constructed for 8-level wavelet decomposition. In the end of this feature extraction process, the three feature vectors (E1, E2, and E3) are obtained for per event. If these three feature vectors are applied to classier inputs, the classier system has some disadvantages such as excessive input dimension (total 27 inputs, 9 features for per phase), more computing time at both the training and testing processes. Besides, a nonsymmetrical event such as phase to ground fault has three different cases, which are AG, BG and CG. Therefore, classier needs the extra training process. Proposed algorithm can reduce these disadvantages mentioned above by an effective feature extraction technique as follows:

V (pu)
1 0.5 0.0 -0.5 -1 0.00 0.02 0.04 0.06 0.08 0.10

Time (s) V (pu)


1 0.5 0 -0.5

Ej

3 X n1

=2 E1 nj ;

j 1; . . . ; 1

15

-1 0.00 0.02 0.04 0.06 0.08 0.10

Therefore, a single feature vector (E) is obtained for each PQ event. The classier has a less input dimension (total 9 inputs). The feature variations in an event having different parameters such as switching time instant is similar to each other, even if event is nonsymmetrical. 4.3. The classication stage In the classication stage, the PQ event is classied using the SVM classier. Firstly, the SVM classier parameters (the parameter C and the kernel parameter) are selected by 10-fold cross-validation. By scanning the chosen parameter range, the parameters resulting in the lowest classication error are determined. In next step, the SVM classier is trained according to these parameters.

Time (s) V (pu)


1 0.5 0 -0.5 -1 0.00 0.02 0.04 0.06 0.08 0.10

Time (s)

Fig. 3. The event waveforms of: (a) single phase to ground fault, (b) phase to phase fault, and (c) three phase fault.

4098

H. Eris ti, Y. Demir / Expert Systems with Applications 37 (2010) 40944102

V (pu)
1 0.5 0.0 -0.5 -1 0.00 0.02 0.04 0.06 0.08 0.10

fault (C4) are chosen. The load switching event (C5), capacitor switching event (C6) and the transformer energizing event (C7) are selected for switching events. 5.1.1. Fault-induced events In the power system, faults are abnormal events which are not part of normal operation and unwanted by the network operator. The single-phase ground fault causes a large voltage drop in the faulted phase and a slight drop (solidly earthed system) or overvoltage (the nongrounded systems) in the non-faulted phases, as in Fig. 3a. A phase-to-phase fault leads to a voltage drop between the faulted phases, as shown in Fig. 3b. Due to three phase fault event, the voltage dips occur with an equal drop in the three phases shown as Fig. 3c. 5.1.2. Switching events Switching events are normal events that are part of the normal operation of the system. The switching of large loads gives an equal drop in the three phase voltages, as shown in Fig. 4a. In the system voltage, capacitor switching causes an oscillatory transient. Fig. 4b illustrates a typical capacitor switching event. The transformer energizing gives a large current, related to the saturation of the core ux, which results in a dip with different voltage magnitude in the three phases. The dip is further associated with a high harmonic distortion, including even harmonics. An example is shown in Fig. 4c. 5.2. Simulation of PQ events
0.02 0.04 0.06 0.08 0.10

Time (s) V (pu)


1 0.5 0.0 -0.5 -1 0.00 0.02 0.04 0.06 0.08 0.10

Time (s) V (pu)


1 0.5 0.0 -0.5 -1 0.00

Time (s)

Fig. 4. The event waveforms of: (a) load switching, (b) capacitor switching, and (c) transformer energizing.

In order to evaluate the performance of the proposed event classication algorithm, different PQ events were generated using ATP/ EMTP, which is commonly used a power system simulation tool (CanAm EMTP User Group., 1992). A test model based on the WECC 9-bus system which is usually used in power ow and transient stability studies (Zhang & Kezunovic, 2007) is implemented in

M1

CB

M2

2 1

Fig. 5. On-line diagram of WECC9-bus system.

H. Eris ti, Y. Demir / Expert Systems with Applications 37 (2010) 40944102 Table 1 Simulation variables data set used in the event simulation. Event types Fault-induced events Single phase to ground Phase to phase Three phase Event types Switching events Load switching Capacitor switching Transformer energizing Incidence angle () 0324 by step of 36 0324 by step of 36 0324 by step of 36 Switching angle () 0324 by step of 36 0342 by step of 18 0342 by step of 18 Loads (%) 50100 50100 50100 Loads For bus 10: 25% + 575% by step of 12.5, and other 100% 50100% For bus 10: 25% + 575% by step of 12.5, and other 100% Fault locations (mile) 12.5187.5 by step of 12.5 12.5187.5 by step of 12.5 12.5187.5 by step of 12.5 Fault resistance (X) 10200 by step of 10 110 by step of 10 110 by step of 10

