Electric Power Components and Systems, 38:182–196, 2010

Copyright © Taylor & Francis Group, LLC

ISSN: 1532-5008 print/1532-5016 online
DOI: 10.1080/15325000903273387

Detection and Classification of Power Quality

Disturbances Using Wavelet Transform and
Support Vector Machines

Z. MORAVEJ,1 A. A. ABDOOS,1 and M. PAZOKI1

1
Department of Electrical and Computer Engineering, Semnan University,
Semnan, Iran

Abstract Recognition of power quality events by analyzing voltage waveform dis-

turbances is a very important task for power system monitoring. This article presents
a novel approach for the recognition and classification of power quality disturbances
using wavelet transform and wavelet-support vector machines. The proposed method
employs wavelet transform techniques to extract the most important and significant
feature from details and approximation waves. The obtained severable feature vectors
are used for training the support vector machines to classify the power quality distur-
bances. Various transient events, such as voltage sag, swell, interruption, harmonic,
transient, sag with harmonic, swell with harmonic, and flicker, are tested. Sensitivity of
the proposed algorithm under different noise conditions is investigated in this article.
The results show that the classifier can detect and classify different power quality
signals, even under noisy conditions, correctly.

Keywords feature extraction, noise, power quality, support vector machines, wavelet
transform

1. Introduction
In recent years, power quality (PQ) has become a significant issue for both utilities and
customers. PQ issues and the resulting problems are the consequences of the increasing
use of solid-state switching devices, non-linear and power electronically switched loads,
unbalanced power systems, lighting controls, computer and data processing equipment,
as well as industrial plant rectifiers and inverters. These electronic-type loads cause quasi-
static harmonic dynamic voltage distortions, inrush, pulse-type current phenomenon with
excessive harmonics, and high distortion. A PQ problem usually involves a variation in
the electric service voltage or current, such as voltage dips and fluctuations, momentary
interruptions, harmonics, and oscillatory transients, causing failure or inoperability of the
power service equipment. Hence, to improve PQ, a fast and reliable detection of distur-
bances and sources and causes of such disturbances must be known before any appropriate
mitigating action can be taken [1, 2].
The proposed algorithm for detection of PQ disturbances are generally decomposed
into main three steps: (1) generation of PQ disturbances, (2) feature extraction and

Received 29 March 2009; accepted 29 June 2009.

Address correspondence to Z. Moravej, Department of Electrical and Computer Engineering,
Semnan University, Semnan, Semnan, 13955, Iran. E-mail: [email protected]

182
 Detection of Power Quality Disturbances Using WT and SVMs 183

(3) training of extracted feature vectors. Mathematical equations and simulation are
usually used for generation of PQ events. Signal processing techniques are widely
used in analyzing PQ events to extract important information of disturbances. Some
examples are the fast Fourier transform method [3], fractal-based method [4], S-transform
method [5], time-frequency ambiguity plane method [6], short time power and correla-
tion transform method [7], wavelet transform (WT) method [8], Hilbert transform [9],
and Chirp-Z transform (CZT) [10]. A neural network [11] with different structures is
traditionally used as a classifier, but recently the probabilistic neural network (PNN) [2,
12] and support vector machines (SVMs) [13] have been introduced as new learning
machines that are more effective. Rule-based expert system and fuzzy logic have also
been employed for the decision-making step in the process of classifying PQ disturbance
types [14, 15].
The major problems of the traditional analyzing methods are:

 the discrete Fourier transform (DFT) and CZT do not provide sufficient informa-
tion on the time, while WT leads to optimal time frequency resolution [3, 10];
 S-transform requires high complexity computation and is very sensitive to the
Gaussian window parameter [5];
 those extracted features based on WT are not as good as the features used in this
article (due to low classification accuracy) [8];
 the neural network classifier needs too many tuning parameters (such as the number
of hidden-layer neurons, learning rate, momentum), and training process is very
time consuming [11];
 more than one cycle is used as time interval for sampling [2, 5, 12]; and
 other techniques do not have high classification accuracy [4, 9].

