Power Quality Disturbance Detection and Monitoring of Solar
Integrated Micro-Grid
DEBASISH PATTANAIK1, SARAT CHANDRA SWAIN1, INDU SEKHAR SAMANTA2,
RITESH DASH3 , KUNJABIHARI SWAIN4
1School of Electrical Engineering, KIIT Deemed to be University, Bhubaneswar, INDIA
2 Department of Electrical and Electronics Engineering, Siksha ‘O’ Anusandhan University, INDIA
3School of Electrical & Electronics Engineering, REVA University, Bangalore, INDIA
4Department of Electrical and Electronics Engineering, NIST Institue of Science and Technology,
Berhampur, INDIA
Abstract: - Due to the popularity of microgrids and power quality disturbances (PQD) induced by renewable
energies, monitoring in microgrids has risen in popularity in recent years. For monitoring the PQD, many
strategies based on artificial intelligence have been proposed. However, when the electrical parameters change,
the need to retrain the Artificial neural network (ANN) becomes a significant issue. This paper presents a new
approach to the power quality disturbance detection and monitoring of integrated solar microgrids. The power
quality event detection is accomplished by analyzing the frequency signal with Wavelet transformation (WT).
The classification of power quality disturbance is achieved based on the features. For the classification of
PQDs, the retrieved features are fed into a Convolutional neural network (CNN) classifier.
Keywords: - Power Quality Disturbances, Artificial Neural Network, Wavelet Transformation, Convolutional
Neural Network, Micro Grid, Renewable Energy Sources, GoogLeNet neural network.
Received: July 8, 2021. Revised: July 21, 2022. Accepted: September 19, 2022. Published: October 6, 2022.
1 Introduction
The present advancement in technology necessitates
an uninterrupted power supply. The uninterrupted
power supply impedes because of the limitation of
the grid due to numerous reasons. Micro-grid with
natural resources integration plays a vital role in
accomplishing this need. The integration of
renewable resources causes the power quality issue
to disturbance and instability. Due to power
electronics devices and inverters connected to
renewable energy (RE) sources, the grid
incorporation of large-scale RE sources, particularly
solar and wind energy, induces voltage and current
harmonics. One of the major issues today is
ensuring acceptable harmonics in the line currents
of RE integrated power systems [1]. As the
penetration capacity of the PV system increases, a
greater level of harmonic distortion is injected into
the grid. Therefore the PV system should only be
integrated up to the network's full capacity. When a
PV system is integrated beyond this maximum
penetration level, it produces considerable harmonic
distortion, which has a negative impact on the
system's performance [2]. As the use of renewable
energy grows, it has an adverse effect on the
distribution system, generating overvoltage, voltage
fluctuations, and reverse power flow to the grid.
Maintaining the voltage and current sinusoidal
waves at the rated frequency and magnitude is
referred to as power quality. Any departure leads to
a loss of power system efficiency, jeopardizing the
power system's economy by putting undue strain on
consumers and suppliers. Practical obstacles such as
voltage control, flicker, harmonic distortion,
stability, and other power quality issues occur when
wind and solar energy are integrated into existing
power systems [3]. Circuit breakers, switches,
converters, and non-linear loads are all being used
more and more these days. The electrical network's
power quality disruptions (PQ) are a common cause.
To avoid the network and its connected equipment
from being disrupted, it is important to identify and
classify the various PQ disturbances (single and
mixed) [4]. This compels the microgrid's rapid and
effective power quality detection and classification.