4099

Fault type AGBGCG ABBCAC ABC

Capacitor size (MVAr) 1030 by step of 1

ATP. The Fig. 5 shows on-line diagram of the WECC 9-bus system. Fault-induced events, capacitor switching events, load switching events, and transformer energizing events were simulated in the lines 47, bus 4, bus 10, and bus 10, respectively. The voltage signals were taken at bus 7 and bus 10. The transmission line of the test model was modeled using distributed parameters without frequency dependence. Table 1 presents the variables used for simulation of PQ events in the test model. 5.3. Simulation results Two hundred event signals of each class were generated for training and testing at interval of simulation variables given in Table 1. The normal signals were varied from 0.95 pu to 1 pu. A hundred signals which randomly selected in two hundred event signals of each class were used for training the SVM. The remaining a hundred signals were used for testing the SVM. The sampling rate for the collection of PQ event data was taken equal to 10 kHz. The LIBSVM which is an efcient tool for the SVM related training was

used to evaluate the classication performance of extracted feature vectors for PQ disturbance signals (Chang & Lin, 2001). An RBF kernel was chosen for it can be have like a linear kernel or a sigmoid kernel under different parameter settings. The db4 wavelet which is widely used in electromagnetic transient analysis (Chen & Zhu, 2007) was selected as the wavelet function and PQ event signals

Table 2 Classication results in terms of confusion matrix for the noiseless event signals. True class C1 C2 C3 C4 C5 C6 C7 C1 C2 C3 C4 C5 0 0 0 0 100 0 0 C6 0 0 0 0 0 100 0 C7 0 0 0 0 0 0 100 Accuracy (%) 100 98 100 100 100 100 100 99.71

100 0 0 0 0 98 2 0 0 0 100 0 0 0 0 100 0 0 0 0 0 0 0 0 0 0 0 0 Overall success rate (%)

0.5 0 -0.5

1 0 -1

-1
-2

-4

-1.5 -2 -2.5
5 0 -5
1

-3 -4
2 1.5

-8

-12

2 1 0 -1 -2 -3

-10
0.5

-15 -20 -25 1 2 3 4 5 6 7 8 9


0 -0.5

1 0 -1 -2 -3 -4

Fig. 6. Variations in feature vectors for different event signals.

4100

H. Eris ti, Y. Demir / Expert Systems with Applications 37 (2010) 40944102

V (pu) Class C2
1 0.5 0.0 -0.5 -1 0.00 0.02 0.04 0.06 0.08 0.10

V (pu) Class C5
1 0.5 0.0 -0.5 -1 0.00 0.02 0.04 0.06 0.08 0.10

Time (s) V (pu) Class C3


1 0.5 0.0 -0.5 -1 0.00 0.02 0.04 0.06 0.08 0.10

Time (s) V (pu) Class C6


1 0.5 0.0 -0.5 -1 0.00 0.02 0.04 0.06 0.08 0.10

Time (s) V (pu) Class C4


1 0.5 0.0 -0.5 -1 0.00 0.02 0.04 0.06 0.08 0.10

Time (s) V (pu) Class C7


1 0.5 0.0 -0.5 -1 0.00 0.02 0.04 0.06 0.08 0.10

Time (s)
Fig. 7. PQ event signals added 30 dB noises.

Time (s)

were analyzed with an 8-level WMRA. The presented approach has been implemented in MATLAB. Fig. 6 illustrates the variations in feature vectors of training set. In this gure, variations are obtained by subtracting 1-pu normal signal feature vector from event feature vectors. In the level-7 which contains main frequency, the cases of voltage increase and voltage decrease are obviously seen for capacitor switching events and remaining event classes, respectively. Besides, distinctive features of transformer event and capacitor switching can be shown in levels 56 and levels 34, respectively. Table 2 gives the classication results in terms of confusion matrix for seven-class PQ event based on the proposed method. In this table, the correct classication results are tabulated at diagonal elements. The misclassication results are tabulated at non-diagonal elements. Two events of the phase to ground fault result in misclassied because of the similar features of the C2C3 classes while the other classes are exactly classied. As it can be seen from Table 2, the proposed method can effectively classify different types of PQ events. 5.4. The performance of proposed method under different noise conditions An another important issue related to the proposed method is noise which is generally omnipresent in an electrical power distribution network. Therefore, the performance of proposed algorithm was tested in different noise environments. Gaussian white noise is widely used in the researches on the PQ issues (He & Starzyk, 2006). Thirty decibels noise exhibits peak noise magnitude of approximately 3.5% of the original signal (Dash & Chilukuri, 2004). To test the performance of the proposed method in different noise environments, different levels of noises with the signal to noise ratio (SNR) values from 20 to 50 dB were considered here. The value of SNR is dened as follows:

