DETECTION AND CLASSIFICATION OF POWER QUALITY

DISTURBANCES USING SIGNAL PROCESSING BASED

ARTIFICIAL NEURAL NETWORKS

Dr. T. Jayasree
Associate Professor, Department of ECE,
Government College of Engineering, Bodinayakanur
&
Dr. T. Selvin Retna Raj
Assistant Professor, Department of ECE,
DMI College of Engineering, Chennai.
1
 Outline

 Introduction

 Power quality problems and objective

 Disturbance detection methods

 . Feature extraction modules

 Classification modules

 Comparison results

 Summary
2
 Introduction – Power quality

 Power quality can be defined as the measure, analysis and improvement of

voltage or current to maintain a sinusoidal waveform at rated voltage and
frequency.
 Ideally, the AC voltage wave is a sine wave alternating from a positive
peak, with 60Hz frequency without any deformations, spikes or surges.

 In reality different factors such as lighting, disturbances coming from the

loads and presence of non – linear devices influence the quality of the
wave.
3
 Any deviation displayed in those power signals from the normal values is
considered as a power quality disturbances
4
Contd …

 Voltage Sag - Occurs due to the faults on the distribution or transmission

network. Faults during the installation of machines, startup of large motors
and connection of heavy loads.

 Voltage Swell - Occurs due to the badly regulated transformers, badly

dimensioned power sources and start/stop of heavy motors.

 Harmonics - arises due to standard sources such as welding machines, DC

brush motors and Rectifiers.

 Transients - occur due to the disconnection of heavy loads, Lightning and

switching of lines.

 Flickers - are due to the presence of arc furnaces, frequent start/stop of

5
electric motors and oscillating loads
 Problem statement

 Electrical power system is a large and complex network, where power quality
disturbances must be monitored, analyzed

 Power quality disturbances are mitigated continuously in order to preserve

and to re-establish the normal power supply without even slight interruption.

 Practically huge disturbance data is difficult to manage and requires the

higher level of accuracy and time for the analysis and monitoring.

 Thus automatic and intelligent algorithm based methodologies are in practice

for the detection, recognition and classification of power quality events.

 This approach may help to take preventive measures against abnormal

6
operations.
Challenges in Power quality disturbance detection and
classification

 The choice of signal processing techniques for the analysis of power

quality disturbance signals in time frequency domain.

 Selection of proper tools for monitoring and detection of power quality

disturbances.

 Detection of disturbances under noisy conditions

 Selection of appropriate feature extraction methods for getting more

information from power quality disturbance signals

 The choice of artificial intelligent technique conferring to classification

performance characteristics are the main concerns and challenges.
7
 Aim and Objective
 Aim: To develop signal based methods for the detection and classification of
power quality disturbances.

 Objectives:
 To develop transformation based methods for the detection of power quality
disturbances in time – frequency domain.

 To develop different feature extraction methods for extracting important information

from various power quality disturbances.

 Development of ANN based classification models for classifying power quality

disturbances.

 Comparative analysis between different power quality disturbance detection, feature

extraction and classification.

8
 Literature survey
Author and Title Method Inference

Kankale, Classification of Power Quality Energy based features Classification of voltage

Disturbances in Emerging Power System using DWT and KNN (K- sag and swell disturbance is
DWT and K-Nearest Neighbour, SPAST Abstracts Nearest Neighbourhood) performed.
(2017),Vol.01, P.No,56-65. classifier
Rajender Kumar Beniwal, Manish Kumar Saini, Anand RMS value of DWT MLP classifier produced
Nayyar, Basit Qureshi & Akanksha Aggarwal, A coefficients are used as less performances
critical Analysis of Methodologies for detection and features and
Classification of Power Quality Events, IEEE classification using Multi
Transaction on Smart Grid, (2019): P.No: 83507-83534. layer perceptron (MLP)

RamaMohana Reddy, Ram Kishore Kumar Reddy, DWPT with Adaptive Classification is performed
Sujatha, Power Quality Classification of Neuro-Fuzzy system only for three disturbances.
Disturbances using Wavelet Packet Transform with
Adaptive Neuro-Fuzzy system, Turkish Journal of
Computer and Mathematics Education (2020), Vol.12,
no.3, pp.4892-4903
Rajendra Mahla, Baseem Khan, Om Prakash Mahela, Wavelet Packet based Difficult to frame rules for
and Anup Singh, Recognition of complex and multi fast kurtogram and ruled classifying disturbances
9
power quality disturbances using Wavelet Packet decision tree
based fast kurtogram and ruled decision tree
 Literature survey
Author and Title Method Inference

