Accepted Manuscript

A new optimal feature selection algorithm for classification of power quality

disturbances using discrete wavelet transform and probabilistic neural network

Suhail Khokhar, Abdullah Asuhaimi Mohd Zin, Aslam P. Memon, Ahmad

Safawi Mokhtar

PII: S0263-2241(16)30561-9
DOI: http://dx.doi.org/10.1016/j.measurement.2016.10.013
Reference: MEASUR 4375

To appear in: Measurement

Received Date: 19 November 2015

Revised Date: 19 September 2016
Accepted Date: 6 October 2016

Please cite this article as: S. Khokhar, A. Asuhaimi Mohd Zin, A.P. Memon, A. Safawi Mokhtar, A new optimal
feature selection algorithm for classification of power quality disturbances using discrete wavelet transform and
probabilistic neural network, Measurement (2016), doi: http://dx.doi.org/10.1016/j.measurement.2016.10.013

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
 A new optimal feature selection algorithm for classification of
power quality disturbances using discrete wavelet transform and
probabilistic neural network
Suhail Khokhar,a,b,*, Abdullah Asuhaimi Mohd Zina, Aslam P. Memonb, Ahmad Safawi Mokhtara
a
Faculty of Electrical Engineering, Universiti Teknologi Malaysia, Malaysia
b
Department of Electrical Engineering, QUEST Nawabshah Pakistan
*Corresponding Author: email: suhail@quest.edu.pk

Abstract: Automatic classification of Power Quality Disturbances (PQDs) is a challenging concern

for both the utility and industry. In this paper, a novel technique of automatic classification of
single and hybrid PQDs is proposed. The proposed algorithm consists of the Discrete Wavelet
Transform (DWT) and Probabilistic Neural Network based Artificial Bee Colony (PNN-ABC)
optimal feature selection of PQDs. DWT with Multi-Resolution Analysis (MRA) is used for the
feature extraction of the disturbances. The PNN classifier is used as an effective classifier for the
classification of the PQDs. However, the two critical concerns such as the selection of the optimal
features and the spread constant value might affect the performance of the classifier. Hence, these
two issues are addressed using a novel technique PNN-ABC based optimal feature selection and
parameter optimization for improving the performance of the classification system. The ABC
algorithm is used to select optimal features from a large feature set and the optimal value of the
PNN spread constant. The optimal feature selection method retains the useful features and discards
the redundant features. The performance of the proposed algorithm is evaluated by
PSCAD/EMTDC simulation of a typical 11 kV underground distribution system of Malaysia. The
noise-riding PQDs have also been analyzed to validate the sensitivity of the proposed algorithm.
The simulation results show that the new PNN-ABC based optimal feature selection algorithm is
proficient and accurate in classifying the PQDs.

Keywords: Power quality disturbances, feature extraction, optimal feature selection, discrete wavelet
transform, artificial bee colony, probabilistic neural network.

1. Introduction

The term Power Quality (PQ) is attributed to the quality of voltage and current waveforms,
concerned to detect the variation in ideal supply voltage and load current waveforms [1]. The power
quality disturbances are defined as the sudden variation or distortion in voltage and/or current amplitude,
phase angle and frequency. These disturbances are produced due to the application of unbalanced and
non-linear loads, switching devices, line faults, industrial grade converters, computers and data
processing equipment, usually applied in both the industrial and domestic appliances [2]. The growing
interest in PQ can further be resulted due to the complicated power distribution networks with the

1
enhancement in grid connection for the renewable energy sources such as photovoltaic, solar and wind
energy systems [3]. The short circuit faults such as Single Line-to-Ground (SLG), Line-to-Line (LL),
Double Line-to-Ground (LLG) and three-phase (LLL) faults occurring in a power system may lead to
variations in magnitude of voltage/current waveforms that create sag, swell and momentary interruption
disturbances [4]. The sag and interruption disturbances occur in faulty phases and subsequently a swell
may happen in non-faulty phase. The transients and spikes PQDs in power networks are usually caused
by lightning as well as a sudden switching off large loads, capacitor banks for power factor improvement
and transformer energizing [5]. The voltage flickers or fluctuations are commonly created due to the arc
furnaces. The application of power electronic converters and nonlinear loads may create the steady state
waveform distortion such as harmonics, noise and notch. In order to maintain the PQ, their sources should
be identified and promptly prohibited before taking a suitable mitigating action. Both the electric utilities
and consumers must certainly recognize the PQ problems in order to identify the origins of these
disturbances. Therefore, continuous measurement of PQDs has recently become a pressing concern in
power transmission and distribution networks to circumvent electrical equipment damage and to identify
the sources of these disturbances [6].

Traditionally PQDs have been detected by visual inspection which is the difficult task for PQ
engineers. A huge amount of data gathered to monitor the PQ waveforms is inflexible to handle.
Therefore, it is attractive to introduce automatic device for measuring the PQDs. Recently, signal-
processing techniques along with artificial intelligent techniques contribute a significant role for the
automatic classification of PQD waveforms [7]. Many researchers have focused on the application of
well-known signal processing techniques for feature extraction from the voltage and current data, and
incorporate an artificial intelligence based classifier to complete the classification process.

The well-known signal-processing techniques such as Fourier Transform (FT), Wavelet Transform
(WT) [8-12], S-transform [13, 14] and Hilbert-Huang transform [15], are usually applied for the detection
and feature extraction of the disturbance waveforms. The FT technique is usually used for the spectral and
harmonics analysis [16]. However, the FT is only suitable for stationary PQDs and provides erroneous
results for time-varying transient disturbances. The short time Fourier transform (STFT) [17] has
overcome the deficiencies of FT by employing a time-frequency window to localize transient in a signal.
Though the fixed resolution of time-frequency window is the limitation of STFT, the modified version of
the STFT known as S-transform has been proposed for the recognition of PQDs. The ST uses moving and
scalable localizing Gaussian window. However, the detection capability of ST also degrades in analysing
some non-stationary PQ transients disturbances [18].
Recently, authors [19] proposed HHT-PNN based classification algorithm for detecting and identifying the
single and multiple PQD signals. Nevertheless, the HHT is ineffective to construct frequency spectrum, which
results in loss of frequency components containing low energy contents. In [20] authors proposed a hybrid approach
of Gabor Transform (GT) and feedforward neural network for the identification of the arcing faults. The GT has
been utilized as an advanced signal-processing technique for optimal feature extraction. The GT is useful for the
signal analysis, it does not provide closed form solution in order to directly obtain the Gabor coefficients. These
coefficients are calculated using another efficient method, which increases the calculation burden.