SNR 10 log

  Ps dB Pn

16

where Ps is the power (variance) of the signal and Pn is that of the noise. The noise was added with three phase signals of the events. Event signals of the training set were divided into ve groups, including 20 events in per group. The noise was not added to rst group. The event signals of second group, third group, fourth group and fth group of the training set were added 20, 30, 40 and 50 dB of noise, respectively. After features were extracted for noisy event signals, the SVM parameters were selected and the SVM was trained. Then proposed method was tested with 20, 25, 30, 35, 40, 45 and 50 dB noise levels. In Fig. 7, PQ event signals added 30 dB noises are shown. The classication results for the noisy event signals are shown in Table 3. Table 4 gives classication results in terms of confusion matrix of event signals with 30 dB noise level. The overall accuracy of classication which reects the extrapolating capability of proposed method is 95.81% as shown Table 3. Testing results of 25 and 35 dB noise levels which are not included during training process have high accuracy. Since 20 and 25 dB are a severe noise, the classication accuracy rate degrades, but proposed algorithm can effectively classify different kinds of PQ events with other noise levels (30, 35, 40, 45 and 50 dB). Hence, classication results of proposed method are satisfactory even if noise levels are different during training and testing. It can be said that automatic classication algorithm have high generalization capability. 5.5. Discussion The greatest advantage of the proposed system based on the WT and the SVM classier has a single feature vector for a three-phase event as different from the similar studies in the literature. Hence, feature vector representing an event is similar to the ones, even if event is nonsymmetrical. To verify the performance of the pro-

H. Eris ti, Y. Demir / Expert Systems with Applications 37 (2010) 40944102 Table 3 Classication results for the noisy event signals. Class Classication accuracy (%) C1 C2 C3 C4 C5 C6 C7 Overall (%) 100 98 100 100 100 100 100 99.71 20 dB 77 67 93 100 62 98 97 84.86 25 dB 78 94 98 100 68 100 71 87.00 30 dB 96 94 100 100 88 100 100 96.98 35 dB 100 96 100 100 97 100 100 99.00 40 dB 100 98 100 100 99 100 100 99.57 45 dB 100 98 100 100 100 100 100 99.71 50 dB 100 98 100 100 100 100 100 99.71 93.87 92.87 98.88 100 89.25 99.75 96 95.81

4101

Overall (%)

Table 4 Classication results in terms of confusion matrix of event signals with 30 dB noise level. True class C1 C2 C3 C4 C5 C6 C7 C1 C2 C3 C4 C5 2 3 0 0 88 0 0 C6 0 0 0 0 0 100 0 C7 0 0 0 0 0 0 100 Accuracy (%) 96 94 100 100 88 100 100 96.86

96 2 0 0 2 94 1 0 0 0 100 0 0 0 0 100 1 11 0 0 0 0 0 0 0 0 0 0 Overall success rate (%)

Besides, this method eliminates the limitations of the existing single-phase based analysis techniques. The proposed algorithm was implemented for accurately distinguishing the seven different PQ classes based on classication regions. In this paper, the performance of proposed approach was tested using event signals under the conditions with noisy and noiseless. Results of this paper show that the algorithm has an effective structure in classication the PQ events, especially under different noise conditions. It has the great potential to improve the performance of the automatic PQ monitoring equipments with on-line classication abilities. Acknowledgments

posed method, PQ events simulated by using the ATP/EMTP were considered. In this case, event parameters which are the fault resistance, incidence angle, fault location, load, etc. are different from the each other as shown in Table 1. In total, 1400 events were simulated. 700 PQ events were used for training of the SVM classier. The proposed algorithm was tested by other 700 PQ events. Moreover, the proposed approach was analyzed under noisy environment. The results are summarized in Tables 24. The overall classication accuracies for noiseless and noisy environment are 99.71% and 95.81%, respectively. Classication results of noiseless and noisy show that proposed algorithm is robust. A wavelet with small support width is fast to compute. Besides, more levels of decomposition will increase the computational cost (He & Starzyk, 2006). Therefore, we have selected the db4 wavelet for analysis with 8-level of decomposition in the feature extraction stage. The db4 wavelet with short wavelet lter length and the 8level of decomposition provide small computational cost. Since the cross-validation process is implemented in the training process of the classication stage, it increases only the training computation cost. Thus, the approach of hardware prototyping for the realization of the proposed algorithm has great advantages because SVM and WMRA is computationally simple, accurate and exhibits a good balance of exibility, speed, size and design cycle time. As a result of this discussion, the proposed algorithm has the great potential for the future development of fully automated monitoring systems with on-line classication capabilities. 6. Conclusions The large number of PQ events occurs in the power systems. Hence, there is a need for a new technique that can be used for automated classication of PQ events. This paper introduced a new algorithm to distinguish the types of PQ events. The technique is based on the WMRA and SVM classier. Unlike most of the literature studies for automatic PQ event classication having three feature vectors for an event or based on single-phase analysis techniques, the proposed technique has single feature vector for an event. Therefore, the proposed classier needs less memory space and less computing time at both the training and testing processes.