One technique that has emerged to overcome the above-mentioned problem is using
WT, whose strength is on handling signals on short time intervals for high-frequency com-
ponents and long time intervals for low-frequency components. By means of its strength,
WT is considered suitable for analyzing signals with localized impulses and oscillations,
particularly for those commonly present in fundamental and low-order harmonics.
In 1994, the usage of WT was proposed to study power system non-stationary
harmonic distortion [16]. In 1996, this technique started to be used as a power tool
that has the ability of analysis in both time and frequency domains to detect and localize
PQ problems [17]. After these works, many studies were done on automatic classification
of PQ disturbances. Gaouda et al. [18] proposed the wavelet-multi-resolution analysis
technique to detect, localize, and classify different PQ problems. A new feature extraction
method based the standard deviation at different resolution levels was applied as inputs to
the neural network to classify the PQ disturbance type. Elmitwally et al. [19] proposed
a new neuro-fuzzy network to classify wavelet statistical features obtained from PQ
disturbances.
Wavelet packet energy entropy is utilized to obtain features for sampling PQ distur-
bance signals in [20]. He and Starzyk [21] proposed a novel approach for PQ disturbance
classification based on WT and a self-organizing learning array (SOLAR) system. The
energy value at each decomposition level using multi-resolution analysis is applied to
SOLAR. The performance of this system under different noise conditions is investigated
and it can correctly classify PQ disturbances. A new classification methodology, based on
machine inductive learning and implemented using the C4.5 algorithm decomposed with
WT of original signals, is proposed in [22]. It is shown that typical PQ disturbances are
184 Z. Moravej et al.

correctly classified. A novel classifier is performed by using a PNN for PQ application [2].
In this scheme, energy distribution features from wavelet analysis and Parseval’s theorem
are used to PQ disturbance recognition and classification.
To utilize a reliable classifier, it is essential to extract a useful feature vector that can
reduce data size as well indicating and recognizing the main characteristics of a signal.
Hence, in this article, a wavelet feature extraction technique based on WT is proposed
for automatic PQ disturbance classification. The disturbance classification schema is
performed with wavelet-SVMs (WSVMs). The WSVM realizes a feature extraction and
a classification algorithm composed of a wavelet feature extractor and an SVM classifier.
The WSVM is applied on a set of different PQ disturbances, such as voltage sag, swell,
interruption, harmonic, transient, sag with harmonic, swell with harmonic, and flicker.
Then, the effect on the classification success of a different wavelet family is investigated.
Sensitivity to noise of the proposed schema is tested under different noise conditions.
Finally, the performance comparison between the proposed method and previous literature
reports are presented for a better validation. The result shows that an WSVM could
analyze the PQ signal efficiently. Using this method, the classification accuracy percentage
of the PQ disturbances can be increased by an easier disturbance classifier based on an
SVM. The advantages of proposed method are:

 high accuracy classification is obtained,

 the new features are extracted based on WT without losing its distinguishing
characteristics [23–25],
 feature vectors are excellently severable,
 an SVM [13, 26–28] is used as a powerful classifier in comparison with other
classifiers such as the PNN and feed forward neural network,
 robust and high classification success in noisy conditions (in practical applica-
tions),
 data generation is based on parametric equations where testing and training signals
can be changed in a wide range and a controlled manner,
 sampling time interval is only one cycle.

This article is organized as follows. In Section 2, feature vector construction using

WT is presented. Section 3 briefly discusses the SVM formulation for pattern recognition
purposes. In Section 4, the proposed method is described in detail. Section 5 presents
numerical results, and the performance of the proposed algorithm is compared to other
presented methods. Finally, some conclusions are given in Section 6.

2. Feature Vector Construction Using WT

2.1. Discrete WT (DWT)

Wavelet-based techniques are powerful mathematical tools for digital signal processing
and have become more and more popular since the 1980s. They find applications in
different areas of engineering due to their ability to analyze local discontinuities of
signals. The main advantage of wavelets is that they have a varying window size, which
is wide for slow frequencies and narrow for fast, thus leading to an optimal time-frequency
resolution in all frequency ranges [23].
The DWT of a signal x is calculated by passing it through a series of filters. First,
the samples are passed through a low-pass filter with impulse response g, resulting in a
 Detection of Power Quality Disturbances Using WT and SVMs 185

convolution of the two:

1
X
yŒn D .x  g/Œn D xŒkgŒn k: (1)
kD 1

The signal is also decomposed simultaneously using a high-pass filter h. The outputs
giving the detail coefficients (from the high-pass filter) and approximation coefficients
(from the low-pass). It is important that the two filters are related to each other and they
are known as a quadrature mirror filter. However, since half the frequencies of the signal
have now been removed, half the samples can be discarded according to Nyquist’s rule.
The filter outputs are:
1
X
ylow Œn D xŒkgŒ2n k; (2)
kD 1

