Received: 1 June 2020 Revised: 29 October 2020 Accepted: 1 November 2020

DOI: 10.1002/2050-7038.12706

SPECIAL ISSUE PAPER

Comparative analysis of machine learning algorithms for

prediction of smart grid stability†

Ali Kashif Bashir1 | Suleman Khan2 | B Prabadevi3 | N Deepa3 |

Waleed S. Alnumay4 | Thippa Reddy Gadekallu3 |
Praveen Kumar Reddy Maddikunta3

1
Department of Computing and
Mathematics, Manchester Metropolitan
Summary
University Manchester, Manchester, UK The global demand for electricity has visualized high growth with the rapid
2
Air University Islamabad, Islamabad, growth in population and economy. It thus becomes necessary to efficiently
Pakistan
distribute electricity to households and industries in order to reduce power
3
School of Information Technology and
Engineering, Vellore Institute of
loss. Smart Grids (SG) have the potential to reduce such power losses during
Technology, Tamilnadu, India power distribution. Machine learning and artificial intelligence techniques
4
Computer Science Department, King have been successfully implemented on SGs to achieve enhanced accuracy in
Saud University, Riyadh, Saudi Arabia
customer demand prediction. There exists a dire need to analyze and evaluate
Correspondence the various machine learning algorithms, thereby identify the most suitable
Thippa Reddy Gadekallu and Praveen one to be applied to SGs. In the present work, several state-of-the-art machine
Kumar Reddy Maddikunta, School of
learning algorithms, namely Support Vector Machines (SVM), K-Nearest
Information Technology and Engineering,
Vellore Institute of Technology, Neighbor (KNN), Logistic Regression, Naive Bayes, Neural Networks, and
Tamilnadu, India. Decision Tree classifier, have been deployed for predicting the stability of the
Email: [email protected] (T. R. G.)
and
SG. The SG dataset used in the study is publicly available collected from UC
Email: [email protected] Irvine (UCI) machine learning repository. The experimentation results
(P. K. R. M.) highlighted the superiority of the Decision Tree classification algorithm, which
Funding information outperformed the other state of the art algorithms yielding 100% precision,
King Saud University, Grant/Award 99.9% recall, 100% F1 score, and 99.96% accuracy.
Number: RSP-2020/250
KEYWORDS
machine learning algorithms, smart grid, stability

1 | INTRODUCTION

The immense growth in the global population and economy, along with the rapid surge in urbanization has high possi-
bilities to increase the demand in energy consumption in the succeeding future years. Electricity being an important
source of energy, can be produced from various sources such as water, wind, solar cells, fossil fuels, thermal, and

List of symbols and abbreviations: AUC, Area Under the Curve, SG, Smart Grid; SVM, Support Vector Machine; KNN, K-Nearest Neighbor; UC,
UC Irvine; ICT, Information and Communication Technologies; AI, Artificial Intelligence; AAN, Artificial Neural Network; LSTM, Long short term
memory; RNN, Recurrent Neural Network; FFNN, Feed-Forward Neural Network; MLSTM, Multi-Directional LSTM; NB, Naive Bayes; DT, Decision
Tree; CM, Confusion Matrix, CR, Classification Report, ROC, Receiver Operating Characteristic.
†
Comparative Analysis of Machine Learning Algorithms for classification of Smart Grid dataset.

Int Trans Electr Energ Syst. 2021;e12706. wileyonlinelibrary.com/journal/etep © 2021 John Wiley & Sons Ltd 1 of 23
https://doi.org/10.1002/2050-7038.12706
2 of 23 BASHIR ET AL.