The Discrete Wavelet Transform (DWT) has been proposed by researchers to overcome the fixed
resolution problem of STFT for analysing the PQDs. In addition, the DWT has been found the most

2
suitable for the automatic detection and feature extraction of the PQDs particularly transients
disturbances. In recent years, due to the transient detection capabilities, the DWT-based technique has
been applied for classification of transmission line faults [21], transformer differential protection [22],
fault location in underground distribution system [23] and islanding detection in distributed generation
[24].

Despite the fact that DWT is more suitable for the detection and feature extraction of PQDs but
DWT alone cannot categorize the various types of PQDs [25]. The DWT can only detect the disturbances
in a normal signal but it cannot accomplish the task of classification without using a suitable classifier.
For the effective classification of disturbances present in power signals, many researchers have proposed
numerous methods based on the combination of DWT and artificial neural network [12, 26, 27].

The neural network provides better classification accuracies than other classifiers, but suffers from
trial and error approach of choosing number of hidden layers, training methods and so on [28]. Among
the different architectures of neural networks existing in literature, the PNN classifier is found the most
accepted one and has been successfully applied to accurately classify the various types of PQDs [27]. For
the PNN classifier, the number of neurons in the hidden layer is equal to the number of training samples.
The proper adjustment of spread constant leads more classification accuracy. Therefore, in this paper, the
optimal value of the PNN spread constant has been evaluated.

Even though the performance of the PNN classifier is likely significant for obtaining maximum
classification accuracy in recognition of PQDs, it is also necessary to acquire efficient features that
represent PQD waveforms. Therefore, in recent studies, optimal feature selection algorithms have been
proposed to achieve effective features that represent PQDs [29]. Few researchers [5, 30-32] have focussed
on optimal feature selection for PQ monitoring, and analysed the recognition rate when they were used
together. The feature selection methods used for obtaining effective features for improving classification
accuracy consist of genetic algorithm (GA) based feature selection [14, 31], PSO based feature selection
[32], k-means based Apriori algorithm [5]. The advantages of implementing optimal feature selection
technique for the classification of PQDs are as follows.

1. The most important features are kept and irrelevant features are discarded, therefore a common and
simple classifier can be used.
2. Optimal feature selection might also be employed to get the maximum classification accuracy of a
classifier.
3. The computational complexity is reduced to some extent.

In this study, the optimal features are obtained using a new PNN-ABC based optimal feature
selection algorithm. The ABC algorithm is one of the most recently introduced swarm intelligence based
optimization techniques [33]. The inspiration of its solution quality and the fast convergence made it
suitable for multi-objective optimization problems. It has the advantages of good memory, multi-
characters, local search and an improved solution mechanism. It offers a balance between complex and
performance in order to obtain the high quality solution [34].

Moreover, recent studies proposed in literature regarding the classification of multiple PQDs are
based on the occurrence of only two classes, for instance, sag with harmonics and swell with harmonics
[2, 29, 35] and only few authors have considered a large number of multiple PQDs [36, 37]. These studies
accomplished promising results; however they are limited to only two types of multiple PQDs. In this

3
paper, six types of multiple PQDs are considered for classification.

The remaining parts of the paper are structured as follows. In section 2, PQDs with parametric
equations are described in detail. The brief introduction and mathematical background of WT and feature
extraction using DWT are discussed in section 3. In section 4, the mechanism of the probabilistic neural
network is described. The optimal feature selection method based on ABC algorithm is described in
section 5. The proposed methodology based on PNN-ABC classifier is discussed in section 6. The results
and discussion method are thoroughly explained in section 7. Finally, the conclusion is discussed in
section 8.

2. Power Quality Disturbances

The classification performance of the proposed algorithm has been assessed using parametric
equations of the sixteen (16) types of PQD signals including pure sine wave. The PQDs consist of 10
single types namely pure sine waveform, sag, swell, interruption, harmonics, impulsive transients,
oscillatory transients, flicker, notch and spikes. The 6 multiple types of PQDs consist of sag with
harmonics, swell with harmonics, interruption with harmonics, flicker with harmonics, flicker with sag
and flicker with swell. The parametric equations with the parameter variations according to IEEE-1159
standard [6] are given in Table 1. The typical PQ waveforms simulated using parametric equations are
shown in Fig. 1.

Table 1 Mathematical Models of single and Multiple Power Quality Disturbances [32]

Label PQ Disturbances Mathematical Equations Parameters

C1 Normal

C2 Sag

C3 Swell

C4 Interruption

C5 Harmonics

C6 Sag with Harmonics

C7 Swell with Harmonics

C8 Interruption with Harmonics

C9 Flicker

C10 Oscillatory Transient ;

C11 Impulsive Transient

;
C12 Periodic Notch ;
;

4
C13 Spike
;

C14 Flicker with harmonics

C15 Flicker with sag

;

C16 Flicker with swell

;

Fig. 1 Power Quality Disturbances waveforms

The waveforms with ten cycles at a sampling frequency of 10 kHz are created for a total of 2000
sampling points. In all the parametric equations of PQDs, the parameter A is a constant value (equal to 1
per unit) representing the amplitude of the waveform. The parameter indicates the intensity of the sag,
swell and interruption having variations according to the type of disturbances. The step function
provides the time duration of disturbance in a pure waveform. Harmonics disturbances represent the 3rd,
5th and 7th order harmonic components of different per unit values varying from 0.05 to 0.15 per unit of

5
the fundamental magnitude in a variety of feasible combinations. Likewise, sag, swell and interruption
disturbances combined with harmonics disturbances provide the multiple PQDs, for example, sag with
harmonics, swell with harmonics and interruption with harmonics. In the flicker disturbance, indicates
the flicker frequency varying from 5 Hz to 20 Hz and is the flicker magnitude varying from 0.1 to 0.2
per unit. For oscillatory transient, a transient frequency range from 300 to 900 Hz and transient time
constant range from 0.008 to 0.04 seconds. The indicates the impulse magnitude for impulsive
transient varying from 0 to 0.414 per unit. The parameter indicates the notch and spike magnitude
varying from 0.1 to 0.4 per unit and their width varying from 0.01 cycle to 0.05 cycle.