1
X
yhigh Œn D xŒkhŒ2n C 1 k: (3)
kD 1

This decomposition has halved the time resolution since only half of each filter output
characterizes the signal. However, each output has half the frequency band of the input, so
the frequency resolution has been doubled. For multi-level resolution, the decomposition
is repeated to further increase the frequency resolution, and the approximation coefficients
decomposed with high- and low-pass filters. This is represented as a binary tree with nodes
representing a sub-space with different time-frequency localization. The tree is known
as a filter bank [24]. In Figure 1, a voltage sag event is decomposed into five levels: a5
is the approximation level containing the fundamental frequency component, and d1 –d5
are detail levels with high frequencies. The fourth-order Daubechies (db4) wavelet filter
was used for wavelet decomposition.

Figure 1. Voltage sag waveform and five levels of wavelet decomposition.

186 Z. Moravej et al.

2.2. Feature Extraction

The critical stage for any pattern recognition problem is the extraction of discriminative
features from the raw observation data. The combination of several scalar features forms
the feature vector. For the classification of these events, the appropriate characteristics
should be extracted. These characteristics should be chosen to have low computation time
and can separate disturbances with high precision. Therefore, the similarities between
these characteristics should be few and the number of them must be low. The details and
approximation signals obtained by wavelet and multi-resolution analysis to disturbance
signals have effective feature information. Most useful features must first be extracted
from those decomposed signals to reduce the dimension of feature vectors in order to
more effectively recognize the type of PQ disturbances. By decomposition PQ events
into several levels with a different mother wavelet, it is found that the best features to be
chosen are as follows:
1. standard deviation of level 2 of detail,
2. minimum value of absolute of level n of approximation (n denotes desirable
decomposition level),
3. mean of average of absolute of all levels of detail,
4. mean of disturbances energy,
5. energy of level 3 of detail, and
6. RMS value of main signal.
These features must be chosen so that in each disturbance, there is at least one
discriminative characteristic that separates it from other types of PQ disturbances. In
Figure 2, feature 3 versus feature 4 is depicted for all selected PQ disturbances. It is
obvious that all of these disturbances are in a separate region, which can be districted
easily in two-dimensional spaces.

Figure 2. Two-dimensional map of feature 3 versus feature 4 with five levels of decomposition.
 Detection of Power Quality Disturbances Using WT and SVMs 187

3. SVMs
The SVM is a supervised learning method that generates input–output mapping functions
from a set of labeled training data. The mapping function can be either a classification
function (i.e., the category of the input data) or a regression function. For classification,
nonlinear kernel functions are often used to transform input data to a high-dimensional
feature space in which the input data become more separable compared to the original
input space. Maximum-margin hyperplanes are then created [26–28].
The SVM uses the principles of statistical learning theory to find a functional as
simple as possible to reach a generalization as good as possible for the description of a
given data set.
The advantages of SVMs over multilayer neural network classifiers are:

—maximization of generalization ability,

—no local minima, and
—robustness to outliers.

The disadvantages of SVMs explained so far are:

—extension to multi-class problems and

—selection of parameters.

The classification problem can be restricted to consideration of the two-class problem

without loss of generality [28]. Consider the problem of separating the set of training
vectors belonging to two linear separable classes, D D f.x 1 ; y 1 /; : : : ; .x l ; y l /g, x 2 <n ,
y 2 f 1; 1g with a hyperplane hw; xi C b D 0. A separating hyperplane in canonical
form must satisfy the following constraints:

y i Œhw; x i i C b  1; i D 1; : : : ; l: (4)

The optimal hyperplane is given by maximizing the margin. Hence the hyperplane
that optimally separates the data is the one that minimizes ˆ D 21 kwk2 . The solution to
the optimization problem of ˆ, under the constraints of Eq. (4) is given by the saddle
point of the Lagrange functional (Lagrangian)

l
1 X
ˆ.w; b; ˛/ D kwk2 ˛i .y i Œhw; x i i C b 1/; (5)
2
i D1

where ˛ are the Lagrange multipliers. Equation (5) can be transformed to its dual problem,
which is easier to solve. The solution to the problem is given by

l l l
 1 XX X
˛ D arg min ˛i ˛j yi yj hxi ; xj i ˛k ; (6)
˛ 2
i D1 j D1 kD1

with constraints

l
X
˛i  0; i D 1; : : : ; l; and ˛j yj D 0: (7)
j D1
188 Z. Moravej et al.