nuclear plants. Also, with the advancement and extensive growth of our population, the demand for electricity is ever
increasing, which automatically impacts the demand for higher electricity production. Power generation, transmission,
and its distribution are the most critical issues involved in electricity management. It is a known fact that electric grid
is an interconnected network that connects the consumers to the producers of electricity and transfers the energy from
producer to consumer. It comprises the power stations generating electricity, substations for stabilizing the electricity
voltage based on use, transmission lines (the carrier of electricity), and the distribution lines which connect the cus-
tomers. The conventional electrical grids adopt a centralized structure with millions of the aforementioned compo-
nents. Increasing the load of an electric grid thus create possibilities of generating additional overhead resulting in
power quality issues. Therefore, the need for installation of new plants emerge. On the other side, these grids do not
have proper prediction system in order to predict intermittent power outages, their causes, response delay, storage
requirements and resource utilization.1
Researchers have identified that the current electrical power system has not experienced any change for the past
hundred years. It is obvious that with the increase in population, there exists a high demand for electricity. The chal-
lenges of the conventional power system are lack of visibility, usage of mechanical switches, which result in reduced
response time, deficiency of monitoring, and power control. The additional inducing factors that need a new grid tech-
nology are change in climatic conditions, need for energy, failure of components, increase in population, lack of storage
for energy, demand for fossil fuels, decrease in electric power generation, unilateral communication, and various other
problems. Thus, in order to overcome such challenges, new grid infrastructure is required. The next-generation electric
power infrastructure, namely the smart grid (SG) emerges as a prominent technology to fulfill such high prioritized
necessities to enhance the quality of modern human life.2
Smart Grid is a new digital electric power grid technology that allows 2-way communication to increase security,3-5
efficiency, reliability of the electric power systems for higher generation of electrical power through contemporary com-
munication technologies. It is a two-way energy delivery and transportation system, which allows its consumers to
make decisions pertinent to energy. SGs basically helps to reduce the electricity bill paid by customers. It also contrib-
utes to an increase in security measures taken during the consequences of natural disasters and other human attacks.
On the contrary, it also ensures a significant decrease in risks resulting in loss of human lives and other physical infra-
structure relevant to conventional grid-related activities. Considering the implementation aspects, SGs assist in integrat-
ing electric vehicles and modernizing the transport division. In the realm of global warming issues and the need for
optimal energy utilization, smart grids help to reduce the energy wastage and also environmental pollution caused by
the discharge of gases from the greenhouse.6,7
From the perspective of comparative analysis, the conventional grid provides one-way communication, which is
limited to energy users, whereas SG provides massive two-way communication. Power quality issues are solved very
slowly in the conventional grid, whereas a rapid self-healing facility is available in the case of SG. The traditional grid
system is more prone to cyber-attacks and natural disasters with a much slower response. The SG, on the contrary, is
much durable during natural calamities and cyber attacks. The conventional grid system responds slowly to system dis-
turbances, whereas the SG provides automatic detection and response to the problems and has a much lesser impact on
customers. In the conventional grid system, power flow control is quite restricted, whereas much vast in SG.
In SG, various components are integrated with sensor nodes8 and communication paths to provide inter-operability
in business, manufacturing, and residential applications. The objective is to avoid power turbulence caused by compo-
nent failures, natural disasters, and capacity constraints by providing online intelligent electric power monitoring and
control system. The SG offers state-of-the-art services with two-way communication, an intelligent system, automatic
monitoring, and self-remedial skills. SG also renders support in demand management by predicting energy usage. The
usage of electricity can act as criteria for providing incentives to the consumers by revising their utilization pattern
thereby, efficiency can be improved. The aforementioned can be achieved by distributing energy to customers with
improved reliability and security features.9
As a huge amount of data from various applications need to be analyzed and controlled, the communication
requirement plays a significant role in SG infrastructure. Therefore it is very critical to identify the best communication
infrastructure to provide cost-effective, reliable, and secure service for the entire system. Numerous technologies have
been found in the case of SG communication with two media, namely wired and wireless, which can be applied for data
communication between smart meters and other electrical components. Wireless communication technologies are low
cost, and connection establishment is comfortable even in unreachable and problematic areas. But the signal may get
weakened due to the nature of the transmission lane, wherein wireless solutions depend on batteries. On the contrary,
wired technologies do not have any intervention problem, as they do not depend on batteries.
BASHIR ET AL. 3 of 23