Dan Su, Kaicheng Li, and Nian Shi, Power Quality Modified S- Transform Power quality disturbance
Disturbances Recognition Using Modified S- based features detection is performed
Transform Based on Optimally Concentrated based on optimally
Window, Integration of Renewable Energy, concentrated window
Sustainability, (2020)Vol.13, P.No. 1-14 function
Nitin Kumawat, Kumar Verma, Zaveri, Comparison DWPT based features Three level decomposition
between Wavelet Packet Transform and M- band are used is performed and energy
Wavelet Packet Transform for identification of Power based features are extracted.
Quality Disturbances, Power Research, (2021)Vol.64,
no.1, P.No. 37-45

Shouxiang Wang & Haiwen Chen, A novel deep Continuous Wavelet CWT based images are
learning method for the classification of power quality Transform (CWT) and used for the discrimination
disturbances using convolutional neural network, Convolutional neural of signals. CNN is used for
Applied energy, (2022), P. No:1126-1140 network (CNN) classification.
Achlerkar, Samantaray, & Manikandan. Variational Variational mode Difficult to frame decision
mode decomposition and decision tree based detection decomposition for rules for classifying
and classification of power quality disturbances in feature extraction and disturbances 10
grid-connected distributed generation system, IEEE decision tree for
Transactions on Smart Grid, (2022), 9(4), 3122-3132. classification
 Literature survey
Author and Title Method Inference
Chirag A Naik & Prasanta Kundu, Wavelet Wavelet Packet Used only for the
Packet Transform Based Parameter for the Transform classification of short
Analysis of Short Duration Power Quality duration power quality
Disturbances, IFAC-Papers, 2020, disturbances
48(30),P.No: 485-48
Kalyana Sundaram, Neela, Automatic S- Transform and MLP The MLP classifier
Classification of Power Quality Disturbances Neural network produced
using S- Transform and MLP Neural
network, International Science Press, (2021),
P.No. 31 -45.
Deokar, Waghmare, Integrated DWT- FFT DWT- FFT , Feed 3 – level decomposition
approach for detection and classification forward neural networks is performed and FFT is
Duration Power Quality Disturbances, applied for finding
International Journal of Electrical Power & spectrogram
Energy Systems, (2022) 61, 594-605.
Li, Jianmin, Zhaosheng Teng, Qiu Tang, and double resolution S- SVM produced less
Junhao Song. Detection and classification of transform and SVM performances 11
power quality disturbances using double
resolution S-transform and SVMs, IEEE
Literature Summary
 Detection of disturbances is carried out by representing the power quality
disturbances in time – frequency domain.

 The application signal processing based transformation techniques are

important for the detection of power quality disturbances.

 Standard mathematical equations are used for the generation of power

quality disturbances.

12
Overall flow diagram

13
 Method 1: DWT based power quality disturbance detection
and feature extraction
 The hierarchical decomposition of an input signal into a succession of gradually lower
resolution signals is called Multi Resolution Analysis (MRA).
 The input signal is passed through a set of high pass and low pass filters. Subsequently, down
sampling is used to remove half of the signal samples
 The low pass filter coefficients are called as approximation coefficients, and the high pass filter
coefficients are called as detailed coefficients,

14
 DWT based feature extraction - Steps
 Step 1: Signal x(n) is decomposed into 4 levels using DWT-MRA technique, in the 4th level, set
of 4 detailed coefficients, {D1, D2, D3, D4} and one set approximation coefficients {A4} are
produced.
 Step 2: Find mean, median, energy and variance for all the coefficients in the 4th level.