The PQDs created by parametric equations are beneficial in some aspects. The disturbances are
very similar to the actual situations. This gives the ability to provide the parameters of the training and
testing datasets in a wide variations and controlled manner. In addition, the signals associated to the same
class can offer the prospect to approximate the generalization capability of the classifiers.

3. Wavelet transform

The Wavelet Transform (WT) is utilized for the analysis of steady-state and non-stationary signals
in assorted fields due to the fact that it can analyse the local discontinuities of the signal. The application
of WT in PQ studies is found more suitable because most of the PQDs in a power system are non-
stationary transients. Mathematically, continuous WT (CWT) for a continuous signal with respect to
the wavelet function is given by

 (1)

In Eq. 1 the constants and stand for the scale and translation parameters, respectively. The scale
parameter provides the oscillatory frequency and the length of the wavelet, while the translation
parameter deposits its shifting position. Hence, the output is a series of wavelet coefficients at each
scale which represents the complete transient signal. In practical applications of CWT there will be
redundant information which is inappropriate for computer analysis. Therefore, DWT as in Eq. 2 is found
the more suitable for the analysis of the PQDs.

(2)

The scaling and translation parameters replaced by functions of and integers, i.e., and
respectively, whereas is the sequence of discrete points of the continuous time signal
In WT applications, selection of a proper mother wavelet plays an important role in analysis
depending on the type of data used. The mother wavelets proposed for the analysis of PQ studies by
various authors are Haar, Morlet, Mexican, Daubechies and Bi-orthogonal. In this paper, Daubechies
mother wavelet at scale 4 is used for the feature extraction of the PQD waveforms.

6
3.1 Multi-resolution analysis

Mallat [38], has introduced the theory of Multi-Resolution Analysis (MRA) algorithm that is used
as an essential part of decomposition and reconstruction of the signal at various resolution levels. The
MRA has been applied for decomposing PQD waveforms, because of its advantages such as easier
implementation and less memory usage. The main objective of MRA is to represent the signals at
different resolution levels. Scaling function and orthogonal wavelet function are used as
the building blocks for decomposing and reconstructing the signals at different resolution levels. Assume
that a time domain signal is passed through two filters in each level which are Low-Pass (LP) and
High-Pass (HP) filters. The HP filter provides the high frequency component known as detail coefficient
and the LP filter provides the low frequency component of approximation coefficient of the
original time-domain signal. Frequency band of both LP and HP filters is same and after each
decomposition cycle, the sampling frequency is divided into two [39]. Afterwards the output of the LP
filter is decomposed repeatedly and two components of the next stage and are produced
as shown in Fig. 2. This process is repeated until the desired decomposition level.

Fig. 2. Multi-resolution analysis

The decomposition of the signal in terms of the approximation and details can be represented
by scaling and wavelet , respectively as in Eqs. (3) and (4).

(3)

(4)

The LP and HP filters are related to the scaling function and the wavelet function, respectively. The
decomposition procedure starts with passing a signal through these filters. The decomposition of a

7
discrete signal with the WT based MRA produce the high frequency and low frequency
coefficients and at instant with scale . In general, the input signal can be expressed
by the sum of all components as follows.

(5)

where represents level of the wavelet decomposition. Therefore the size of the feature
vector for a signal is of length (Eq. 6).

(6)

Fig. 3 shows the plot of approximation level 4 and detail levels 1 to 4 DWT coefficients for the
pure sine wave, sag, harmonics and notch disturbances waveforms. The decomposed waveforms specify
the PQDs individually from which the disturbance duration and type can be easily identified. The pure
sine wave is a smooth waveform, therefore any variation is not indicated.

(a) (b)

(c) (d)

8
 Fig. 3. Plot of DWT coefficients for (a) Normal (b) Sag (c) Harmonics and (d) Notch

3.2 Feature extraction using Statistical Parameters

The statistical parameters have been obtained from the literature [5, 40] that are used for the feature
extraction. The eight statistical features such as energy (E), entropy (Ent), standard deviation ), mean
value ( ), kurtosis (KT), skewness (SK), root-mean square (RMS) and range (RG) value of the detail (D)
and approximation (A) coefficients are calculated using the mathematical equations given in Table 2.

Table 2 Statistical Parameters

Statistical Parameters Equations

Energy

Entropy

Standard
Deviation

Mean

Kurtosis

Skewness

RMS
Value

Range

where i = 1, 2 , …l, represents the number of wavelet decomposition at level l. Here Nis the
number of coefficients in each decomposed data.

The expressions and in the Kurtosis and Skewness calculations

indicate the expected values of the quantity, which is also known as kth moment about mean.

In the proposed featue selection method, the PQD waveforms are decomposed upto eight levels
which provide eight detail coefficietns and one approximation coefficient . The total
features of each detail and approximation coefficients obtained are 72 out of which the best 9 features
have been selected which provide the best classification accuracy. The statistical feature vectors are given
by

(7)

9
 where F1, F2, F3, …, F8 represent the feature vectors of energy, entropy, mean, standard deviation,
kurtosis, skewness, RMS values and range values, respectively of the detail and approximation
coefficients of the DWT. The entire data of features for all sixteen types of PQDs is a large feature set
which can affect the performance of the classifier. Therefore, the data must be normalized [between 0 and
1]. Before the data presented as input to the classifiers, the feature vectors obtained from MRA are
normalized between 0 and 1 using the min-max normalization method.

(8)

Where is the normalized data, and are the minimum and maximum values in the
feature vector . After normalization of the data, the overall feature set is given as

(9)

4. Classification using probabilistic neural networks

An artificial neural network presents an interconnected group of simple processing-elements

known as artificial neurons which are trained to perform a particular task by adapting the values of
weights and biases connections between the neurons of different layers. The interconnection pattern
between different layers of neurons, learning process and the activation function are the main
characteristics of a typical neural network. A pair of input and target is required to train a neural network.
The number of neurons in the input layer is equal to the number of extracted features of each signal. The
number of neurons in the output layer is equal to the target vector which is equivalent to the number of
PQDs to be classified. The weights are adapted based on a comparison of the output and the target
vectors, until the network output matches the target.