In the SG infrastructure, two types of information flows are involved. The first information flow is from electrical
applications and sensors to smart meters. The second flow is from smart meters and the SG utility data centers. Wireless
communication technologies such as 6LowPAN, Z-wave, ZigBee, can be utilized for the first information flow, and
internet or cellular technologies can be applied for the second type of information flow. The same communication tech-
nology may not be suited for various other applications. Based on the application domain, the choice of communication
technologies should thus be made.2
Smart Grid provides smart solutions to almost all spheres and activities concerned with electricity.10-12 It offers the
following features: real-time monitoring of electricity consumption as per the type of application, dynamic pricing (on-
demand pricing), faster and effective restoration of electricity after the power outage, in-house electrical displays, alter-
ing the electricity usage during day time based on the pricing signals (offering an electricity incentive to consumer) and
usage levels, transferring of the role of consumer to producer (both consume and produce power), tracking the electric-
ity usage through the use of web apps and mobile apps.
SGs ranging from small sizes to large ones have been deployed in most of the developed countries.13 As an example,
a small-sized grid network was used at Gazi University, Turkey, which connected the wind, solar, battery storage and
diesel-powered microgrid systems. Similarly, a large-sized SG network was constructed at Jeju island, South Korea. All
these SGs incorporated the features that have been discussed earlier. It is important to mention the fact that cloud com-
puting technologies have extensive applications in SGs. Smart Grid uses Information and Communication Technologies
(ICT) to enable communication among the various resources. SG involves enormous data accumulation during electric-
ity generation, transmission, and distribution. Cloud computing14,15 can thus be utilized in energy management, secu-
rity services and information management in SGs.16 Zigbee technology can be utilized in an SG for monitoring and
controlling information, fault locating and transmission lines monitoring.17
The Internet of Things18 play an important role in the deployment of energy meters. In connection to this, SGs can
be used in smart home automation, smart building automation, smart city automation, smart substation, and feeder
automation.9
Although SG encompasses various technologies to resolve the issues with classical electricity networks, it has
associated problems that need to be addressed. There are basically two categories, technical and socio-economic
issues.19,20 The technical challenges include lack of policies, storage concern, cybersecurity vulnerabilities while con-
necting the grid to cyber-physical systems, inadequacy in grid infrastructure to accommodate the future needs and
demands in the storage of intermittent power generation, voluminous data management from different components
of the grid, grid stability concerned with power-sharing, system inertia, power oscillation and power reservation.
Technical challenges, on the other hand, include energy management in using electric vehicle involving power flow
from vehicle to grid, grid to vehicle and vehicle to vehicle. Some of the socio-economic challenges include stake-
holder management, lack of awareness, lack of policies and substantial capital investments. This also creates addi-
tional issues relevant to electricity charges, new tariff, health issues related to radiofrequency usage, privacy, fear of
obsolescence and power theft. Security remains to be one of the primary concerns as the SG can be breached by both
wired and wireless communication networks. Artificial intelligence and its subset machine learning algorithms21-23
can be employed in predicting the problems in SG that aids in taking precautionary steps. In this work, the most
prominent technical challenge, predicting the stability of an SG, is considered because it determines the reliable
power transmission in almost 50% of the SGs.24 Figure 1 depicts the SG environment incorporated with artificial
intelligence (AI) technology.
The main contributions of this work include:

• A detailed review of the existing literature on application of ML algorithms on SG is presented.

• A thorough investigation on performance of several ML algorithms for predicting the stability of SG.
• Comparative analysis of the performance of ML algorithms implemented on SG dataset against the recent work on
Deep Learning based model.25 The results highlighted the superiority of the ML algorithms, which outperformed the
deep learning based model considering the size of the dataset being very less.

The rest of the article is organized as follows: A thorough review of the existing works on application of ML algo-
rithms on SG is presented in Section 2. Section 3 presents the proposed model. The performance analysis of the ML
algorithms is presented in Section 4. The conclusion and future directions are presented in Section 5.
4 of 23 BASHIR ET AL.