Monitoring the power quality disturbance in order to
take corrective action has become a hot topic,
particularly in renewable energy source integrated
micro-grid systems. The principal source of PQDs
in today's distribution networks is rapid
industrialization, the use of sensitive electrical
equipment on a large scale, massive non-linear
loads, and significant usage of power electronics
devices. Due to the move from conventional
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2022.17.31
Debasish Pattanaik, Sarat Chandra Swain,
Indu Sekhar Samanta, Ritesh Dash, Kunjabihari Swain
E-ISSN: 2224-350X
306
Volume 17, 2022
distributed systems to smart distributed systems, the
integration of non-conventional based distributed
generators (DGs), energy storage systems (ESSs),
power electronics converters, and electric car
charging stations exacerbates this issue. PQ
degradation leads to voltage sag, swell, impulse,
oscillatory transients, and other operational
difficulties in real-time circumstances. Power
quality (PQ) issues in real-time systems can cause
various control and protection devices to
malfunction or fail, affecting the system's overall
performance [4]. The signals of these PQDs are
non-stationary and statistically time-varying in
nature. Signal processing techniques are widely
used to examine non-stationary signals in various
power system challenges, including fault
identification, islanding detection, differential
protection, and the detection of PQEs. Fourier
transforms (FT), discrete Fourier transform (DFT),
and fast Fourier transform (FFT) were used in the
early stages of signal processing applications for PQ
analysis (FFT), short-time Fourier transform
(STFT), Curvelet transform, Hilbert transform,
Empirical mode decomposition (EMD), and
variational mode decomposition (VMD) are gaining
popularity as a computational approach for
extracting spectra for stationary signals at various
frequencies [5-13]. The power quality classification
is equally essential in PQE monitoring. With
minimal human intervention, automatic
classification is critical in modern power quality
classification. There are several techniques have
been employed for the classification of PQEs. One
of the most effective ANN approaches is ELM,
frequently used to resolve classification difficulties
[14-16]. Several other techniques, such as
probabilistic neural networks and support vector
machines, gained popularity for PQE classification
[17-22]. Due to numerous advantages over other
renewable energy sources, solar energy proved to be
the optimum energy source among all existing
renewable resources. The PV systems are a gift to
modern society. Numerous power quality challenges
develop when connecting an extensive PV system
with the grid. Poor or insufficient power quality
could lead to financial losses and end-user
disturbance. The power system components become
overheated and behave unfavorably due to the low
power quality issue, resulting in substantial damage
[23].
An increasing issue of power quality disturbances
due to integration of renewable energy sources like
solar and wind to micro-grids may have greater
impact on the operation of end user equipments. For
mitigation of PQDs, an efficient monitoring system
is much needed. The monitoring system is mainly
consists of the process of detection and
classification of PQDs. Different signal processing
techniques like FFT, ST, WT, EMD, and VMD are
used for the feature extraction process and different
machine learning techniques like ELM, PNN, SVM,
DT, FL have been used for establishment of an
efficient monitoring system. However, a better
monitoring system is always a need for the
conditions of real-time, non-stationary, noisy,
robust, faster computation, and cost effectiveness.
Here, the hybrid approach of CWT and CNN shows
better results in many of the above mntioned
conditions.
Power quality is the measure of correctness of the
power signal without any deviation from the
specified range for amplitude, frequency, and phase.
Different power quality disturbances are voltage
sag, swell, interruption, flicker, transients, and
harmonics etc. There is a possibility of simultaneous
happening of multiple PQDs. The impact of these
PQDs on consumer electronic appliances, power
electronics and control instruments based on micro-
controllers is severe. To mitigate these disturbances,
there is an urgent need of detection, classification,
and monitoring of PQDs. Several research work
have been done on this issue and this study is an
alternative method to answer this issue. The
proposed hybrid method is a novel technique
consisting of CWT and CNN shows faster
computation and better accuracy in comparison to
contemporary methods.
The micro-grid uses the environment friendly
energy sources to reduce transmission loss, manage
the power supply and demand, improve the
operation and stability, and to provide dynamic
responsiveness. However, the use of power
electronics instruments and devices is one of the
major reason of PQDs. To mitigate the PQDs, a
robust monitoring system is required and for which
detection, classification, and monitoring of PQDs
are very much necessary. Here, the suggested
method is a meaningful approach for the above
processes to build a robust monitoring system for
mitigation of PQDs.
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2022.17.31
Debasish Pattanaik, Sarat Chandra Swain,
Indu Sekhar Samanta, Ritesh Dash, Kunjabihari Swain
E-ISSN: 2224-350X
307
Volume 17, 2022
This article presents the PQD event detection and
classification using wavelet transform (WT) and
Convolutional neural network (CNN). WT is
utilized to extract the prominent features, and CNN
is used to classify PQ events. The proposed
methodology is validated in a physical real-time PV
integrated microgrid.
From decades and over, research work is going on
the detection, classification, and monitoring for the
mitigation of power quality disturbances efficiently.
Here, in this study a hybrid approach consisting of
Wavelet transform and CNN is considered for the
above work due to the following reasons: (i) faster
computation (ii) dealing a large data set (iii) for
better classification accuracy (iv) to work on real-
time environment and noisy conditions.
The background theory of the WT is presented in
the next section. Section 3 describes CNN-based
event classification. An experimental setup is
presented in section 4. The result analysis is
presented in Section 5, and the conclusion is given
in the final section.
2 Back Ground Theory of Wavelet
Transform
The wavelet transform could provide both the
frequency and the time associated with signals,
making it extremely helpful in a variety of
applications. It gives an STFT generalization. Like
DFT and STFT in signal theory, wavelet
transformation can be thought of as the projection of
a signal into a series of essential functions called
wavelets. In the frequency domain, these basis
functions provide localization.