The PNN classifier is a well-known Bayesian classifier technique usually used in pattern
recognition systems [19]. The PNN model is one of the supervised learning networks which is
implemented using the probabilistic model such as Bayesian classifier. If PNN is given enough training
data, it is guaranteed to converge to a Bayesian classifier. The initial weights of the network are adjusted
automatically. The learning processes and the recalling processes are not associated to each other. The
training speed of the PNN model is very high which makes it suitable for real-time fault detection and
signal classification systems. The PNN model consists of radial basis layer between the input and hidden
layers and the competitive layer between the hidden layer and output layer as shown in Fig. 4. For the
application of signal classification, the training datasets are classified according to their distribution
values of the basic principle called probability density function . A general is given as follows:

10
 (10)

The output vector of the hidden layer in the PNN can be obtained by adjusting Eq. (10) as follows:

(11)

(12)

(13)

where and are the number of input, output and hidden layers respectively. is the number of
training samples. N is the number of classification patterns. is the spread constant. The vector
represents the input vector. is the Euclidean distance between the vectors and , i.e.,
; is the connection weight between the input layer and the hidden
layer ; and represents the connection weight relating the hidden layer and the output layer
[41].

Fig. 4 Architecture of probabilistic neural network

5. Feature selection using artificial bee colony

Generally, a feature selection algorithm consists of four factors: a subset feature generation, an
evaluation function, stopping criterion, and validation process. The main issue for the feature selection
process is how to search the optimal set of feature subsets [33]. The ABC algorithm evaluates this overall
process of feature selection. Being inspired by the intelligent foraging behaviour of honey bee swarms,
Karaboga [42] introduced the ABC algorithm which belongs to a swarm intelligence based optimization
techniques. It is applied for clustering, classification and optimization of datasets. The colony of artificial
bees contains three groups. First group is an employed bees , second group
is an onlooker bees and the third group is a scout bees
. The number of the employed bees and onlooker bees is the same which is
equal to the number of the solutions. The employed bees explore and utilize food source positions around
the hive and send information about food source positions to the hive. In consideration of this
information, the onlooker bees investigate the best food source locations explored by the employed bees
in order to find out a food source with better quality. A "greedy selection" technique is used for choosing

11
the best food source position. In the case, when both the employed and onlooker bees could not be able to
improve the food source position in a predetermined number of trials called limit, the employed bee
assigned to this food source is changed into a scout bee which carries out a random search to discover a
new food source. The limit is a control parameter which controls the occurrence of scout bees. Every
cycle of the bees search includes three phases. Firstly, modification of the position of the employed and
onlooker bees near the food sources. Second, deciding their nectar amount and specifying scout bees.
Third, change the position of the scout bees randomly towards the available food sources. The location of
each food source provides a possible solution to the problem to be optimized. The nectar amount of a
food source corresponds to the quality (fitness) of the associated solution by that location of food source.

Fig. 5 Flowchart of PNN-ABC method

12
6. Proposed PNN-ABC algorithm for classification

This section presents an approach of optimal feature selection for the classification of the PQDs.
The steps of the proposed PNN-ABC based feature selection algorithm are well explained in Fig. 5.

6.1 Initialization

In the ABC algorithm, a randomly distributed initial population of solutions (food source
positions) is generated, where indicates the number of solutions. Each solution in
the food source is a D-dimensional vector. Here, D is the number of optimal features to be analyzed that
is input size for each dataset. Once the initial food is generated, the food source positions (solutions) is
subjected to repeated cycles, of the search processes for
the bees groups namely the employed bees, onlooker bees and scout bees. The initializing pre-defined
parameters of the ABC algorithm are given Table 3. The succeeding process estimates the food source by
PNN classifier using an objective function or fitness function. The formulation of fitness function is a
critical point, which decides what ABC should optimize.

Table 3 ABC Control Parameter setting

Parameter Value
Colony Size (CS) 20
Food Source (FS) CS/2
Maximum Cycle
20
Number
Feature Dimension (D) 9
Lower bound 1, 9,17,25,33,41,49,57,65
Upper bound 8,16,24,32,40,48,56,64,72
Run time 10
Limit (L)

6.2 Evaluation of features and food sources

Let as the fitness function value related with the food position. The selection of the fitness
function is a significant issue because it is on the basis that the ABC evaluates the quality of each food
source to the proposed PNN-ABC classification system. Spread constant of PNN alongwith optimal
features for training has also been optimized using the proposed method. Therefore, the food source
consists of two parts, spread constant and the optimal features mask as shown in Fig. 6. The classification
accuracy and the optimal features are the two criteria to design a fitness function. For the fitness
definition, the PNN classification accuracy ( ) denoting the percentage of correctly classified
patterns [25] can be evaluated by Eq. (14).

(14)

Hence, for the food sources with high classification accuracy and an optimal feature subset

13
provides higher fitness value. The fitness fuction for the PNN classifier is defined as in Eq. (15)

(15)

The nectar amount or the fitness function which corresponds to the quality of the related solution
can be calculated by

(16)

In this algorithm, the total number of bees in the colony are equally divided into employed bees
and onlooker bees. Therefore, the number of solutions is equal to the number of employed bees or
onlooker bees.

Fig. 6. Particle Representation

6.3 Employed bees phase

In this phase of ABC algorithm, each employed bee creates a modification on the position
(solution) in its memory based on the neighbour food source and evaluates the nectar amount (fitness
value) of the new source (new solution). Each employed bee is allocated for each feature and assesses the
fitness value of each feature using Eq. (16). After the new food source generation and evaluation of the
fitness, a greedy selection is applied between the old and new food positions. If the fitness function
obtained by the new food source is more than that of the previous one, the bees remember the new food
source position and forget the previous one. Otherwise, they memorize the old food source position. In
this phase, each food source is improved based on the process of keeping informed feasible solution.