FIGURE 1 SG environment incorporated with artificial intelligence (AI) technology

2 | LITERATURE S URVEY

A survey on various research solutions adopted methodologies, outcomes, and limitations of existing works on smart
grids are presented in this section. The SG, which replaces the conventional electricity grid, promises to carry out a
2-way communication. A complex system is used with the help of electric vehicles to distribute the power disseminated
from heterogeneous sources.26 This adds additional overhead to the SG modeling, controlling of its components in
order to optimize its performance. Therefore it requires consistent monitoring of stability robustness, efficiency, and
reliability in different operating conditions. Various researchers have used machine learning algorithms like support
vector machine (SVM), linear regression, K-nearest neighbors (KNN), Ridge Regression, Artificial neural network
(ANN), random forest, stochastic gradient descent, gradient boosting, Extra trees Regressor to predict the load in
SG. Deep learning models like Long short term memory (LSTM) neural network, recurrent neural network (RNN),
feed-forward neural network (FFNN), Backpropagation neural network, Multi-Layered Perceptron (MLP) are compared
with machine learning models in these studies. Feature selection algorithms are applied for selecting the best input
values, and hyperparameter tuning is performed to avoid the overfitting of data in machine learning models.
BASHIR ET AL. 5 of 23

Bouktif et al. have used the LSTM-based RNN model for predicting the electric load to handle demand related
aspects in SGs.27 They have applied best-of-breed machine learning algorithms to choose a benchmarking model and a
wrapper for feature selection (namely Extra tree regressor and regressive feature elimination) to select the best features
as input to the model. The genetic algorithm was used to find the layers and optimal time lags in the LSTM model. They
have compared their results with machine learning algorithms and found that LSTM-RNN has less forecasting error
than others. But this may not perform well in varying data sets as deep learning selects only good patterns from huge
data in different time series and more training data set. Therefore this model has to be evaluated with different
datasets.
SGs are transiting towards demand-based power supply services to the consumers. Therefore there is a mandatory
need to predict the consumer load. An attempt is made to identify whether the existing short term load
forecasting(STLF) model provides better accuracy or the anthropologic-structural data provides better accuracy in fore-
casting individual consumer household load.28 To perceive the best forecasting model for individual consumer load, a
short term multiple load forecasting (STMLF) model was proposed based on anthropologic structural data of the con-
sumer's house. SLMTF has the potential to forecast multiple loads at different time series using a single model. The
study used backpropagation neural network and SVM to compute predictions on the STMTF model. The results
highlighted the fact that STMLF is 7% more accurate than SLTF and also reduced 50% of the error. The study justified
the enhanced accuracy in the case of anthropologic structural data in comparison to the SLTF model. ANN was trained
with three weeks of data in the household, which predicted the power load per hour for the next day.
Normally SLTF is implemented on nation wise or region wise data. Hernandez et al., proposed an ANN-based
approach for SLTF within microgrids,29 which consisted of three stages: Pattern recognition using Self Organizing Map
(SOM), K-means clustering for partition and MLP for demand forecasting in the individual cluster. The ANN model
was validated using real-time data from a Spanish company. The model was trained with periodic values (weekday and
months). This model was compared with similar models using radial basis function neural network and generalized
regression neural network and proved to be extremely efficient.
The stability of the SG depends on its ability to provide a constant power supply based on demand. Ahmed and
Chen have employed three different machine learning models to predict the long-term and medium-term energy
demand in the SG.30 The models used are ANN with non-linear autoregressive exogenous multivariate inputs model
(ANN_NAEMI), Ada Boost and multivariate linear regression (MLR). The study classified the load considering three
different intervals, namely 1-month ahead, seasonal forecast and 1-year ahead forecast based on the data about aggre-
gated data consumption. The models not only increased the prediction accuracy but also adequately described the
Spatio-temporal use of energy inconsistencies, its variations and future perspective of energy prediction. Ada Boost
model, through its preponderance in prediction, outperformed the other models. The variations in the prediction results
helped to identify the irregularities in prediction operation.
Instability in electricity prices is one of the socio-economic factors determining the usage of electricity by the consumer
and often price impacted the electricity load. Therefore, Shayeghi et al. presented a multi-input multi-output(MIMO)
model that correlated the relation between electricity price and its load.31 The model used the wavelet packet, which was
transformed for decomposing the price and load signals into numerous subsets at different frequencies, and generalized
mutual information with an objective to select the best features. Also, the model made predictions on electricity load and
prices simultaneously using the least-squares SVM (LSSVM) based MIMO model. Furthermore, the Quasi Oppositional
Ant Bee Colony algorithm was used for parameter optimization. The simulation results showed that the proposed LSSVM
model outperformed ANN as it considered the prediction indices for evaluating the forecasting error.
Khan et al. presented a detailed study of dynamic pricing, load prediction in the SG.32 The study highlighted the cor-
relation among real-time dynamic pricing of electricity, critical peak pricing and time of use. Two different ways of load
forecasting, namely Artificial intelligence (AI) models and computational models, were presented. AI models used
ANN, Generalized Regression Neural Network(GRNN), RNN, Auto-Regressive Integrated Moving Average (ARIMA)-
SVM, SVM, Wavelet Transformation Error Correction(WTEC)-ANN, Wavelet Transformation(WT)-ANN, Probabilistic
Neural Network(PNN), Expert System and Fuzzy logic. It was evident from the survey that AI-based forecasting tech-
niques were found to be more accurate than other statistical models.
Muhammad et al. conducted a survey on Photovoltaic (PV) output forecast.33 Although the majority of the
researchers have attempted to forecast PV output using traditional methods, mathematical models and AI methods, this
particular study identified ANN to be capable of generating more accurate forecasting when compared to other conven-
tional and statistical models. The study also revealed the fact that the accuracy of any prediction techniques changed
based on the day, seasonal variation, input features and other evaluation matrices.
6 of 23 BASHIR ET AL.