2.1 Wavelet Transform
The ability of Wavelet transform (WT) to analyze
the local discontinuities of the signals can be best
used for steady-state analysis and the analysis of
signals in various fields having non-stationary
characteristics. The power quality events (PQEs) of
the power system have non-stationary features;
hence the WT is preferred as a suitable tool to apply
for the PQEs detection. The continuous signal u(t)
in the continuous Wavelet transform (CWT) form
can be mathematically expressed with the wavelet
function as:
(1)
In Eq. (1) the scale and translation parameters are
represented by the constants a and b, respectively.
The oscillatory frequency and wavelet length are
provided by the scale parameter a. The shifting
position is well represented by b, the translation
parameter. Each scale has a series of wavelet
coefficients at each scale which are the output and
hence represent the comprehensive PQ signal. The
superfluous information in practical applications of
CWT makes it unsuitable for signal analysis. The
discrete wavelet transform (DWT) is found more
appropriate for analysis of the PQEs and can be
expressed as in Eq. (2):
(2)
where
and
Represent the scaling
parameter and the translation parameter,
respectively. The discrete point sequences
represented by are the discrete form of the
continuous-time signal u(t). Depending on the type
of data used in WT applications, the type of mother
wavelet selection has a significant role in analyzing
the signal. Among different types of mother
wavelets, the Daubechies mother wavelet at scale 4
(db4) has a substantial role in feature extraction and
is widely used for various applications.
3 Classification using CNN
CNN, a sub-class of artificial neural networks
currently being extensively prominent in various
computational vision processing, is increasing
interest among researchers across different research
areas comprising power systems. CNN is
automatically and adaptively adjusted to
hierarchical spatial features using backpropagation
techniques implementing multiple building blocks,
involving various stratified blocks as convolution
layers, pooling layers, and fully connected layers.
CNN is a sub-program of deep learning models for
the computation of information employing a grid
pattern as pictures, which can be correlated to the
topology of the animal pictorial cortex and
formulated to adjust to hierarchical spatial features
automatically and adaptively from low- to high-
level patterns. CNN is usually made up of three
types of layers such as convolution, pooling layer,
and fully connected layers. The feature extraction
process is involved in the first two layers of CNN
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2022.17.31
Debasish Pattanaik, Sarat Chandra Swain,
Indu Sekhar Samanta, Ritesh Dash, Kunjabihari Swain
E-ISSN: 2224-350X
308
Volume 17, 2022
topology, whereas in the third layer, the features are
mapped into the latest output. The first layer is a
significant block in CNN, and It is made up of a
series of mathematical processes involving
convolution, a type of linear operation technique. If
a square neuron layer of size is followed by a
convolutional layer, then the size of the output of
the convolutional layer will be
, where is the size of the filter
. The pre-nonlinearity input can be calculated as
, where
is a
unit of the layers and
is
the sum of contributions by weighted filter
components. In digital images, pixel data is
assimilated in a 2D grid, i.e., a number array (Fig.
2), and a parameter of a small grid referred to as
kernel, which is a feature extractor that can be
optimized accordingly to requirements, which is
used at every image position, which renders CNNs
as being significantly effective in image processing,
since a feature may arise at any stage in the image.
As one layer relays the output data into the next
layer, there can be a progressive complexity with
increasing hierarchy being observed in extracted
features. The process of optimizing parameters
involved in kernel data arrangement is called
training. It is the process of minimizing the
difference between outputs and ground truth labels
by employing optimization techniques such as
backpropagation and gradient descent, among
others.
Fig. 1: CNN architecture and its training process.
The performance characteristics of a learning model
under specific kernel models and weights are
computed using a loss function and a forward
propagation technique applied to a training dataset,
with parameters, such as kernels and weights, that
are updated based on the loss value via
backpropagation using the gradient descent
optimization algorithm ReLU, rectified linear unit.
3.1 Building Blocks of CNN Architecture
The CNN topology comprises several
building blocks such as convolutional, pooling, and
fully connected layers. A typical CNN
topology entails a stack of convolutional layer and
pooling layers that repeats, followed by one or more
fully connected layers. The phase in which the input
data is processed into the output data by these layers
is called forward-ing propagation. This section
describes the convolution and pooling procedures
configured for 2D-CNNs, but similar operations can
be applied for 3D-CNNs.
3.2 Training a Network
Network training includes calculating
kernel data for convolutional layers and
weights for fully connected layers to minimize
differences between training dataset output
predictions and specific ground truth labels.