Solving Eq. (6) with constraints Eq. (7) determines the Lagrange multipliers, and the
optimal separating hyperplane is given by
l
X 1 
w D ˛i yi xi and b D hw ; xr C xs i; (8)
i D1
2

where xr and xs are any support vector (SV) from each class satisfying

˛r ; ˛s > 0; yr D 1; ys D 1:

The hard classifier is then

f .x/ D sgn.hw  ; xi C b/I (9)

only the points xi that satisfy y i Œhw; x i iCb D 1 will have non-zero Lagrange multipliers.
These points are SVs for the linearly non-separable to enable the optimal separating
hyperplane method to be generalized. The non-negative variables i  0 and penalty
function is introduced [26–28]:
X
F .i / D i ;  > 0; (10)
i

where i are measures of the misclassification errors. The constraints of Eq. (4) are
modified for the non-separable case to

y i Œhw; x i i C b  1 i ; i D 1; : : : ; l: (11)

where i  0. The generalized optimal separating hyperplane is determined by the vector

w that minimizes the functional
1 X
ˆ.w; / D kwk2 C C i (12)
2 i

subject to the constraints of Eq. (13). The solution is given by the saddle point of the
Lagrangian. Hence the solution to the problem is given by
l l l
1 XX X
˛  D arg min ˛i ˛j yi yj hxi ; xj i ˛k (13)
˛ 2 i D1 j D1
kD1

with constraints
l
X
0  ˛i  C; i D 1; : : : ; l; ˛j yj D 0: (14)
j D1

The uncertain part of the approach is that the coefficient C has to be determined.
This parameter introduces additional capacity control within the classifier. C must be
chosen to reflect the knowledge of the noise on the data.
The SVM can also be used in non-linear classification tasks with the application
of kernel functions. The data to be classified are mapped onto a high-dimensional
feature space, where the linear classification is possible. The change of optimization
problem of Eq. (13) is hxi ; xj i 7! K.xi ; xj /, where K.xi ; xj / is the kernel function
 Detection of Power Quality Disturbances Using WT and SVMs 189

performing the non-linear mapping into feature space, and the constraints (Eq. (14))
are unchanged. Solving the optimization problem and a hard classifier implementing the
optimal separating hyperplane in the feature space is given by
!
X
f .x/ D sgn ˛i yi K.xi ; x/ C b : (15)
i 2SVs

If the kernel function is a symmetric positive definite function, it satisfies Mercer’s con-
ditions. Some of the kernel functions are polynomial, Gaussian radial basis, exponential
radial basis, and splines.

4. Proposed Method

4.1. Simulation of Different PQ Disturbances Using

Parametric Equations
Data generation by parametric equations for classifier tests is advantageous in some ways.
It was possible to change training and testing signal parameters in a wide range and in a
controlled manner. The signals simulated this way were very close to the real situation.
On the other hand, different signals belonging to the same class gave the possibility to
estimate generalization ability of classifiers based on the SVM.
The input data to the SVM based on the PQ disturbance classification system was
generated based on the model in the literature. Nine classes (S1–S9) of different PQ
disturbances, named pure sine (normal), sag, swell, interruption, harmonics transients,
sag with harmonic, swell with harmonic, and flicker, were considered. Table 1 gives the
signal generation models and their control parameters. One hundred cases of each class
with different parameters were generated for training, and 100 cases were generated for
testing. Both the training and testing signals are sampled at 100 points/cycle, and the
normal frequency is 50 Hz. One power frequency cycle that contains the disturbance is
used for the extraction of the features vector. A waveform of PQ events is shown in
Figure 3.

4.2. Training Feature Vectors

Each disturbance signal is decomposed with a different wavelet filter for three to six
levels. The MATLAB (wavelet toolbox) software [29] was used to calculate the wavelet
decomposition. Based on the feature extraction mentioned above, six-dimensional feature
sets for training and testing data were constructed. The dimensions here describe different
features resulting from WT; that is, the total size of the training data and testing data set
is 6  900, where 6 is the dimension of the feature size of each case, and 900 is derived
from 100 cases per class multiplied by 9 classes.
The input of the SVM is a preprocessed signal. In this case, the PQ disturbance signal
in the time domain is transformed into the wavelet domain before applying as am input to
the SVM. This process is represented by the feature extraction stage in Figure 4. Feature
extraction is the key for pattern recognition; thus, it is the most important component of
designing the intelligent system based on pattern recognition, since even the best classifier
will perform poorly if the features are not chosen well. A feature extractor should reduce
the pattern vector (i.e., the original waveform) to a lower dimension, which contains most
of the useful information from the original vector.
 Table 1
PQ disturbance model