6.4 Onlooker bees phase

Once each employed bee completes the search process; they share the nectar amount information
of the food sources and their positions with the onlooker bees. As in the case of the employed bee, it
produces a modification on the position in its memory and ensures the nectar amount of the candidate
source. If the nectar amount is higher than that of the previous one, the bee memorizes the new position
and forgets the old one. An onlooker bee estimates the information of nectar amount taken from all
employed bees and selects food source with a probability related to its nectar amount. The probability
value ( ) is calculated in Eq. (17).

(17)

14
 where is the fitness value of the solution and is the number of food sources equal to the
number of employed bees and the fitness of the solution given in Eq. (16). Similar to the employed
bees, the onlooker bees change the position from memory and analyse the nectar quantity of the candidate
food source based on the wining probability value as calculated in Eq. (17). The probability value is
based on the roulette wheel selection scheme in which every portion is comparative in size to the fitness
value is employed. If nectar amount of the onlooker bees is more than the previous one, the new position
is incorporated into memory and the old one is forgotten. A novel food position can be calculated by
using the Eq. (18)

(18)

Here and j are arbitrarily chosen indices. Although k is a random

value, it should be different from . The value is a random number between −1 and 1 which is used to
control the production of neighbour food sources around to represent the comparison of the two food
sources to a bee.

6.5 Scout bees phase

Despite the fact that onlooker and employed bees perform exploitation in the search area, scout
bees control the discovery process and substitute the consumed nectar food source with a new food
source. The food source is accepted as abandoned, if the fitness value is not improved by predefined
iterations. The predefined iterations or maximum cycle number is defined as the “limit” for rejection.
Assume that be an abandoned source and j , the scout bee searches for a new food source
to substitute according to Eq. (19):

(19)

The pseudo-code of the ABC algorithm is as follows [33]:

1. Initialization of the food source positions

2. Evaluate the nectar amount (fitness)
3. Set Cycle=1
4. repeat
5. i) Employed bees phase
6. ii) Onlooker bees phase
7. iii) Scouts bees phase
8. Save the best food source discovered so far
9. Cycle=Cycle+1
10. until (Cycle=maximum cycle number)

7. Results and discussion

The proposed algorithm has been designed to perform the following three main experiments. The
first experiment was intended for assessing the effectiveness of the PNN approach to identify the PQD

15
waveforms without optimal feature selection approach. The performance of the PNN approach was
evaluated with the energy, entropy features, energy and entropy features, and all eight features together.
The second experiment was performed to analyse the generalization capability of the PNN classifier with
the feature selection based on ABC optimization technique. The third experimental performed as an
objective to evaluate the capability of the proposed PNN-ABC classification system to boost further the
accuracy of the PNN classifier using automatic feature selection and spread constant selection
optimization process. The described simulation based experiments were implemented using the
MATLAB environment.

7.1 Dataset generation

A dataset of single and hybrid PQDs has been obtained from the most widely used synthetic
parametric equations for the classification of the PQDs. The most common single stage PQDs include
sag, swell, interruption, harmonics, impulsive transient, oscillatory transients, flicker, spike and notch.
The combination of two or more PQDs occurring simultaneously is known as hybrid PQD. The
parametric equations of the PQDs based on IEEE standard 1159 [6] were simulated in MATLAB
(R2014a). Considering the parameter variations of the different types of the PQDs, overall 3200
waveforms have been randomly created for the 16 types with 200 samples of each disturbance. The
simulated PQDs obtained from the parametric models mostly bear a resemblance to real-time data.

7.2 PNN-ABC based feature selection

The best optimal features of all the wavelet based MRA coefficients selected by ABC algorithm are
reported in Table 4. The statistical parameters for each wavelet coefficient are
optimally selected by the combination of the proposed PNN-ABC algorithm. The best optimal features
based on classification accuracy and the spread constant for PNN classifier are reported in Table 5. The
best optimal features RMS
and the best spread constant are obtained which give the highest classification
accuracy of 99.875%. In order to compare the performance of the only PNN classifier, it is evaluated with
the data set of energy and entropy features as shown in Table 6. The recognition rate for only energy and
entropy (9 features for each) as a feature set for all coefficients is also obtained. It is observed that either
energy or entropy as a feature data set for training the PNN classifier offers almost same percentage of
average accuracy. When both energy and entropy (18 features) are chosen as features, the average
accuracy is enhanced but the feature dimension is enlarged to 18 features. If all features selected having
the feature set dimension 72, the overall accuracy further increases, but the computation time is enhanced
due to the large feature dimension set and the size of the feature set is increased.

16
 Table 4 Features selected by ABC

Statistical Features
Wavelet

Coefficients Standard RMS

Energy Entropy Mean Kurtosis Skewness Range
Deviation Value

1 2 3 4 5 6 7 8

9 10 11 12 13 14 15 16

17 18 19 20 21 22 23 24

25 26 27 28 29 30 31 32

33 34 35 36 37 38 39 40

41 42 43 44 45 46 47 48

49 50 51 52 53 54 55 56

57 58 59 60 61 62 63 64

65 66 67 68 69 70 71 72

Table 5 PNN-ABC based optimal feature selection and spread constant

Classification Spread
Iter Optimally selected features
Accuracy Constant

1 92.59375 0.00054124 4 13 24 25 37 41 54 61 69

2 99.875 0.0000682 1 13 23 30 38 41 55 58 69

3 85.34375 0.001569672 4 9 19 29 34 41 49 57 69

4 95.59375 0.000975222 3 16 21 29 36 41 49 57 69

5 78.8125 0.000461903 3 14 20 27 37 46 50 60 69

6 97.25 0.0012946 1 13 18 30 33 47 50 63 71

7 96.75 0.000746155 6 12 22 26 35 46 52 63 66

8 88.65625 0.00238121 2 13 23 30 34 46 53 63 69

9 85.21875 0.002649649 3 13 17 25 38 47 53 58 70

10 64.5625 0.001243033 7 10 18 29 37 43 52 63 70

17
 Table 6 Comparison of classification accuracy results

Classification Accuracy (%)

Types of
Energy and ABC selected
disturbances Energy (9) Entropy (9) All (72)
Entropy (18) features (9)

C1 96 97 97 98 100
C2 95 95 96 96 99
C3 95 95 97 95 100
C4 95 96 95 97 100
C5 96 96 97 97 100
C6 95 93 94 94 100
C7 94 96 94 95 100
C8 94 95 95 94 100
C9 93 94 95 95 100
C10 92 90 92 98 100
C11 94 94 94 97 99
C12 96 96 95 99 100
C13 98 98 98 99 100
C14 97 96 95 96 100
C15 96 97 96 97 100
C16 96 97 97 98 100

Average 95.07 95.20 95.33 96.47 99.875

7.3 Performance under noisy environment

An actaul power system consists of random noise riding data due to electromagnetic inference or
communication lines passing through distribution lines. Therefore, the proposed approach has been
investigated under noisy atmosphere. The performance of the proposed classifier has been tested by
adding uniformly distributed Gaussian noise to the PQ signals. A white Gaussian noise ratio of different
values 20db, 30db and 40db is uniformly applied to all parameteric equations of the PQDs. The results for
classification accuracy are shown in Table 7.