DWT Coefficients Mean Median Energy Variance

D1 M0 N0 E0 V0

D2 M1 N1 E1 V1

D3 M2 N2 E2 V2

D4 M3 N3 E3 V3

A4 M4 N4 E4 V4

15
 Generation of DWT based distribution patterns
 Step 1: Apply Multi Resolution Analysis (MRA) based decomposition
technique in order to obtain the DWT coefficients.
 Step 2: In the Nth level of decomposition, the DWT coefficients are {D1, D2,
D3, ………DN, AN}.
 Step 3: Extract statistical features from the coefficients.
 Mean
 Variance DWT coefficients Mean Variance
D1 M0 V0
D2 M1 V1
D3 M2 V2
D4 M3 V3
……… ………. ……….
DN MM VM
AN MN VN

 Step 4: Plot the values in order to obtain various distribution patterns. 16

 Generation of power quality disturbances

 Power quality disturbance signals are using mathematical equations

based on IEEE 1159 - 2009 standards.
Disturbance Mathematical model
Model parameters
type expression
Sinusoidal
signal

Swells

Sags

Transient

Harmonics 17

Flickers
 Results – Normal and Power quality disturbances
 Implemented in Matlab 2018b software. Length of the signal = 4000 samples,
Sampling frequency = 10kHz
 The normal signal and the five types of disturbances such as voltage sag,
voltage swell, transients, flicker and harmonics considered in this research
work.

18
Results - DWT Cofficients for voltage sag signal

19
The start and end of the of the voltage disturbance is clearly detected in
all the DWT coefficients D1, D2, D3, D4 and A4
DWT Coefficients for voltage swell

 The start and end of the of the voltage disturbance is clearly

detected in all the DWT coefficients D1, D2, D3, D4 and A4 20

 Magnitude of the spikes vary and depends on the signal

DWT Coefficients for transients and harmonics signal

 Transients: only one spike is seen in D1, D2, D3, D4 and A4,
which represents the presence of disturbance. 21

 Harmonics: Uniform pattern is attained in all the cases. It is

observed that magnitude is approximately zero for some cases.
DWT Coefficients for Flicker and normal signal

 It is observed that uniform pattern is attained in all the cases.

22 It
is also seen that magnitude is approximately zero for some cases.
 Disturbance detection in the presence of noise
Input signals are added with Additive Gaussian noise and the DWT is applied.

Voltage sag/swell with AWGN of 30 db and DWT coefficients

 Disturbances are detected in the coefficients A1 and A2. i.e.,23at

low frequency sub bands.
 Voltage transient signal AWGN of 30 db and DWT coefficients

 It is observed that transient disturbances are detected in all the

24
coefficients.
 Mean distribution patterns

 DWT based
decomposition is
performed upto nine
levels.
 Mean and variance
are calculated at the
9th level and plotted.

Variance distribution patterns

 It is observed that
Mean and variance
distribution patterns
show unique
distribution patterns
for voltage sag, swell,
transient, harmonics25
and flicker
disturbances.
Method 2 :Discrete Wavelet Packet Transform (DWPT) based
Disturbance detection
 Here, input signal x(n) is passed through LPF, g(n) and high pass
filter h(n), and then decimation operation is applied.
 Both the LPF coefficients, i.e., approximations and HPF
coefficients, i.e., detailed coefficients are decomposed successively.

26
 Wavelet Packet Tree after performing 4 – level decomposition

 The tree structure contains 31 nodes represented by (0,0), (1,0), (1,1), (2,0),
(2,1), (2,2), (2,3), (3,0), (3,1), (3,2), (3,3), (3,4), (3,5), (3,6), (3,7), (4,0),27(4,1),
(4,2), (4,3), (4,4), (4,5), (4,6), (4,7), (4,8), (4,9), (4,10), (4,11), (4,12), (4,13),
(4,14) and (4,15)
 DWPT based features
Nodes Mean Median Energy Variance
 After performing 4 – level (4,0) M0 N0 E0 V0
decomposition, find mean, (4,1) M1 N1 E1 V1
median and variance for all the (4,2) M2 N2 E2 V2
(4,3) M3 N3 E3 V3
nodes. (4,4) M4 N4 E4 V4
(4,5) M5 N5 E5 V5
(4,6) M6 N6 E6 V6
(4,7) M7 N7 E7 V7
(4,8) M8 N8 E8 V8
(4,9) M9 N9 E9 V9
(4,10) M10 N10 E10 V10
(4,11) M11 N11 E11 V11
(4,12) M12 N12 E12 V12
(4,13) M13 N13 E13 V13
(4,14) M14 N14 E14 V14
(4,15) M15 N15 E15 V15

28
 DWPT based disturbance detection – Voltage sag signal
(after performing 4 – level decomposition)

 Voltage sag disturbance is detected in all sub bands.