Backpropagation algorithms are commonly used
in neural network training, where loss function
and gradient descent optimizati-
on techniques are essential.The model's perfor-
mance characteristics under a given kernel and
weights are estimated using the training dataset and
the forward propagation loss function of the
trainable parameters. The kernel and weights are
calculated and updated with loss values
by optimization techniques such as backpropa-
gation and gradient descent.
The available data is generally divided into three
sets: training, validation, and testing, but there
are several variations on the following sets, such as
cross-validation. It then trains the network using
the training set, calculates the loss value using the
forward propagation technique, and updates the
parameters that backpropagation can
learn. Validation sets are used to evaluate models,
fine-tune hyperparameters, and apply
model selection during the training phase. Ideally,
the test set should be used only in the last stage of
the project, and the training set and validation set
should be used to evaluate the performance of
the fini-shed model improved and selected during
the training phase.
Model training inevitably necessitates fine-tuning
hyperparameters and model selection tasks,
necessitating several validations and test sets.
This method is calculated depending on the
validation set's performance, so some info about this
validation set is reflected in the model. The model
is not directly trained on trainable parameters
but overfitting to the validation set. This ensures
that a model with hyperparameters fine-tuned in
a validation set will work appropriately in a similar
validation set. Therefore, control datasets should be
used to properly evaluate the performance of the
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2022.17.31
Debasish Pattanaik, Sarat Chandra Swain,
Indu Sekhar Samanta, Ritesh Dash, Kunjabihari Swain
E-ISSN: 2224-350X
309
Volume 17, 2022
model. The model performance of the new random
data is essential and requires a separate test set.
Fig. 2: Available data are typically split into three
sets: training, validation, and a test set
Training sets train the network, calculate loss
values with forwarding
propagation techniques, and update
trainable parameters with backpropagation
techniques. Validation sets are used to
monitor model performance during the ongoing
training phase, fine-tune hyperparameters,
and perform model selection. Ideally, the test
set should be used only once at the end of the
project to evaluate the performance of
the finished model with a finely tuned and selected
strategy after the training process with
the training set and validation set. The available data
is generally divided into three sets: training,
validation, and testing. In the training the training
process the training data and validation data are
used to train the network. According to the training
accuracy of the network, the selection and tuning of
the parameters is done by selecting the optimized
parameters of the network for enhanced
performance. So, in the last phase of the process, the
test set is used to find the accuracy of the CNN
network.
4 Experimental Setup
Figure 3 shows the complete experimental setup for
the proposed work. It consists of 550kwatt of
rooftop solar PV array and 2200 PV cells, each
250watt. A REFUsol three-phase inverter. A
changer to switch between grid and solar power in
case of insufficient solar power. Connection with
the main grid. Three-phase loads. Three single
phases step down transformer. National Instrument
(NI) USB data acquisition card (DAQ). A personal
computer with LabVIEW and MATLAB software.
Fig. 3: Experimental setup
Maximum power, voltage at maximum power,
open-circuit voltage, maximum current, and short
circuit current for each solar cell is 250Wp, 30.72V,
37.05V, 8.15A, and 8.58A, respectively. The solar
panels are connected in series and connected at the
input terminal of the three-phase inverter. And the
output of the inverter is connected to the loads
through the changer switch. The role of the changer
switch is to switch the load between inverter output
and grid connections. By default, the loads are
connected with inverter output utilizing solar power.
The solar power is less than the load requirement
due to clouds covering the panel or in the rainy
season. The changer will switch the loads to grid
mode. The three-phase voltage and current signals
are captured by NI USB DAQ and continuously
monitored. The NI USB DAQ is connected to the
computer to analyze the power quality event using
machine learning technique. NI LabVIEW is used to
take the voltage signal through the NI USB DAQ to
the computer. The machine learning algorithm is
implemented in MATLAB. The data transfer
between LabVIEW and MATLAB occurs using
TCP/IP protocol.
5 Result Analysis and Discussion
5.1 Time-Frequency Representations
The experiment was performed on
the PQE signals obtained from the MATLAB
simulation environment and the real-time
signal obtained from the experimental setup.
The time-frequency representation of a PQE signal
is called a scalogram and represents the absolute
value of the signal's CWT coefficient. Pre-
calculation of the CWT filter bank is required to
create the scalogram data. The re-calculation of the
CWT filter bank is preferabley selected to
to acquire the CWTs of multiple signals using the
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2022.17.31
Debasish Pattanaik, Sarat Chandra Swain,
Indu Sekhar Samanta, Ritesh Dash, Kunjabihari Swain
E-ISSN: 2224-350X
310
Volume 17, 2022
same parameters. The filter bank is used to get the
CWT of the first 1000 samples of the signal, and the
scalogram of the coefficients is taken.