Table 7 Performance under noisy environment

Classification Accuracy (%)

Features
Without 20 dB
30dB Noise 40dB Noise
noise Noise
Energy (9) 95.07 90.1 93.1 94.1
Entropy (9) 95.20 90.5 92.5 94.5
Energy and
Entropy 95.33 91.4 93.4 94.4
(18)
All (72) 96.47 93.1 94.1 96.1
ABC
selected 99.875 93.6 95.2 98.6
features (9)

18
7.4 Classification performance using distribution network data

The proposed optimal feature selection approach was validated by different PQDs created in a
typical underground distribution network of TNB Malaysia using PSCAD/EMTDC power system
simulation software. Fig. 7 shows the schematic diagram of a typical 11 kV radial distribution network
system in Malaysia. The network consists of two 11 kV feeders outgoining from 132 kV grid system
through two 132/11 kV, Υ/Δ stepdown transforms.

n1

132/11kV 132/11kV
Υ/Δ Υ/Δ

Bus 1 n2 Measurement
Node
Feeder 2
n20 Feeder 1 n3

n4
n34 n21

n35 n14
n22 n5

n36 n23 n15 n17 n6

n24
n7
n37 n16 n18

n38 n26 n25 n19 n8 n12

n27
n9 n13

n28 n30 n10

n29 n31 n11

n32

n33

Fig. 7. Schematic diagram of a typical underground distribution system

The PQDs produced in the distribution system due to line faults, capacitor switching incidents and
nonlinear loads have been captured at bus 1 using PSCAD/EMTDC (ver 4.5) power system software. The
main objective is to monitor the PQDs at 11 kV bus (node 1). The PQDs have been obtained by applying
various types of faults and events. Sag, swell and interruption have been generated by applying line faults
such as single line-to-ground (SLG) faults, line-to-line (LL) faults, double line-to-ground (LLG) faults
and three-phase (LLL) faults. The spikes, impulsive and oscillatory transients have been created by
applying switching-of a large capacitor bank in order to maintain the power factor of the distribution
system. The harmonics, flicker and notches disturbances have been created by applying non-linear power
electronic based converters such as six pulse rectifiers. Similarly, multiple PQDs have been generated due
to the simultaneous applications of line faults and other loads in distribution network. In this way 16 types
of PQDs including normal sinusoidal signal have been generated. The overall 1600 disturbances were
generated consisting of 100 for each type. The classification performance of the proposed PNN-ABC
based algorithm for the simulated PQ data is reported in Table 8. Also the the performance of the PNN is
compared with the RBF and MLP neural networks using the optimal features obtained by the proposed
PNN-ABC algorithm. The PNN shows the highest accuracy as compared to the RBF and MLP neural
networks.

19
 Table 8 Classification results for real-time simulated PQ data

Types of Training /testing Classification Accuracy (%)

Disturbances data set
PNN MLP RBF

Normal 100 99 97 98
Sag 100 98 96 98
Swell 100 99 97 99
Interruption 100 100 95 96
Harmonics 100 99 97 97
Sag with Harmonics 100 98 94 94
Swell with Harmonics 100 96 94 96
Interruption with Harmonics 100 99 95 95
Flicker 100 100 93 95
Oscillatory Transient 100 98 92 96
Impulsive Transient 100 99 94 97
Periodic Notch 100 97 94 95
Spike 100 99 98 98
Flicker with harmonics 100 100 95 98
Flicker with sag 100 98 96 97
Flicker with swell 100 99 97 97

Average 98.625 95.25 96.625

7.5 Comparison of computational weight

Since, a large feature vector requires a high computational cost. The proposed technique is useful
due to the reduced features representing the core characteristics of signal without losing its distinctive
properties. The optimal feature selection and PNN spread constant optimization algorithm can decrease
the computational time of monitoring the disturbances as well as enhance the classification accuracy. The
selection of an optimal feature set and an optimal PNN spread constant decrease the computational time
for classification. Moreover, it decreases memory space, reduces pre-processing requirements, and
increases the computational speed. Table 9 shows the time comparison of classification of 100
disturbances of each type consisting of all 72 features, 18 features and optimally selected 9 features.
Likewise, the time comparison of other techniques such as PNN, MLP and RBF classifiers is shown in
Table 9. The PNN is found the fastest classifier for the automatic classification of PQDs.

Table 9 Comparison of computational time for proposed method

Architecture of All features 18 features Optimal features

neural network (sec) (sec) (sec)

PNN 1.200 0.8324 0.1762

MLP 6.025 2.5634 1.5454

RBF 2.548 1.243 0.8654

20
7.6 Comparative analysis

The classification performance of the proposed WT based PNN-ABC technique has been
compared with other recently reported methods, which are presented in Table 10. These methods include
Wavelet Packet Transform (WPT) and Genetic Algorithm/Simulated Annealing-Support Vector Machine
( GA/SA-SVM0) [40], Kalman Filter (KF) and Fuzzy Expert System (FES) [43], WT and Particle Swarm
Optimization-Extreme Learning Machine (PSO-ELM) [2] and WT and Fuzzy ARTMAP Neural Network
(FANN) [44]. The comparison in Table 10 shows that the proposed algorithm is capable of classifying
the 16 types of single stage and multiple PQDs with the highest classification accuracy of 99.875%.
Therefore, this method is suitable for practical implementation for automatic classification of PQDs.