29
 Two spikes are seen in all the sub bands which represents stat and
end of the disturbance.
 DWPT based disturbance detection – Voltage swell signal
(after performing 4 – level decomposition)

 Voltage swell disturbance is detected in all sub bands.

 Two spikes are seen in all the sub bands which represents stat 30
and
end of the disturbance.
DWPT based disturbance detection – Voltage transient signal
(after performing 4 – level decomposition)

 Voltage transient disturbance is detected in all sub bands.

 The presence of spike in all the sub bands represents the existence
31

of transient disturbance.
 DWPT based disturbance detection – Voltage Harmonics
signal

 Harmonics signal shows uniform magnitude variations in the32 sub

bands. For some of the nodes, the magnitude is approximately zero
as shown in Figure.
 DWPT based disturbance detection – Voltage Flicker signal

 Voltage flicker signal shows uniform magnitude variations in

33 the
sub bands. For some of the nodes, the magnitude is approximately
zero as shown in Figure.
 Comparison of features for different types of power quality
disturbances (using DWT and DWPT)

 Comparing DWT and DWPT features, it is found that DWPT feature values are
high for all the disturbances.
34
 Energy value is found to be high for all the cases and for the transient
disturbance, energy is maximum.
 Method 3: S – Transform based Power quality disturbance
detection
 The S-transform (ST) is used for generating the temporal frequency
representation of the signal. It was proposed by Stockwell R.G. in 2001

35
Algorithm for computation of S – Transform

36
S – Transform based feature extraction

37
 Results: Disturbances detection using ST contour and 3-D
plots (Voltage swell)

 As shown in the Figure, the presence of contour represents in the

presence of Voltage swell disturbance.
 The ST 3D plot shows increase in amplitude during the time 38 of
voltage swell disturbance.
 Disturbances detection using ST contour and 3-D plots
(Voltage sag)

 As shown in the Figure, the presence of contour represents in the

presence of Voltage sag disturbance.
 The ST 3D plot shows decrease in amplitude during the time of
voltage sag disturbance. 39
 Disturbances detection using ST contour and 3-D plots
(Voltage harmonics and Flicker)

 As shown in the Figure, the magnitude of harmonics and flicker

signal varies uniformly.
 Hence, ST contour and ST 3D plot show uniform amplitude
variations. 40
 Disturbances detection using ST contour and 3-D plots
(Voltage transient signal)

 As shown in the Figure, the presence of contour represents in the

presence of Voltage transient signal with noise with SNR 25 dB.41
 The ST 3D plot shows increase in amplitude during the time of
voltage transient signal with noise.
 Disturbances detection using ST contour and 3D plots
(Voltage sag signal with noise)

 As shown in the Figure, the presence of contour represents the

presence of Voltage transient signal with noise with SNR 20 dB.
 The ST 3D plot shows increase in amplitude during the time of
voltage sag signal with noise. 42
 Comparison results of S Transform based for different power
quality disturbances
 Features such as mean, median, energy and variance are calculated
for all the power quality disturbances from the ST matrix.
 As shown in Figure, the values of all the features are different for
each disturbance.

43
Method 4: DWT-ST based power quality disturbance detection
and feature extraction
Step 1: First find the DWT of
the power quality disturbance
signal.
Step 2: Obtain the
approximation and detailed
coefficients by performing 4
level decomposition level.
Step 3: Find the ST of the
DWT coefficients in the fourth
level.
Step 4: Derive the ST Matrix,
ST Contour, ST 3D plots.
Step 5: Extract the features
such as Mean, Median,
Variance, Energy.
44
 DWT-ST contours and 3D plots – Voltage swell signal

 As illustrated in Figure, the start and end of the voltage swell

disturbance is represented by the presence of contours in all the
sub bands. 45
 ST 3D plots show increase in signal magnitude at the start and
end of the swell disturbance.
 DWT-ST contours and 3D plots – Voltage sag signal

 As illustrated in Figure, the start and end of the voltage sag

disturbance is represented by the presence of contours in the sub
bands. 46
 ST 3D plots show increase in signal magnitude at the start and end
of the swell disturbance.
 DWT-ST contours and 3D plots – Voltage transient signal