Muhammad and Abbas conducted a survey on AI-based load forecasting models in the SG.34 The study highlighted
that the performance of the forecast model depended on its architecture, input features, activation function, ML algo-
rithms, which was used for training and generating of forecast errors. It was observed that Back-propagation
(BP) training algorithm was commonly used to train NN, but it had numerous associated challenges whereas ANN was
best suited for STLF yielding better performance than BP. It could be finally concluded that integrated approaches pro-
vided better results.
It is a known fact that Blackout events of the traditional power grid lead to many cascading failures. In order to
address this issue in SG, early warning of blackout events becomes a mandatory necessity. Gupta et al. applied the time
series model, namely SVM, to predict the blackout events earlier and validated using a 30-bus testbed from IEEE.35
SVM was trained using a historical database constructed by evaluating the system performance in a steady-state and
dynamic state. This database recorded the normal cases and abnormal situation (cascading failure condition).
Pan and Lee performed a comparative analysis of SVM and ANN in the midterm load forecast of the SG.36 ANN
was widely used for this type of forecast, whereas SVM was adopted by the researchers in recent times. The factors
affecting the load prediction, which was carried out for the daily power load in one year was analyzed. Mitchell et al.
used these models for STLF on different load types, namely batch load, continuous load and batch-continuous load.37
The results showed that SVM produced the global minimum repeatedly. Both the algorithms performed with extreme
inferiority with more than 3% deviation on erratic load and 1.2% deviation on a continuous load.
Electricity demand forecast depends on various factors such as climatic changes, seasonal changes, sea level and cat-
astrophic events by nature. Therefore, demand management in the SG determines its reliability and stability for provid-
ing consistent power demands of the consumer. Demand scheduling can be effectively derived by the effective forecast
of the consumers' electricity usage pattern. Ali and Azad have used machine learning algorithms, namely Linear regres-
sion, SVM and MLP for demand management and load prediction38 wherein it outperformed the other models. Support
vector regression (SVR) employed the constrained quadratic optimization problem, which mapped the input features
into high dimensional space using a kernel. SVR outperformed the NN trained with BP algorithm and other linear
regression methods. It also produced high-quality outputs with time series missing data. The study thus recommended
SVM for load forecasting.
Similar to demand management, SVM performed exceptionally better in other forecasts too. In the prediction of
lake water levels, SVM showed compatibility and best results (long-term forecast) when compared with ANN and Sta-
tistical model, namely the Seasonal autoregressive model in Reference 39. SVM was employed in the time series forecast
for financial analysis, which helped to overcome two common problems, namely noisy data and non-stationary.40 The
study used fuzzy-based Support vector regression basically to serve the purpose. Li et al. used PCA and SVM with rough
sets for long term electricity load forecast.41 C-ascending SVM in non-stationary financial time series provided better
results with fewer support vectors.42 Cao and Ju used dynamic SVM for non-stationary time series forecast.43 The study
modified the traditional SVM with a regularized risk function, which enabled the model to effectively track the struc-
tural change in financial time series. Also, SVM with rough sets outperformed the statistical models and GRNN.44 SVR
and chaotic GA simulated annealing algorithm which was integrated to improve the prediction accuracy of Chaotic
load.45 Also, SVM showed better accuracy in the case of MLP for wind speed forecasting.46 Alazab et al., developed a
multi-directional LSTM (MLSTM) model to predict the stability of the SG and the results shown that MLSTM out-
performed the traditional LSTM, Deep learning models gated recurrent units and RNN.25 An overview of the survey
describing the problems was addressed. The models used and inferences on their methodology is presented in Table 1.
From the literature survey, it is evident that most of the research on SG was done using deep learning algorithms
irrespective of the dataset size. But ML algorithms performed better than the deep learning models being subjected to
the SG dataset due to its size. The present work thus emphasizes on the use of ML algorithms on the SG dataset.