Table 10 Comparison of classification accuracy with the existing proposed methods

References

PQ Disturbance WPT and GA/SA-SVM [40] WT and

KF and Fuzzy WT-FANN Proposed WT and
PSO-ELM
ES [43] [44] PNN-ABC
GA-SVM SA-SVM [2]
Normal --- --- --- --- --- 100
Sag 96.67 96.67 100 98 98.46 99
swell 100 100 100 100 100 100
Interruption 93.33 93.33 100 100 100 100
Harmonics 100 100 97 100 99.73 100
Sag with
96.67 96.67 98 98 99.15 100
Harmonics
Swell with
93.33 90.00 98 100 100 100
Harmonics
Interruption with
--- --- --- --- --- 100
Harmonics
Flicker 100 100 98 100 100
Oscillatory
100 100 96 100 100
Transient
Impulsive
--- --- --- --- --- 99
Transient
Periodic Notch --- --- --- 86 --- 100
Spike --- --- 98 100 --- 100
Flicker with
--- --- --- --- --- 100
harmonics
Flicker with sag --- --- --- --- --- 100
Flicker with swell --- --- --- --- --- 100
No. of PQ
08 08 07 10 08 16
Disturbances
Overall Accuracy 98.33 97.92 98.71 97.60 99.66 99.875

8. Conclusion

A novel approach of optimal feature selection has been presented for the classification of the PQDs
in order to identify the sources of the disturbances. The proposed PNN-ABC classification system has the
capability of discriminating the single and multiple PQDs. The PNN based classifier in conjunction with
the ABC optimization technique evaluates the classification accuracy of the optimally selected feature
subsets and the PNN spread constant. The best suitable features and the spread constant have been
selected based on the highest classification accuracy obtained by the proposed PNN-ABC classification
system. The performance of the obtained optimal features is also evaluated with the RBF and MLP neural

21
network classifiers. The simulation results show that the proposed PNN-ABC algorithm is a significantly
superior technique for identifying the single and multiple PQDs as compared to traditional classification
systems. In future work, the capability of the proposed technique can be verified using various aspects of
PQ data such as using inter-harmonics disturbances and applying the real-time data of PQDs.

Acknowledgements
This work was financially supported by the ministry of higher education (MOHE) Malaysia under
Research University Grant (RUG) Program (vote number: 00M31) and Universiti Teknologi Malaysia
(UTM) Malaysia. We also would like to thank the higher education commission (HEC) of Pakistan and
Quaid-e-Awam University of Engineering Science & Technology for their financial support.

References
[1] G. Singh, Power system harmonics research: a survey, European Transactions on Electrical Power, 19 (2009) 151-172.
[2] R. Ahila, V. Sadasivam, K. Manimala, An integrated PSO for parameter determination and feature selection of ELM and its
application in classification of power system disturbances, Applied Soft Computing, 32 (2015) 23-37.
[3] P.K. Dash, M. Padhee, T.K. Panigrahi, A hybrid time–frequency approach based fuzzy logic system for power island
detection in grid connected distributed generation, International Journal of Electrical Power & Energy Systems, 42 (2012) 453-
464.
[4] S. Khokhar, A.A.M. Zin, A.S. Mokhtar, N.A.M. Ismail, MATLAB/Simulink based modeling and simulation of power quality
disturbances, 2014 IEEE Conference on Energy Conversion (CENCON), , IEEE, Johor Baharu, 2014, pp. 445-450.
[5] H. Erişti, Ö. Yıldırım, B. Erişti, Y. Demir, Optimal feature selection for classification of the power quality events using
wavelet transform and least squares support vector machines, International Journal of Electrical Power & Energy Systems, 49
(2013) 95-103.
[6] I. Standard, IEEE Recommended Practice for Monitoring Electric Power Quality, IEEE Std 1159-2009, Institute of Electrical
and Electronics Engineers., New York, USA, 2009, pp. c1-81.
[7] S. Khokhar, A.A.B. Mohd Zin, A.S.B. Mokhtar, M. Pesaran, A comprehensive overview on signal processing and artificial
intelligence techniques applications in classification of power quality disturbances, Renewable and Sustainable Energy Reviews,
51 (2015) 1650-1663.
[8] S. Avdakovic, A. Nuhanovic, M. Kusljugic, E. Becirovic, Applications of wavelets and neural networks for classification of
power system dynamics events, Turkish Journal of Electrical Engineering & Computer Sciences, 22 (2014) 327-340.
[9] B. Biswal, P. Smieee, M. Biswal, Nonstationary Power Signal Time Series Data Classification Using LVQ Classifier,
Applied Soft Computing, 18 (2014) 158-166.
[10] L.C.M. Andrade, M. Oleskovicz, R.A.S. Fernandes, Adaptive threshold based on wavelet transform applied to the
segmentation of single and combined power quality disturbances, Applied Soft Computing, (2015).
[11] D. De Yong, S. Bhowmik, F. Magnago, An effective Power Quality classifier using Wavelet Transform and Support Vector
Machines, Expert Systems with Applications, 42 (2015).
[12] P. Kanirajan, V.S. Kumar, Power Quality Disturbance Detection and Classification using Wavelet and RBFNN, Applied
Soft Computing, 35 (2015) 470-481.
[13] H. Shunfan, L. Kaicheng, Z. Ming, A Real-Time Power Quality Disturbances Classification Using Hybrid Method Based on
S-Transform and Dynamics, IEEE Transactions on Instrumentation and Measurement, 62 (2013) 2465-2475.
[14] P.K. Ray, S.R. Mohanty, N. Kishor, J.P.S. Catalao, Optimal Feature and Decision Tree-Based Classification of Power
Quality Disturbances in Distributed Generation Systems, Ieee T Sustain Energ, 5 (2014) 200-208.
[15] D. Saxena, S. Singh, K. Verma, S.K. Singh, HHT-based classification of composite power quality events, International
Journal of Energy Sector Management, 8 (2014) 146-159.
[16] G.T. Heydt, P.S. Fjeld, C.C. Liu, D. Pierce, L. Tu, G. Hensley, Applications of the windowed FFT to electric power quality
assessment, IEEE Transactions on Power Delivery, 14 (1999) 1411-1416.
[17] F. Jurado, J.R. Saenz, Comparison between discrete STFT and wavelets for the analysis of power quality events, Electric
Power Systems Research, 62 (2002) 183-190.
[18] B. Biswal, P. Dash, B. Panigrahi, Non-stationary power signal processing for pattern recognition using HS-transform,
Applied Soft Computing, 9 (2009) 107-117.