Class
PQ disturbance symbol Model Parameters
Pure signal S1 f .t/ D A sin.!t/ Frequency: 50 Hz,
AD1
Sag S2 f .t/ D A.1 ˛.u.t t1 / u.t t2 /// sin.!t/ 0:1  ˛  0:9,
T  t2 t1  9T
Swell S3 f .t/ D A.1 C ˛.u.t t1 / u.t t2 /// sin.!t/ 0:1  ˛  0:8,
T  t2 t1  9T
Interruption S4 f .t/ D A.1 ˛.u.t t1 / u.t t2 /// sin.!t/ 0:9  ˛  1,
T  t2 t1  9T
Harmonic S5 f .t/ D A.˛1 sin.!t/ C ˛3 sin.3!t/ C ˛5 sin.5!t/ C ˛7 sin.7!t// 0:05  ˛3  0:15,
0:05  ˛5  0:15,
0:05  ˛7  0:15,
†˛i2 D 1

190
Transient S6 f .t/ D .sin.!t/ C ˛osc exp. .t t1 /=osc // sin.!nosc .t t1 // osc D 0:008–0.04 sec,
!nosc D 100–400 Hz
Sag C harmonic S7 f .t/ D A.1 ˛.u.t t1 / u.t t2 ///.˛1 sin.!t/ 0:1  ˛  0:9,
C˛3 sin.!t/ C ˛5 sin.5!t// T  t2 t1  9T ,
0:05  ˛3  0:15,
0:05  ˛5  0:15,
0:05  ˛7  0:15,
†˛i2 D 1
Swell C harmonic S8 f .t/ D A.1 C ˛.u.t t1 / u.t t2 /// 0:1  ˛  0:8,
.˛1 sin.!t/ C ˛3 sin.!t/ C ˛5 sin.5!t// T  t2 t1  9T ,
0:05  ˛3  0:15,
0:05  ˛5  0:15,
0:05  ˛7  0:15,
†˛i2 D 1
Flicker S9 f .t/ D .1 C ˛f sin.ˇf !t// sin.!t/ ˛f D 0:1–0.2,
ˇf D 5–10 Hz
 Detection of Power Quality Disturbances Using WT and SVMs 191

Figure 3. PQ events signals.

The basic idea of the WSVM classifier is to perform waveform recognition in the
wavelet domain using the SVM. The WSVM is a pattern classification algorithm that
is tried to integrate aspects of the wavelet transformation for the feature extraction and
SVM approaches with the characteristic decision capabilities for the classification. In
this algorithm, after realizing the feature extraction (preprocessing) stage using detail
and approximation in each decomposition level obtained from WT and multi-resolution
analysis, the classification (processing) stage is implemented by using the SVM with
different kernel functions such as linear, radial basis, and sigmoid. The input to the clas-
sifier is the time domain disturbance signal, and the output is the type of the disturbance
along with its corresponding degree of belief. Therefore, it can be said that the proposed
WSVM for PQ disturbance classification is a relatively new approach.

5. Network Performance and Results

In the literature, different wavelet families have been used to decompose the distorted
signals and to extract the feature vector. For testing the classification performance of
feature vectors extracted from different wavelet families, five different wavelet families
(Daubechies [db], coiflets [coif], biorthogonal [bior], haar [haar], and symlet [sym]) are
selected. Table 2 gives the classification performance of feature vectors extracted from
db4, db12, coif5, bior3.7, bior5.5, haar, sym8, and sym12 wavelet filters, although the
192 Z. Moravej et al.