 The presence of voltage transient disturbance is represented by

the presence of contours in all the sub bands.
 ST 3D plots show increase in signal magnitude during the time
47 of
occurrence of the transient disturbance.
 DWT-ST contours and 3D plots – Voltage harmonics signal

 Voltage harmonics signal shows uniform magnitude variations in

the contours in all the sub bands. 48
 ST 3D plots also shows uniform variations in signal magnitudes.
 DWT-ST contours and 3D plots – Voltage flicker signal

 Voltage flicker signal shows uniform magnitude variations in the

contours in all the sub bands. 49
 ST 3D plots also shows uniform variations in signal magnitudes.
 Method 5: DWPT-ST based power quality disturbance
detection and feature extraction

Step 1: First find the

DWPT of the power quality
disturbance signal.
Step 2: Obtain the
approximation and detailed
coefficients by performing 4
level decomposition level.
Step 3: Find the ST of the
DWT coefficients in the
fourth level.
Step 4: Derive the ST
Matrix, ST Contour, ST 3D
plots.
Step 5: Extract the
features such as Mean,
Median, Variance, Energy. 50
 DWPT-ST contour for voltage swell signal

 After performing 4 – level decomposition using DWPT, in the 4th level, the
nodes are N(4,0), N(4,1), ………….N(4,15)
 In the above figure, it is observed that, two contours are present at the 5051th and
100th sample, in most of the nodes which represents start and end of the sag
disturbance.
 DWPT-ST 3D plots for voltage swell signal

 In all DWPT-ST 3D plots it is observed that the signal magnitude is increased

to some extend which denotes the presence of voltage swell signal.
52
 DWT-ST contours and 3D plots – Voltage transient signal

 The presence of contours in all the sub bands represents the

presence of transient disturbance.
 Similarly, ST 3D plots show variations in signal magnitude at53 the
time of occurrence of the transient disturbances.
 DWT-ST contours and 3D plots – Voltage harmonics signal

 The magnitude of the contour shows uniform variations for the

harmonics signal in most of the sub bands.
 Similarly, ST 3D plots show significant variations in signal
54
magnitude in some of the sub bands.
 DWT-ST contours and 3D plots – Voltage flicker signal

 The magnitude of the contour shows uniform variations for the

voltage flicker signal in most of the sub bands.
 Similarly, ST 3D plots show significant variations in signal
55
magnitude in some of the sub bands.
 Feature extraction using DWT-ST and DWPT-ST
After performing DWT-ST and DWPT-ST, the average values of mean,
median, variance and energy are computed for all the power quality
disturbances.

 Comparing DWT-ST and DWPT-ST features, it is found that DWPT-ST feature

values are high for some of the disturbances. 56
 Energy value is found to be high for all the cases and for the transient
disturbance, energy is maximum.
 Classification of Power quality disturbances
using ANN models

 The classification of power quality disturbances is done by means of

Artificial Neural Network (ANN) classifiers.

 Artificial Neural Networks (ANNs) can be considered to be computer

systems that use sets of algorithms that are inspired by and loosely modeled
after biological neural networks that are found in human brains.

 Artificial neural networks have the ability to perform multiple functions

simultaneously.

 The different set of features are extracted using the DWT, DWPT and ST,
DWT-ST and DWPT-ST are given as input to the ANN classifiers.
57
 The performance of the network is evaluated by finding various
performance metrics.
 Probabilistic Neural Network (PNN) Classifier
 PNN is a feed forward network consisting of 4 layers: 1) Input layer, 2) Pattern layer, 3)
Summation layer 4) Output layer.
o When an input is applied, the distance from the input vector to the training input vectors is
computed by pattern layer. This layer produces a vector, whose elements specify how close the
input is to the training input.
o The summation layer sums the involvement for each class of inputs and generates the net output
as a vector of probabilities.

o The output layer contends transfer

function on the output of the
second layer and choose the
maximum of these probabilities,
and produces a 1 (positive
credentials) for that class and a 0
(negative credentials) for non-
targeted classes.
o The PNNs are faster and more
accurate than multilayer
perceptron networks (MLPs).
58
 Classification Algorithm using PNN
Step 1: Form the input vector and assign the class numbers. Divide the input into
k sets. Each set contains one class of vectors.