(a)
(b)
(c)
Fig. 4: (a) Harmonic signal (b) Flicker signal (c)
Impulsive transient signal
Fig. 5: Scalogram
5.2 Division of Data Into Training and
Validation
The scalogram image data is loaded into the
program routine as saved image
data. Image data store allows programs to store
extensive image data, including data that cannot be
allocated in memory space, and efficiently
read batches of images during the ongoing training
phase of the CNN. The images are then
categorized and randomly divided into two
separate groups. One is for the training dataset, and
the other is for the validation dataset. 80% of the
images will be used for CNN training, and the
rest will be used for validation.
5.3 GoogLeNet
The pre-trained GoogLeNet neural network is
loaded. The layer graph is extracted and displayed
from the network.
Fig. 6: GoogLeNet Layer Graph
Evaluation of GoogLeNet Accuracy
The network is then evaluated using the network
data.
GoogLeNet Accuracy: 100%
Table 1. Initialization of input data normalization
(GoogLeNet)
*See the Table 1 at the Annex section
Fig. 7: Training progress (GoogLeNet)
It can be seen that the numbers obtained here are the
same as the validation accuracy observed in the
training visualization. Next, scale
grams were divided into training and validation
categories. Both category scalograms
were used in GoogLeNet training. The ideal way to
evaluate the results obtained after training is that the
network is responsible for classifying data that
has never been observed before. The
calculated validation accuracy is referred to here as
network accuracy because not enough data is
needed for training, validation, and testing
subroutines. The network accuracy is found out by
using the data which the network never seen before,
and it is done after training the network. The
network accuracy found out in this method can be
considered as the validation accuracy.
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2022.17.31
Debasish Pattanaik, Sarat Chandra Swain,
Indu Sekhar Samanta, Ritesh Dash, Kunjabihari Swain
E-ISSN: 2224-350X
311
Volume 17, 2022
5.4 SqueezeNet
SqueezeNet is a type of deep
CNN whose architecture supports image
sizes 227x227x3 pixels in size. GoogLeNet has
different image dimensions but doesn't need to
generate a new image in RGB format according to
SqueezeNet's dimensional specifications. It can
use the original RGB image.
Loading
The pre-trained Squeeze Net neural network is then
loaded.
Fig. 8: First Convolutional Layer Weights
5.5 Preparation of RGB Data for Squeeze
Net
RGB images have a size suitable for the GoogLeNet
architecture. Next, the extended image data is
created, and the existing RGB image resized for
the Squeeze-Net architecture data is automatically
saved.
The setting of Training Options and Training
Squeeze Net
Next, new training options will be created
for use with SqueezeNet. Then the
random seed value is set to the default value, and
the network is trained. The training
process typically takes 1-5 minutes on a well-
designed desktop CPU.
Next, the network's last layer is inspected to see
if the classification output layer contains three
classes.
During the training process 80% of the data (image)
are used for training the SqueezeNet. For the
reproduceable purpose the random seed value is set
to the default value. The neural network training
process is an iterative process which minimizes the
loss function. In iteration the used gradient descent
alogorithm evaluates the loss function and updates
its own weights.
Squeeze Net Accuracy: 93.75%
Table 2. Initialization of input data normalization
(Squeeze Net)
*See the Table 1 at the Annex section
Evaluation of Squeeze Net Accuracy
The network is then to be evaluated using
network data.
Fig. 9: Training Progress(Squeeze Net)
5.6 Discusion
The suggested method is a hybrid method of CWT
and CNN. The scalogram obtained from the wavelet
coefficients are fed to the GoogLeNet and
SqueezeNet for classification of PQDs. The
GoogLeNet and SqueezeNet are pretrained CNN,
which are trained with 1000 different kinds of
images. The classification accuracy for GoogLeNet
is 100% and for SqueezeNet is 93.75%. This
classification accuracy is found for a solar
integrated micro-grid in real-time environment. This
proposed method gives a higher classification
accuracy and having a faster computational time. A
comparison table is given for the comparison of the
suggested method with available present day
alternative methods. The limitations, suggested
improvements, and future scope of this study is
highlighted in the conclusion section.