3 | P R O P O S ED M O D E L

The work-flow of the proposed model is depicted in Figure 2. The steps involved in the proposed methodology can be
summarized as follows:

• SG dataset from UCI ML repository is loaded.

• Preprocessing of the dataset is performed by using min-max method for normalization and label encoding for data
transformation.
BASHIR ET AL. 7 of 23

TABLE 1 Survey of the literature on SGs

Problem definition Models used Inferences

Prediction of electric load in SG 27
↣ Deep Learning LSTM-based LSTM RNN may not perform well on different dataset as
RNN model for prediction. train split was more.
↣Recursive feature
elimination for feature
selection.
↣Extra trees regressor as
benchmark machine
learning model.
↣Genetic algorithm for
identifying the time lages
and layers in LSTM
Demand forecast on individual ↣ANN 3 layered (60-20-1 The STMLF was not compared with other model, impact of
consumer's household load28 neuron) data during prediction and the price demand relationship
↣ SVM was not discussed.
↣ Regression model
STLF within microgrids29 ↣Self-Organizing Map System was not evaluated with microgrid-sized
↣K-Means Clustering environment. Also number of patterns considered for
↣ Multilayer Perceptron evaluation against other models was lesser
Predicting long-term and medium- ↣ANN-NAEMI Prediction results have enormous errors with large data and
term energy demand in SG at ↣ MLR input parameter sets
district level30 ↣ AdaBoost
Correlating the non-linear pattern in ↣ LSSVM-MIMO The model can predict the load and power signals without
the electricity load and price ↣ ANN considering the prior data on particular forecast day but it
signals31 ↣ QOABC optimization might not be more accurate
algorithm
↣ GMI
Comprehensive review on load ↣ ANN, WT-ANN,ARIMA- AI-based forecasting techniques are more accurate than
prediction and dynamic pricing of SVM, Regression,WTEC- statistical models
electricity32 ANN, Fuzzy logic
↣ Statistical models
Review on AI-based load forecasting ↣ ANN Integrated approaches can be used for training the NN to
models34 ↣ BP, gradient descent attain better forecast results. ANN performs better than BP
Early warning system for prediction ↣ SVM Grid resilience feature of the SG can be determined
of blackout events in SG to avoid
cascading failure35
Midterm load forecasting36 ↣ SVM NA
↣ ANN
Short-term load forecasting37 ↣ SVM SVM attain global minimum repeatedly. But both models
↣ ANN shows deviations in forecast on continuous load and
erratic loads
Demand and load forecasting in SG38 ↣ SVM, SVR SVR proved to be the best choice for the context (even in
↣ Linear Regression case of huge data set, with the given data set it was
↣ NN with BP algorithm for compatible with the other algorithms) whereas MLP have
training higher computational complexity
↣ MLP
Forecasting Lake water Levels39 ↣ SVM SVM showed yet compatible and better results than other
↣ ANN model in long-term
↣ Seasonal Autoregressive
model
Financial tie series forecast40 ↣ Fuzzy based SVR The model provided best results in noise and non
-stationarity data
(Continues)
8 of 23 BASHIR ET AL.