Step 2: Generate a cumulative Gaussian output function for each k.

Step 3: Read the input vector and assign Gaussian function for each class.
Compute all Gaussian functional values for each class, for each cluster of hidden
nodes.

Step 4: Feed all the Gaussian functional values from the hidden node cluster to
the single output node of cluster.

Step 5: Add all the inputs and multiply by a constant for each category of the
output node. Find the most valuable of all the useful values added together at the
59
output nodes.
 Radial Basis Function Neural Network (RBFNN) Classifier
 The RBFNN consists of an input layer, a hidden layer and an output layer
 The hidden layer contains a transfer function called the multivariate Gaussian
Density function

j

h
yk ( x) = w 
j =1
kj j ( x ) + wk 0

60
classification Algorithm using RBFNN

61
 Recurrent Neural Network (RNN) classifier

 Recurrent Neural Networks(RNN) are a type of Neural Network where the

output from the previous step is fed as input to the current step.
 The output from the previous step are fed as input to the present step. The
most important feature of RNN is the presence of hidden state, which
recollects some information about a sequence.

62
 Classification Algorithm using RNN
 The input feature vector is represented as xᵢ. The weights associated with the input is denoted
using a vector U
 The hidden representation (sᵢ) of the network is computed as a function of the output of the
previous time step and current input along with bias.
 In this approach, we will predict the output after each time step.

63
 Performance metrics

F Score
F Score is a measure used for finding the accuracy of the model on a dataset.
64
 ANN model parameters

 Input
 Five disturbances (5 x 500) + normal (500) – Totally 3000 signals
 Input feature vector – mean, median, standard deviation, variance ( 4 x
500 for each type)
 Training data – 75%, testing – 25%
 Implemented in MATLAB
 No of neurons in the hidden layer – chosen by trial and error
 No. of epochs: 500, MSE attained – 0.0125
 Output
 Six classes (normal + 5 disturbance types)

65
 Comparison results using DWT + PNN and RBFNN classifiers
(using different mother wavelets)
 The performances are evaluated using DWT based features as input. DWT is computed
using different mother wavelets.
 It is observed that, Db8 mother wavelet produced better performance for both PNN and
RBFNN classifiers.
Accuracy Specificity Sensitivity F-Score
PNN classifier
Db4 81.2 76.3 74.3 80.4
Db8 85.6 83.4 80.2 81.5
Sym 4 82.4 78.4 80.4 77.8
Sym8 76.2 80.5 81.4 80.0
Coif 81.4 79.2 80.6 81.7
haar 80.8 81.7 83.2 79.5
RBFNN classifier
Db4 82.4 77.9 75.2 81.6
Db8 86.1 84.2 81.2 82.6
Sym 4 83.6 79.5 81.7 78.6
Sym8 77.6 81.7 82.7 81.366
Coif 82.8 80.5 81.7 82.8
haar 81.8 82.8 84.1 80.6
 Comparison results using DWT + RNN classifier
(using different mother wavelets)

 It is observed that, Db8 mother wavelet produced better performance for RNN classifier.

Accuracy Specificity Sensitivity F-Score

RNN classifier
Db4 83.6 80.4 76.2 82.7
Db8 87.2 87.5 82.2 83.6
Sym 4 84.7 80.5 82.5 79.4
Sym8 78.6 82.7 83.5 82.3
Coif 83.5 81.6 82.8 83.8
haar 82.4 83.6 85.2 81.6
67
 Comparison results using DWT + PNN classifier under noisy conditions
(using different mother wavelets)

 The performances are evaluated under noisy conditions. Then DWT based features are
extracted from the power quality disturbance signals added with noise. DWT is
computed using different mother wavelets.
 It is observed that, Db8 mother wavelet produced better performance using PNN
classifier.