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2022.17.31
Debasish Pattanaik, Sarat Chandra Swain,
Indu Sekhar Samanta, Ritesh Dash, Kunjabihari Swain
E-ISSN: 2224-350X
312
Volume 17, 2022
5.7 Comparison of Classification Accuracy
Table 3. Overall Accuracy Comparison Statistics
with other Conventional Methods
Classification Methods
Accuracy Percentage
DWT with ANN
DWT with NFS
WPT with MSVM
DWT with RBES
ST with DT
ST with RBES
ST with PNN
Proposed CWT with
CNN (GoogLeNet)
Proposed CWT with
CNN (SqueezeNet)
94.37
96.5
96.8
98.7
98.5
98.2
97.4
100
93.75
6 Conclusion
The above studies show the use of transfer learning
and continuous wavelet analysis to classify three
classes of PQE signals using pre-trained CNNs
i.e., GoogLeNet and SqueezeNet. The wavelet-
based time-frequency
representation of the PQE signal is used to create
scalograms. RGB images of scalograms are plotted
using a computer. The image is then processed
for fine-tuning of both deep
CNNs. Various network layer activations
were also investigated. The above study also shows
a workflow that can be used to classify signals using
a pre-trained CNN model. The above studies also
show the efficiency of a hybrid model based on
CWT and CNN for detecting power
quality events in a solar-integrated microgrid
environment with 100% and 93.75% accuracy
of GoogLeNet and SqueezeNet, respectively. Here,
GoogLeNet and SqueezeNet are two deep CNNs
pretrained to recognize the images for classification
of PQD signals based on time-frequency
representation. This network architecture of CNN is
reused for the classification of PQD signals based
on images obtained from CWT of the time series
data.
The limitation of this study is that its SqueezeNet
accuracy is only 93.75%. Further research can be
done on the use of efficient feature extraction
techniques and parameter optimization of CNN to
have better efficiency and lesser computational time
in a real-time environment. The detection and
classification process of PQDs to be done by deep
learning alone may be the future direction of
research of this study.
References:
[1] G. M. Shafiullah and A. M. T. Oo, “Analysis
of harmonics with renewable energy
integration into the distribution network,”
Proc. 2015 IEEE Innov. Smart Grid Technol. -
Asia, ISGT ASIA 2015, 2016, doi:
10.1109/ISGT-Asia.2015.7387191.
[2] E. A. Sharew, H. A. Kefale, and Y. G. Werkie,
“Power Quality and Performance Analysis of
Grid-Connected Solar PV System Based on
Recent Grid Integration Requirements,” Int. J.
Photoenergy, vol. 2021, 2021, doi:
10.1155/2021/4281768.
[3] A. S. P. and S. K. S. Varun Kumar1, “Grid
Integration and Power Quality Issues of Wind
and Solar Energy System: A Review,” in
International Conference on Emerging Trends
in Electrical, Electronics and Sustainable
Energy Systems (ICETEESES–16), 2013, vol.
316–317, pp. 345–348. doi:
10.4028/www.scientific.net/AMM.316-
317.345.
[4] S. Khokhar, A. A. B. Mohd Zin, A. S. B.
Mokhtar, and M. Pesaran, “A comprehensive
overview on signal processing and artificial
intelligence techniques applications in
classification of power quality disturbances,”
Renew. Sustain. Energy Rev., vol. 51, pp.
1650–1663, 2015, doi:
10.1016/j.rser.2015.07.068.
[5] M. Szmajda, K. Górecki, and J. Mroczka,
“DFT algorithm analysis in low-cost power
quality measurement systems based on a DSP
processor,” 2007 9th Int. Conf. Electr. Power
Qual. Util. EPQU, 2007, doi:
10.1109/EPQU.2007.4424081.
[6] G. T. Heydt, P. S. Fjeld, C. C. Liu, D. Pierce,
L. Tu, and G. Hensley, “Applications of the
windowed FFT to electric power quality
assessment,” IEEE Trans. Power Deliv., vol.
14, no. 4, pp. 1411–1416, 1999, doi:
10.1109/61.796235.
[7] F. Jurado and J. R. Saenz, “Comparison
between discrete STFT and wavelets for the
analysis of power quality events,” Electr.
Power Syst. Res., vol. 62, no. 3, pp. 183–190,
2002, doi: 10.1016/S0378-7796(02)00035-4.
[8] P. S. Wright, “Short-time fourier transforms
and Wigner-Ville distributions applied to the
calibration of power frequency harmonic
analyzers,” IEEE Trans. Instrum. Meas., vol.
48, no. 2, pp. 475–478, 1999, doi:
10.1109/19.769633.
[9] I. S. Samanta, P. K. Rout, S. Mishra, K. Swain,
and M. Cherukuri, “Fast TT transform and
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2022.17.31
Debasish Pattanaik, Sarat Chandra Swain,
Indu Sekhar Samanta, Ritesh Dash, Kunjabihari Swain
E-ISSN: 2224-350X
313
Volume 17, 2022
optimized probabilistic neural network-based
power quality event detection and
classification,” Electr. Eng., 2022, doi:
10.1007/s00202-022-01505-8.