Figure 4. The structure of WSVMs for disturbance classification.

best classification performance is achieved with biorthogonal wavelet filters. A detailed

discussion about these wavelet characteristics can be found in [24].
After several experiments to obtain the best performance of the SVM, different
values of parameters and various kernel functions are tested. It is found by choosing
640,000 and 2 for C and , respectively. The radial basis function (RBF) as a kernel

Table 2
Maximum accuracy by SVM prediction and different wavelet function
at levels 3, 4, 5, and 6

Wavelet function

Level number bior3.7 bior5.5 coif5 db4 db12 haar sym8 sym12

Overall accuracy (%)

Three levels 97.33 98.22 98.22 97.33 97.78 94.22 98.00 97.78
Kernel function Linear rbf rbf rbf rbf Sigmoid rbf rbf
Four levels 97.44 97.33 96.78 97.44 98.67 95.78 98.22 97.89
Kernel function rbf rbf rbf rbf Sigmoid rbf Sigmoid rbf
Five levels 97.11 98.89 96.22 96.33 96.44 94.11 96.67 96.67
Kernel function rbf Sigmoid rbf Sigmoid rbf rbf rbf rbf
Six levels 97.89 97.67 97.44 97.11 96.78 93.50 97.11 97.44
Kernel function rbf rbf rbf rbf rbf rbf rbf rbf

Note. The best classification accuracy is distinguished by bold and shading.

 Detection of Power Quality Disturbances Using WT and SVMs 193

Table 3
Classification results of WSVM algorithm without considering noise

Accuracy

This
True article Reference Reference Reference
class S1 S2 S3 S4 S5 S6 S7 S8 S9 (%) [1] (%) [21] (%) [22] (%)

S1 100 0 0 0 0 3 0 0 1 96.15 100 100 100

S2 0 100 0 0 0 0 0 0 0 100.00 88 87 76.5
S3 0 0 100 0 0 0 0 0 0 100.00 96.5 100 97
S4 0 0 0 100 0 0 0 0 0 100.00 85.55 80.5 90
S5 0 0 0 0 98 0 0 0 0 100.00 100 100 100
S6 0 0 0 0 0 95 0 0 2 97.94 — — —
S7 0 0 0 0 0 0 100 0 0 100.00 100 97 71.5
S8 0 0 0 0 0 0 0 100 0 100.00 100 100 98
S9 0 0 0 0 2 2 0 0 97 96.04 — — —
Overall success rate 98.89 95.71 94.93 90.42

function has been selected, and the best classification accuracy of 98.89% is obtained.
As seen in Table 3, classification results are described in terms of a 9  9 confusion
matrix. The diagonal elements represent the correctly classified PQ disturbances. The off-
diagonal elements represent the misclassified PQ disturbances. The system can correctly
classify 890 of the 900 different PQ disturbance signals in the testing set, as shown
in Table 3. A 100% correct classification rate is obtained for six disturbances signals
(S1, S2, S3, S4, S7, and S8). Signals S5, S6, and S9 are classified correctly with a
96.71% mean success rate. This means that the overall success rate is found to be about
98.89%. The reason for the lower classification accuracy, especially in the S5, S6, and
S9 classes compared to those of the other classes, is the similar wave shapes of these
three classes.

5.1. Performance of WSVM under Different Noise Conditions

The signals in real electric power systems usually have noise. Additive white Gaussian
noise (AWGN) is commonly considered in the research of PQ issues [1, 22]. In order to
test the sensitivity of a WSVM under different noise conditions, different levels of noises
with the signal-to-noise ratio (SNR) values between 20 and 50 dB are added. The value
of the SNR is defined as follows:
 
Ps
SNR .dB/ D 10  log ; (16)
Pn

where Ps is the power (variance) of the signal and Pn is that of the noise. Twenty, 30, 40,
and 50 dB of noise was added to the testing vectors to see how well the proposed scheme
would work in the presence of noise. A typical SNR value of 30 dB is equivalent to a
peak noise magnitude of nearly 3.5% of the voltage signal. From Table 4, the average
classification accuracy rate is 97.52%. As noted, the tolerance to the noise declines as
the SNR becomes smaller. The WSVM has a robust anti-noise performance, and high
classification success in a noisy condition can be still attained.
194 Z. Moravej et al.

Table 4
Percentage of correct classification results of WSVM
under different SNR values

Class 20 dB 30 dB 40 dB 50 dB

S1 98 100 100 100

S2 100 100 100 100
S3 100 100 100 100
S4 100 100 100 100
S5 88 87 90 93
S6 92 94 97 100
S7 99 100 100 100
S8 99 99 99 99
S9 91 93 98 95
Overall 96.33 97.00 98.22 98.55