Accuracy Specificity Sensitivity F-Score

PNN classifier

Db4 80.6 74.2 73.4 78.2

Db8 83.4 81.2 78.8 80.2

Sym 4 80.6 76.2 78.2 75.2

Sym8 74.3 78.6 78.3 78.2

Coif 88.2 77.6 78.4 79.868

Haar 78.4 79.4 87.7 77.4

Performance results using DWT + RBFNN and RNN classifiers under noisy
conditions (using different mother wavelets)

Accuracy Specificity Sensitivity F-Score

RBFNN classifier
Db4 80.6 75.2 73.9 79.5
Db8 80.9 75.2 73.8 79.2
Sym 4 84.7 82.8 80.9 80.7
Sym8 81.7 77.2 79.5 76.4
Coif 75.8 79.0 80.4 80.0
haar 80.6 79.2 79.2 80.4
RNN classifier
Db4 81.4 79.4 74.3 80.6
Db8 85.9 85.2 80.8 81.2
Sym 4 82.7 78.7 80.4 77.2
Sym8 76.5 80.5 81.2 80.4
Coif 81.7 79.6 80.8 81.2
haar 80.4 80.6 83.5 79.4
69
 Table shows that, Db8 and Sym4 mother wavelets produced better
performance using RBFNN and RNN classifiers.
 Performance results using DWPT + RBFNN. PNN and RNN classifiers

 The performances are evaluated using DWPT based features under normal and noisy
conditions using PNN, RBFNN and RNN classifiers. DWPT is computed using Db8
mother wavelet.
 It is observed that, RNN produced better outcomes.

Classifier Accuracy Specificity Sensitivity F-Score

type
PNN 88.8 87.4 83.2 86.5

RBFNN 89.3 87.4 86.2 84.6

RNN 90.2 91.7 88.6 87.5

Classification under noisy conditions

PNN 85.4 84.2 78.8 81.2

RBFNN 86.7 85.8 83.9 81.7

70

RNN 87.9 85.2 81.8 82.2

Performance results using S – Transform based features using
RBFNN. PNN and RNN classifiers
 The performances are evaluated using ST – based features under normal and noisy
conditions using PNN, RBFNN and RNN classifiers.
 Input feature vector:

 It is observed that, RNN produced better outcomes.

Classifier Accuracy Specificity Sensitivity F-Score

PNN 90.2 89.6 85.4 88.7
RBFNN 91.3 89.5 87.2 86.6
RNN 91.2 93.7 89.6 89.7
Classification under noisy conditions
PNN 89.5 90.6 87.6 89.5
71
RBFNN 92.5 90.7 92.4 85.2
RNN 90.8 92.4 90.6 88.7
 Performance results using DWT-ST features using RBFNN.
PNN and RNN classifiers

 The DWT-ST based Input feature vector:

are given as input to the classifiers and the performance metrics are evaluated.
o Highest performance is obtained using RNN classifier.

Accuracy Specificity Sensitivity F-Score

PNN 92.6 91.7 94.4 95.2

RBFNN 94.8 93.5 94.6 92.5

RNN 95.6 94.5 96.3 94.3

Classification under noisy conditions

PNN 91.7 92.3 93.5 92.6
RBFNN 93.5 92.8 93.6 91.8
72

RNN 94.8 93.6 95.4 93.7

 Performance results using DWPT-ST features using RBFNN.
PNN and RNN classifiers

 The DWT-ST based Input feature vector:

are given as input to the classifiers and the performance metrics are evaluated.
o Highest performance is obtained using RNN classifier.

Accuracy Specificity Sensitivity F-Score

PNN 96.8 95.4 96.3 94.3

RBFNN 97.5 96.8 98.4 97.9
RNN 99.2 99.4 98.3 98.4
Classification under noisy conditions
PNN 95.7 95.7 95.83 94.9
73
RBFNN 96.8 95.4 97.3 96.6
RNN 98.3 98.6 98.7 98.4
 Summary
 Different signal processing methodologies and ANN models adopted for the
detection and classification of power quality disturbance waveforms is
explored in this research work.
 1. DWT based disturbance detection
 Time - frequency representation approach in which low frequency components of the
signal are divided successively.

 Only five frequency sub bands are considered for the extraction of features.

 Mean and variance distribution patterns show unique patterns which are helpful for
distinguishing various power quality disturbances.

 2. DWPT based disturbance detection

 Time - frequency representation approach in which low and high frequency components of
the signal are divided successively.
74
 Fifteen frequency sub bands are considered for the extraction of features, thus improving
the performance.
 Summary (contd..)