TABLE 1 (Continued)

Problem definition Models used Inferences

Short-term load forecast 41
↣ SVM with polynomial basis Single kernel SVM function performed weakly
function and Radial basis
function
↣ PCA
↣ Rough set theory
Non-stationary financial time series ↣ C-ascending SVM with Model utilized less support vectors than the traditional one.
forecast42 regularized risk function Model should be explored with sophisticated weight
function.
Exchange Rate Prediction44 ↣ Linear and non-linear SVM SVM outperforms conventional NN and applied structural
↣ GA for feature selection risk minimization principle to reduce the generalization
error.
Cyclic electric load forecasting45 ↣SVR Outperformed ARIMA and Tensor Flow SVR simulated
↣ Chaotic genetic algorithm annealing model
(Simulated Annealing
algorithm)

FIGURE 2 Proposed methodology

• The dataset is then split into training and testing data.

• The dataset is trained by various machine learning algorithms.
• The performance of the ML algorithms is then evaluated with several metrics.

3.1 | Preprocessing

Pre-processing plays an essential role in improving data quality and also the performance of the ML algorithms.47,48
The two common pre-processing techniques used in any machine learning model are normalization and data
BASHIR ET AL. 9 of 23

FIGURE 3 Maximum marginal hyperplane with support

vectors

transformation. The data in an SG dataset are scattered with different ranges which often lead to bias towards values
having higher weights, thereby degrading the performance of the proposed model. In order to avoid this, min-max nor-
malization is used in the present study to normalize the SG dataset. The min-max normalization fits the data into a
common scale, which improves the performance of the classifiers. redMachines use mathematical formulae to process
the data and hence requires data to be in numeric format. Since most of the dataset contains both numerical and cate-
gorical values, data encoding is done during data pre-processing,49 which converts the non-numeric values to numeric
ones before being fed to the ML models.
The pre-processed data is then split into training and testing datasets. The ML algorithms are trained by the training
dataset and then the trained algorithms are tested with new data set to evaluate its performance. In this work, 70% of
the dataset is used to train the ML algorithms and the remaining 30% of the dataset is used for evaluating the perfor-
mance of the trained ML algorithms.
Due to the minimal size of the dataset, several ML algorithms are used for the purpose of classification instead of
Deep Learning based algorithms. Some of the popular ML algorithms, namely SVM, Logistic Regression, Naive Bayes,
Neural Networks and Decision Tree algorithms, are used in this work to classify the SG dataset. The performance of the
ML algorithms are then evaluated using metrics - precision, recall, F1-score, Receiver Operating Characteristic and
accuracy. The results obtained are then compared with the recent works on the SG datasets. The ML algorithms used
in this work are discussed in the below subsections.