22
[19] R. Kumar, B. Singh, D.T. Shahani, Recognition of Single-stage and Multiple Power Quality Events Using Hilbert–Huang
Transform and Probabilistic Neural Network, Electric Power Components and Systems, 43 (2015) 607-619.
[20] T.A. Kawady, N.I. Elkalashy, A.E. Ibrahim, A.-M.I. Taalab, Arcing fault identification using combined Gabor Transform-
neural network for transmission lines, International Journal of Electrical Power & Energy Systems, 61 (2014) 248-258.
[21] J. Upendar, C.P. Gupta, G.K. Singh, Statistical decision-tree based fault classification scheme for protection of power
transmission lines, International Journal of Electrical Power & Energy Systems, 36 (2012) 1-12.
[22] O. Ozgonenel, S. Karagol, Transformer differential protection using wavelet transform, Electric Power Systems Research,
114 (2014) 60-67.
[23] A. Bakar, M. Ali, C. Tan, H. Mokhlis, H. Arof, H. Illias, High impedance fault location in 11kV underground distribution
systems using wavelet transforms, International Journal of Electrical Power & Energy Systems, 55 (2014) 723-730.
[24] A. Samui, S. Samantaray, Wavelet Singular Entropy-Based Islanding Detection in Distributed Generation, IEEE
Transactions on Power Delivery, 28 (2013) 411-418.
[25] R. Ahila, V. Sadasivam, K. Manimala, Particle swarm optimization-based feature selection and parameter optimization for
power system disturbances classification, Applied Artificial Intelligence, 26 (2012) 832-861.
[26] M.A.S. Masoum, S. Jamali, N. Ghaffarzadeh, Detection and classification of power quality disturbances using discrete
wavelet transform and wavelet networks, IET Science, Measurement & Technology 4(2010) 193-205.
[27] B. Biswal, S. Mishra, Power signal disturbance identification and classification using a modified frequency slice wavelet
transform, IET Generation, Transmission & Distribution, 8 (2014) 353-362.
[28] N. Huang, D. Xu, X. Liu, L. Lin, Power quality disturbances classification based on S-transform and probabilistic neural
network, Neurocomputing, 98 (2012) 12-23.
[29] M. Hajian, A. Akbari Foroud, A new hybrid pattern recognition scheme for automatic discrimination of power quality
disturbances, Measurement, 51 (2014) 265-280.
[30] S. Gunal, O.N. Gerek, D.G. Ece, R. Edizkan, The search for optimal feature set in power quality event classification, Expert
Systems with Applications, 36 (2009) 10266-10273.
[31] B.K. Panigrahi, V.R. Pandi, Optimal feature selection for classification of power quality disturbances using wavelet packet-
based fuzzy k-nearest neighbour algorithm, IET Generation, Transmission & Distribution, 3 (2009) 296-306.
[32] C.-Y. Lee, S. Yi-Xing, Optimal Feature Selection for Power-Quality Disturbances Classification, IEEE Transactions on
Power Delivery, 26 (2011) 2342-2351.
[33] F.G. Mohammadi, M.S. Abadeh, Image steganalysis using a bee colony based feature selection algorithm, Engineering
Applications of Artificial Intelligence, 31 (2014) 35-43.
[34] T.-J. Hsieh, H.-F. Hsiao, W.-C. Yeh, Forecasting stock markets using wavelet transforms and recurrent neural networks: An
integrated system based on artificial bee colony algorithm, Applied soft computing, 11 (2011) 2510-2525.
[35] M. Hajian, A. Akbari Foroud, A.A. Abdoos, New automated power quality recognition system for online/offline monitoring,
Neurocomputing, 128 (2014) 389-406.
[36] D.D. Ferreira, J.M. de Seixas, A.S. Cerqueira, A method based on independent component analysis for single and multiple
power quality disturbance classification, Electric Power Systems Research, 119 (2015) 425-431.
[37] R. Kumar, B. Singh, D.T. Shahani, A. Chandra, K. Al-Haddad, Recognition of Power-Quality Disturbances Using S-
Transform-Based ANN Classifier and Rule-Based Decision Tree, Ieee T Ind Appl, 51 (2015) 1249-1258.
[38] S.G. Mallat, A theory for multiresolution signal decomposition: the wavelet representation, IEEE Transactions on Pattern
Analysis and Machine Intelligence, 11 (1989) 674-693.
[39] H. Dehghani, B. Vahidi, R.A. Naghizadeh, S.H. Hosseinian, Power quality disturbance classification using a statistical and
wavelet-based Hidden Markov Model with Dempster–Shafer algorithm, International Journal of Electrical Power & Energy
Systems, 47 (2013) 368-377.
[40] K. Manimala, K. Selvi, R. Ahila, Optimization techniques for improving power quality data mining using wavelet packet
based support vector machine, Neurocomputing, 77 (2012) 36-47.
[41] S. Mishra, C.N. Bhende, K.B. Panigrahi, Detection and Classification of Power Quality Disturbances Using S-Transform
and Probabilistic Neural Network, IEEE Transactions on Power Delivery, 23 (2008) 280-287.
[42] D. Karaboga, Artificial bee colony algorithm, scholarpedia, 2010, pp. 6915.
[43] A.A. Abdelsalam, A.A. Eldesouky, A.A. Sallam, Characterization of power quality disturbances using hybrid technique of
linear Kalman filter and fuzzy-expert system, Electric Power Systems Research, 83 (2012) 41-50.
[44] J.G.M.S. Decanini, M.S. Tonelli-Neto, F.C.V. Malange, C.R. Minussi, Detection and classification of voltage disturbances
using a Fuzzy-ARTMAP-wavelet network, Electric Power Systems Research, 81 (2011) 2057-2065.

23
 Research Highlights

 Power quality disturbances are classified using optimal feature selection method.
 Discrete wavelet transform extracts the set of features successfully.
 Artificial bee colony has been proposed for the selection of optimal features.
 16 types of single and multiple power quality disturbances are classified.
 Superior results are achieved as compared to the similar existing techniques.

24