[10] I. S. Samanta, P. K. Rout, and S. Mishra,
“Feature extraction and power quality event
classification using Curvelet transform and
optimized extreme learning machine,” Electr.
Eng., vol. 103, no. 5, pp. 2431–2446, 2021,
doi: 10.1007/s00202-021-01243-3.
[11]Y. Lei, J. Lin, Z. He, and M. J. Zuo, “A review
on empirical mode decomposition in fault
diagnosis of rotating machinery,” Mech. Syst.
Signal Process., vol. 35, no. 1–2, pp. 108–126,
2013, doi: 10.1016/j.ymssp.2012.09.015.
[12] Samanta, I.S., Rout, P.K., Mishra, S. et al. Fast
TT transform and optimized probabilistic
neural network-based power quality event
detection and classification. Electr Eng (2022).
https://doi.org/10.1007/s00202-022-01505-8
[13] M. Sahani, P. K. Dash, and D. Samal, “A real-
time power quality events recognition using
variational mode decomposition and online-
sequential extreme learning machine,” Meas. J.
Int. Meas. Confed., vol. 157, p. 107597, 2020,
doi: 10.1016/j.measurement.2020.107597.
[14] I. S. Samanta, P. K. Rout, K. Swain, M.
Cherukuri, and S. Mishra, “Power quality
events recognition using enhanced empirical
mode decomposition and optimized extreme
learning machine,” Comput. Electr. Eng., vol.
100, no. March, p. 107926, 2022, doi:
10.1016/j.compeleceng.2022.107926.
[15] I. S. Samanta, P. K. Rout, and S. Mishra, “An
optimal extreme learning-based classification
method for power quality events using
fractional Fourier transform,” Neural Comput.
Appl., vol. 33, no. 10, pp. 4979–4995, 2021,
doi: 10.1007/s00521-020-05282-y.
[16] I. S. Samanta, P. K. Rout, and S. Mishra,
“Power Quality Events Recognition Using S-
Transform and Wild Goat Optimization-Based
Extreme Learning Machine,” Arab. J. Sci.
Eng., vol. 45, no. 3, pp. 1855–1870, 2020, doi:
10.1007/s13369-019-04289-5.
[17] K. Thirumala, A. C. Umarikar, and T. Jain, “A
new classification model based on SVM for
single and combined power quality
disturbances,” 2016 Natl. Power Syst. Conf.
NPSC 2016, 2017, doi:
10.1109/NPSC.2016.7858889.
[18] P. Thanthirige et al., “Classification if Power
Quality Events Using Support Vector Machine
and S- Transform,” in 2016 2nd International
Conference on Contemporary Computing and
Informatics (ic3i), 2016, vol. 7, no. August, pp.
279–284.
[19] Ç. Arikan and M. Özdemir, “Classification of
power quality disturbances at power system
frequency and out of power system frequency
using support vector machines,” Prz.
Elektrotechniczny, vol. 89, no. 1A, pp. 284–
291, 2013.
[20] S. Mishra, “Detection and Classification of
Power Quality Disturbances Using S-
Transform and Probabilistic Neural Network,”
J. Ind. Inf. Integr., vol. 22, p. 4244, 2021, doi:
10.1016/j.jii.2021.100204.
[21] R. Sharma and L. Srivastava, “Power quality
disturbance prediction using PNN,” 2018 2nd
IEEE Int. Conf. Power Electron. Intell. Control
Energy Syst. ICPEICES 2018, pp. 299–304,
2018, doi: 10.1109/ICPEICES.2018.8897482.
[22] A. Aggarwal and M. K. Saini, “Designed
orthogonal wavelet based feature extraction
and classification of underlying causes of
power quality disturbance using probabilistic
neural network,” Aust. J. Electr. Electron.
Eng., vol. 18, no. 3, pp. 161–171, 2021, doi:
10.1080/1448837X.2021.1948166.
[23] A. Khandelwal and P. Neema, “State of Art for
Power Quality Issues in PV Grid Connected
System,” in 2019 International Conference on
Nascent Technologies in Engineering (ICNTE),
2019, no. Icnte, pp. 1–4.