5.2. Performance Comparison

In order to evaluate the effectiveness and feasibility of the proposed method, a com-
parison in terms of percentage of the classification accuracy between the results of this
study and results of the classification scheme in [1, 21, 22] is made and comparatively
presented in Table 3. The shceme in [1, 21, 22] included only seven types of disturbances
(excluded are transient and flicker) but use same mathematical equations as used in
this article. As seen from Table 3, the performance of the classification process with
the proposed feature extraction method exceeds the performance of the classification
process proposed in [1, 21, 22].
In addition, a comparison for the presence of noise is made between the results of
this work and that of [1, 21]. The classification accuracies of [1, 21] are about 90–95
and 90–94%, respectively, for SNR values of 20–50 dB. As seen from Table 4, the
classification results of this study are close to those of [1]; the classification accuracy
under different noise conditions was not presented in [22]. However, the classification
accuracy in SNR values of 20 dB belonging to this study has exceeded those without
noise shown in [21]. The comparison results show that the classification system used in
this article is highly immune to the presence of noise.

6. Conclusions
In this article, a wavelet-based effective feature extraction method is proposed for the
automatic PQ disturbances classification. The disturbance classification scheme is per-
formed with the WSVM. It performs a feature extraction and a classification algorithm
composed of a wavelet feature extractor and SVM classifier. In addition, the comparison
of wavelet families (Daubechies, coiflets, biorthogonal, haar and symlet) for PQ distur-
bance classification by using the proposed method is presented. Finally, a performance
comparison with previous studies in terms of the classification accuracy is presented.
Often, many practical applications, including PQ disturbance classification, require
a large computational ability to cope with complexity or real-time limitations. Often,
traditional computers cannot quickly achieve such ability or they are too expensive.
 Detection of Power Quality Disturbances Using WT and SVMs 195

The most important advantage of the proposed method is the reduction of data
size, as well as indicating and recognizing the main characteristics of the signal with-
out losing its distinguishing characteristics. Furthermore, it can reduce memory space,
shorten preprocessing needs, and increase computation speed for the classification of
PQ disturbances. The analyses and the results presented in this article clearly reveal the
potential capability of the proposed method in classifying the distorted PQ signals. It has
great potential to improve the performance of the automatic PQ monitoring equipments
with on-line classification abilities.
Future work will include the following:
 Expanding the data set of PQ disturbance types with many different disturbances
such as spike, notch for the detection, and the classification process. Thus, the
characteristics of the different disturbances could be evaluated for the proposed
method.
 Employing the classification process of PQ disturbances by using real distorted
signals obtained from a distribution system or an arbitrary waveform generator.

References
1. Uyara, M., Yildirima, S., and Gencoglub, M. T., “An effective wavelet-based feature extraction
method for classification of power quality disturbance signals,” Elect. Power Syst. Res., Vol. 78,
pp. 1747–1755, 2008.
2. Gaing, Z. L., “Wavelet-based neural network for power disturbance recognition and classifi-
cation,” IEEE Trans. Power Delivery, Vol. 19, No. 4, pp. 1560–1568, 2004.
3. Heydt, G. T., Fjeld, P. S., Liu, C. C., Pierce D., Tu, L., and Hensley, G., “Applications of the
windowed FFT to electric power quality assessment,” IEEE Trans. Power Delivery, Vol. 14,
No. 4, pp. 1411–1416, 1999.
4. Huang, S. J., and Hsieh C. T., “Feasibility of fractal-based methods for visualization of power
system disturbances,” Int. J. Elect. Power Energy Syst., Vol. 23, No. 1, pp. 31–36, 2001.
5. Nguyen, T., and Liao, Y., “Power quality disturbance classification utilizing S-transform and
binary feature matrix method,” Elect. Power Syst. Res., Vol. 79, No. 4, pp. 569–575, 2009.
6. Wang, M., Ochenkowski, P., and Mamishev, A., “A classification of power quality disturbances
using time-frequency ambiguity plane and neural networks,” IEEE Power Eng. Soc. Summer
Mtg., Vol. 2, pp. 1246–1251, 2001.
7. Wen, J., and Liu, P., “A method for detection and classification of power quality disturbances,”
Automat. Elect. Power Syst. China, Vol. 26, No. 1, pp. 42–44, 2002.
8. Santoso, S., Powers E. J., Grady, W. M., and Hofmann, P., “Power quality assessment via
wavelet transform analysis,” IEEE Trans. Power Delivery, Vol. 11, No. 2, pp. 924–930, 1996.
9. David, G. L., “Comments on Hilbert transform based signal analysis,” BYU Microwave Remote
Sensing (MERS) Laboratory Technical Report No. MERS 04-001, Brigham Young University,
Provo, UT, 2004.
10. Aiello, M., Cataliotti, A., and Nuccio, S., “A Chirp-Z transform-based synchronizer for power
system measurements,” IEEE Trans. Instrument. Meas., Vol. 54, No. 3, pp. 1025–1032, 2005.
11. Kaewarsa, S., Attakitmongcol, K., and Kulworawanichpong, T., “Recognition of power quality
events by using multi wavelet-based neural networks,” Elect. Power Energy Syst., Vol. 30,
pp. 254–260, 2008.
12. Mishra, S., Bhende, C. N., and Panigrahi, B. K., “Detection and classification of power quality
disturbances using S-transform and probabilistic neural network,” IEEE Trans. Power Delivery,
Vol. 23, No. 1, pp. 280–287, January 2008.
13. Cerqueira, A. S., Ferreira, D. D., Ribeiro, M. V., and Duque, C. A., “Power quality events
recognition using a SVM-based method,” Elect. Power Syst. Res., Vol. 78, pp. 1546–1552,
2008.
196 Z. Moravej et al.