 3. ST based disturbance detection

 Time frequency representation approach in which the presence of power quality
disturbance is represented in the form of ST contours and ST 3D plots.
 S-Transform provided better disturbance detections under noisy conditions.
 The features are extracted from the ST matrix derived from the S – Transform.
 4. DWT-ST based disturbance detection
 The disturbance detection is performed by combining both DWT and ST.
 Features are extracted using DWPT-ST.
 5. DWPT-ST based disturbance detection
 The disturbance detection is performed by combining both DWT and ST.
 Features are extracted using DWPT-ST.

 Classification of power quality disturbances is performed using PNN, RBFNN

and RNN classifiers.

 Performance of the classifiers are evaluated using various performance metrics

75

and results are compared.

 Major research contributions

 Detection of power quality disturbances using DWT–ST

 Extraction of features using DWT–ST

 Classification of power quality disturbances using DWT-ST and PNN, DWT-

ST. RBFNN and DWT-ST and RNN

 Detection of power quality disturbances using DWT–ST

 Extraction of features using DWT–ST

 Classification of power quality disturbances using DWT-ST and PNN, DWT-

76
ST. RBFNN and DWT-ST and RNN
 Future work

 The present research work may be extended by applying other signal

processing techniques such as Hilbert transform, complex Wavelet
Transforms, frequency spectrum estimation techniques for power quality
disturbance detection and feature extraction.

 The proposed signal processing algorithms could be implemented using

digital signal processors and VLSI processors for real-time applications.
Moreover, other ANN architectures like Convolutional Neural Network
(CNN) and other Deep Neural Networks (DNN) could be explored for
automatic disturbance classification.

77
 References

 Achlerkar, PD, Samantaray, SR & Manikandan, M S 2018, ‘Variational mode

decomposition and decision tree based detection and classification of power
quality disturbances in grid-connected distributed generation system’, IEEE
Transactions on Smart Grid, vol. 9, no. 4, pp. 3122-3132.

 Ali, M, Zied, B, Djaffar, OA & Alain, D 2015, ‘A new optimized Stockwell

transform applied on synthetic and real non-stationary signals’, Digit, Signal
Process, vol. 46, pp. 226-238.

 Jaiswal, S & Ballal, MS 2017, ‘FDST-based PQ event detection and energy

metering implementation on FPGA-in-the-loop and NI-LabVIEW’, IET
Science, Measurement and Technology, vol. 11, no. 4, pp. 453-463.

 Jaramillo, SH, Heydt, GT & O’Neill-Carrillo, E 2000, ‘Power quality indices

78

for aperiodic voltages and currents’, IEEE Trans, Power Deliv, vol. 15,
 References

 Alqam, SJ & Zaro, FR 2018, ‘Power Quality Detection and Classification

Using S-Transform and Rule Based Decision Tree’.

 Bhuiyan, SM, Khan, J & Murphy, G 2018, ‘WPD for detecting disturbances
in presence of noise in smart grid for PQ monitoring’, IEEE Trans Ind Appl,
vol. 54, no. 1, pp. 702‐711.

 Biswal, M & Dash, PK 2013, ‘Measurement and classification of

simultaneous power signal patterns with an s-transform variant and fuzzy
decision tree’, Industrial Informatics, IEEE Transactions, vol. 9, no. 4, pp.
1819-1827.

 Cai, D, Li, K, He, S, Li, Y & Luo, Y 2018, ‘A highly accurate and fast power
quality disturbances classification based on dictionary learning 79
sparse
decomposition’, Transactions of the Institute of Measurement and Control.
 Publications

 Chirag A Naik & Prasanta Kundu 2015, ‘Wavelet Packet Transform Based
Parameter for the Analysis of Short Duration Power Quality Disturbances’,
IFAC-Papers On Line, vol. 48, no. 30, pp. 485-489.

 Dan Su, Kaicheng Li & Nian Shi 2021, ‘Power Quality Disturbances
Recognition Using Modified S-Transform Based on Optimally Concentrated
Window’, Integration of Renewable Energy, Sustainability, vol. 13, pp. 1-14.

 Dekhandji, FZ 2017, ‘Signal processing deployment in power quality

disturbance detection and classification’, Acta Phys Pol, vol. 13, no. 2, pp.
415‐419.

 Eristi, H & Demir, Y 2012, ‘Automatic classification of power quality events

and disturbances using wavelet transform and support vector machines’,
80 IET

Gener, Transm, Distrib, vol. 6, pp. 968-976.

Thank you

81