This work is supported by Firat University Scientic Research Unit (FUBAP) (Project No. FUBAP-1605). References
Abe, S. (2005). Support vector machines for pattern classication. Berlin: Springer. Axelberg, P. G. V., Irene, Y. H. G., & Bollen, M. H. J. (2007). Support vector machine for classication of voltage disturbances. IEEE Transactions on Power Delivery, 22(3), 12971303. Bertsekas, D. P. (1999). Nonlinear programming (2nd ed.). Athena Scientic. Bollen, H. J., Gu, I. Y. H., Axelberg, P. G. V., & Styvaktakis, E. (2007). Classication of underlying causes of power quality disturbances: Deterministic versus statistical methods, EURASIP Journal on Advances in Signal Processing, article id: 79747, 17pp. Bollen, M. H. J., & Gu, I. Y. H. (2006). Signal processing of power quality disturbances. New York: IEEE Press. CanAm EMTP User Group. (1992). Alternative transient program (ATP) rule book, Portland. Chang, C. C., & Lin, C. J. (2001). LIBSVM: A library for support vector machines. Available from http://www.csie.ntu.edu.tw/~cjlin/libsvm. Chen, S., & Zhu, H. Y. (2007). Wavelet transform for processing power quality disturbances. EURASIP Journal on Advances in Signal Processing, article ID: 47695, 20pp. Cortes, C., & Vapnik, V. N. (1995). Support vector networks. Machine Learning, 20, 273297. Dash, P. K., & Chilukuri, M. V. (2004). Hybrid S-transform and Kalman ltering approach for detection and measurement of short duration disturbances in power networks. IEEE Transactions on Institutions and Measurements, 53(2), 588596. Daubechies, I. (1992). Ten lectures on wavelets. Philadelphia, USA: CBMSNSF Regional Conference Series, SIAM. El-Saadany, E. F., Abdel-Galil, T. K., & Salama, M. M. A. (2004). Power medium voltage industrial distribution system power quality assessment utilizing multi-resolution decomposition techniques. Canadian conference on Electrical and Computer Engineering, 2, 869874. Gaing, Z. L. (2004). Wavelet-based neural network for power disturbance recognition and classication. IEEE Transactions on Power Delivery, 19(4), 15601568. Gargoom, A. M., Ertugrul, N., & Soong, W. L. (2007). Investigation of effective automatic recognition systems of power-quality events. IEEE Transactions on Power Delivery, 22(4), 23192326. He, H., & Starzyk, J. A. (2006). A self-organizing learning array system for power quality classication based on wavelet transform. IEEE Transactions on Power Delivery, 21(1), 286295. Hong, Y. Y., & Wang, C. W. (2005). Switching detection/classication using discrete wavelet transform and self-organizing mapping network. IEEE Transactions on Power Delivery, 20(2), 16621668. part 2. Huang, J., Negnevitsky, M., & Nguyen, D. T. (2002). A neural-fuzzy classier for recognition of power quality disturbances. IEEE Transactions on Power Delivery, 17(2), 609616.

4102

H. Eris ti, Y. Demir / Expert Systems with Applications 37 (2010) 40944102 Styvaktakis, E., Bollen, M. H. J., & Gu, I. Y. H. (2002). Expert system for classication and analysis of power system events. IEEE Transactions on Power Delivery, 17(2), 423428. Uyar, M., Yildirim, S., & Gencoglu, M. T. (2008). An effective wavelet-based feature extraction method for classication of power quality disturbance signals. Electric Power Systems Research, 78(10), 17471755. Vapnik, V. (1998). Statistical learning theory. New York: Wiley. Zhang, N., & Kezunovic, M. (2007). A real time fault analysis tool for monitoring operation of transmission line protective relay. Electric Power Systems Research, 77(34), 361370.

Mallat, S. G. (1989). A theory for multiresolution signal decomposition: The wavelet representation. IEEE Transaction on Pattern Analysis and Machine Intelligence, 11(7), 674693. Santoso, S., Lamoree, J., Grady, W. M., Powers, E. J., & Bhatt, S. C. (2000). A scalable PQ event identication system. IEEE Transactions on Power Delivery, 15(2), 738743. Silva, K., Souza, B., & Brito, N. (2006). Fault detection and classication in transmission lines based on wavelet transform and ANN. IEEE Transactions on Power Delivery, 21(4), 20582063.