14. Liao, Y., and Lee, J. B., “A fuzzy-expert system for classifying power quality disturbances,”
Elect. Power Energy Syst., Vol. 26, pp. 199–205, 2004.
15. Andami, H., and Jalilian, A., “Voltage notch detection using fuzzy expert system,” Proceeding
of IEEE International Conference, pp. 479–482, Montreal, Canada, 4–7 May 2003.
16. Ribeiro, P., “Wavelet transform: An advanced tool for analyzing non-stationary harmonic
distortion in power systems,” Proc. IEEE Int. Conf. Harmon. Power Syst., Vol. 6, pp. 21–
23, Bologna, 1994.
17. Santoso, S., Powers E. J., Grady, W. M., and Hofmann P., “Power quality assessment via
wavelet transform analysis,” IEEE Trans. Power Delivery, Vol. 11, No. 2, pp. 924–930, 1996.
18. Gaouda, A. M., Salama, M. M. A., Sultan, M. R., and Chikhani, A. Y., “Power quality
detection and classification using wavelet-multiresolution signal decomposition,” IEEE Trans.
Power Delivery, Vol. 14, No. 4, pp. 1469–1476, 1999.
19. Elmitwally, A., Farghal, S., Kandil, M., Abdelkader, S., and Elkateb, M., “Proposed wavelet-
neurofuzzy combined system for power quality violations detection and diagnosis, IEE Proc.
Generat. Transm. Distrib., Vol. 148, No. 1, pp. 15–20, 2001.
20. Hu, G. S., Zhu, F. F., and Ren, Z., “Power quality disturbance identification using wavelet
packet energy entropy and weighted support vector machines,” Expert Syst. Appl., Vol. 35,
pp. 143–149, 2008.
21. He, H., and Starzyk, J. A., “A self-organizing learning array system for power quality classifi-
cation based on wavelet transform,” IEEE Trans. Power Delivery, Vol. 21, No. 1, pp. 286–295,
2006.
22. Abdel-Galil, T. K., Kamel, M., Youssef, A. M., El-Saadany, E. F., and Salama, M. M. A.,
“Power quality disturbance classification using the inductive inference approach,” IEEE Trans.
Power Delivery, Vol. 19, No. 4, pp. 1812–1818, 2004.
23. Daubechies, I., Ten Lectures on Wavelets, Philadelphia, PA: Society for Industrial and Applied
Mathematics, 1992.
24. Strang, G., and Nguyen T., Wavelets and Filter Banks, Wellesley, MA: Cambridge Press, 1996.
25. Heydt, G. T., and Galli, A. W., “Transient power quality problems analyzed using wavelets,”
IEEE Trans. Power Delivery, Vol. 12, No. 2, pp. 908–991, 1997.
26. Wang, L., Support Vector Machines: Theory and Applications, Nanyang Technological Uni-
versity, School of Electrical & Electronic Engineering, Singapore: Springer, 2005.
27. Steinwart, I., and Christmann, A., Support Vector Machines, New York: Springer, 2008.
28. Christopher, J. C., A Tutorial on Support Vector Machines for Pattern Recognition, Boston:
Kluwer Academic Publishers, Bell Laboratories, Lucent Technologies, manufactured in The
Netherlands, 1998.
29. MATLAB 7.4 Version Wavelet Toolbox, Math Works Company, Natick, MA.