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
Conceptualization, Indu Sekhar Samanta,
Kunjabihari Swain, Debasish Pattnaik;
methodology, Debasish Pattnaik, Ritesh Dash and
Indu Sekhar Samanta; software, Indu Sekhar
Samanta and Debasish Pattnaik; validation, Indu
Sekhar Samanta, Ritesh Dash and Kunjabihari
Swain; investigation, Sarat Chandra Swain and Indu
Sekhar Samanta; resources, Sarat Chandra Swain
and Debasish Pattnaik ; data curation, Debasish
Pattnaik , Indu Sekhar Samanta and Kunjabihari
Swain; writing—original draft preparation,Debasish
Pattnaik, Indu Sekhar Samanta and Kunjabihari
Swain, supervision, Sarat Chandra Swain; formal
analysis, Sarat Chandra Swain; Visualization, Sarat
Chandra Swain; figure and table, Kunjabihari
Swain and Indusekhar Samanta. All authors have
read and agreed to the published version of the
manuscript.
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2022.17.31
Debasish Pattanaik, Sarat Chandra Swain,
Indu Sekhar Samanta, Ritesh Dash, Kunjabihari Swain
E-ISSN: 2224-350X
314
Volume 17, 2022
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
Annex
Table 1. Initialization of input data normalization (GoogLeNet)
Epoch
Iteration
Time Elapsed
(hh:mm:ss)
Mini-batch
Accuracy
Validation
Accuracy
Mini-batch
Loss
Validation Loss
1
1
00:00:21
33.33%
46.88%
1.7424
1.2343
2
10
00:01:27
46.67%
59.38%
1.3216
1.0275
3
20
00:02:38
40.00%
65.62%
1.0883
0.7985
4
30
00:03:45
60.00%
75.00%
0.6761
0.5229
5
40
00:04:51
86.67%
78.12%
0.2971
0.4800
7
50
00:05:57
86.67%
71.88%
0.3501
0.4566
8
60
00:07:05
80.00%
81.25%
0.2811
0.3469
9
70
00:08:10
100.00%
84.38%
0.2021
0.3480
10
80
00:09:13
100.00%
96.88%
0.1737
0.2811
12
90
00:10:10
93.33%
96.88%
0.3201
0.2464
13
100
00:10:58
93.33%
84.38%
0.1404
0.2614
14
110
00:11:47
86.67%
96.88%
0.2538
0.1880
15
120
00:12:35
93.33%
96.88%
0.2127
0.1942
17
130
00:13:22
100.00%
96.88%
0.0712
0.1986
18
140
00:14:10
93.33%
96.88%
0.1597
0.2011
19
150
00:14:57
100.00%
84.38%
0.0674
0.2103
20
160
00:15:45
100.00%
100.00%
0.0546
0.1405
Here the base learning rate in all cases of input data normalization is taken to be 1.0000e-04
Table 2. Initialization of input data normalization (Squeeze Net)
Epoch
number
Itera-
tion
Time Elapsed
(hh:mm:ss)
Mini-batch
Accuracy
Validation
Accuracy
Mini-batch
Loss
Validation
Loss
t
1
00:00:07
20.00%
53.12%
4.2032
1.2722
1
13
00:00:27
60.00%
62.50%
0.9195
0.9299
2
26
00:00:48
60.00%
59.38%
0.7726
0.8316
3
39
00:01:09
60.00%
62.50%
0.7032
0.7386
4
50
00:01:26
90.00%
0.7458
4
52
00:01:30
70.00%
87.50%
0.5993
0.6630
5
65
00:01:51
90.00%
84.38%
0.4828
0.5314
6
78
00:02:12
90.00%
87.50%
0.3731
0.3828
7
91
00:02:33
90.00%
84.38%
0.2220
0.3695
8
100
00:02:47
90.00%
0.3257
8|
104
00:02:54
90.00%
90.62%
0.2096
0.3013
9
117
00:03:15
100.00%
90.62%
0.0694
0.2194
10
130
00:03:36
90.00%
90.62%
0.2280
0.1890
11
143
00:03:57
100.00%
90.62%
0.0280
0.1959
12
150
00:04:07
100.00%
0.1082
12
156
00:04:18
90.00%
90.62%
0.1536
0.3638
13
169
00:04:39
80.00%
81.25%
0.7354
0.7383
14
182
00:04:59
100.00%
90.62%
0.0744
0.2016
15
195
00:05:20
100.00%
93.75%
0.0211
0.1672
Here the base learning rate in all the cases is taken to be of the value of 0.0003
WSEAS TRANSACTIONS on POWER SYSTEMS
DOI: 10.37394/232016.2022.17.31
Debasish Pattanaik, Sarat Chandra Swain,
Indu Sekhar Samanta, Ritesh Dash, Kunjabihari Swain
E-ISSN: 2224-350X
315
Volume 17, 2022