3.2 | Support vector machine

SVM aims to solve classification and regression problems and most researchers prefer to use SVM due to its capability
to with high accuracy using minimal computing power. The versatility of SVM lies in the kernelization process n which
it uses the kernel trick to model nonlinear decision boundaries.
SVM aims to classify points using hyperplanes and ensures that after developing hyperplanes, two margin lines are
produced creating classification points that are linearly separable.50 These margin lines are created to move one of the
margin lines to the nearest positive point and another margin line to the nearest negative point as shown in Figure 3.
The distance between the two parallel margins is referred to as the marginal distance. The primary aim of this approach
is to maximize the marginal distance by selecting the best hyperplane.
Support Vectors: The vector points extremely closer to the hyperplane are classified as support vector points, and
these two data points specifically contribute to the results of the algorithm whereas other data points have no signifi-
cant impact. Also, it is important to highlight that the removal of support vectors alters the hyperplane's position.
Good Margin: The support vector points of the positive class and the negative class maintain the maximum dis-
tance to the hyperplane.
Bad Margin: This refers to the hyperplane that is close to either positive support class vectors or negative support class vectors.
Hard Margin: This refers to data points that can be separated from the positive class and the negative class by
maintaining the maximum marginal distance between the parallel hyperplane.
10 of 23 BASHIR ET AL.

Soft Margin These are data points that cannot be separated from the positive class and the negative class by draw-
ing a hyperplane. Soft margins are very difficult to manage as positive and negative data points are intermixed wherein
the accuracy efficiency is diminished. The line equation is given in Equation 1.

b = xa + y ð1Þ

The two vectors are v, a.

0 0 11
−y 1
B C B C
vB C B C
@ −x Aanda@ a A
1 b ð2Þ
v a = −y × ð1Þ + ð −x Þ × a + 1 × b
T

vT a = b −xa −y

The hyperplane equation is given in Equation 3.

vT a + y = 0 ð3Þ

where, y is bias and v, a is vector.

Consider two support vectors, a+ (Positive Support Vector) and a− (Negative Support Vector). The distance between
the two margin lines should be the maximum for scattering the data points linearly. Maximum margin helps to mini-
mize loss. The equation for the approximation of the margin is shown in Equation 4.

v v v
ða + − a − Þv = ða + −a − Þ = a+  − a−  ð4Þ
kvk kvk kvk

Cost Function and Gradient Updates SVM's aim is to maintain maximum margin distance between data points and
hyperplane. Hinge loss function allows maximum margin between data points and hyperplane which is given in Equation 5.
X   
M ðvÞ = max 0, 1 −bi vT ai + y + λ k vk22 ð5Þ
i=1

P
1 − bi[vTai + y]) is used to minimize misclassification,
i = 1max(0,
is regularization used to avoid over-fitting. λ is known as the regularization factor to maximize the marginal
λ k vk22
difference and holds ai on the appropriate margin side. The difference between positive and negative hyperplane is
measured using Hinge loss as given in Equation 6.
  
max 0, 1 −bi vT ai + y = 0
  ð6Þ
) bi vT ai + y = 1


+ 1 va + y≥1
bi
− 1 va + y ≤ −1

Equation 6 is further derived as in Equation 7.

v v 1 −y −y −1 2
a+  −a −  = − = ð7Þ
kvk kvk kvk kvk kvk

Finally, the objective function is given in Equation 8.

BASHIR ET AL. 11 of 23

2 1 1
max ! max ! min k v k! min k vk2 ð8Þ
kvk kvk 2

Polynomial kernel: The polynomial kernel is a kernel functions used along with SVM to represent similar data
points in a dataset.51 The polynomial kernel for degree-d is delineated as in Equation 9.
 s
F ða, bÞ = aT b + z ð9Þ

where a and b are vectors of the data set, z ≥ 0 is a weight vector comparing higher-order and lower-order polynomial
values. As a kernel, F correlates to an internal data points in a higher dimensional space that is based on a certain map-
ping φ as shown in Equation 10.

F ða, bÞ = hφðaÞ,φðbÞi ð10Þ

The influence of φ is shown below in Equation 11, where s = 2.

!2
X
m X
m    X i − 1 pffiffiffi
m X pffiffiffi
F ða, bÞ = ai bi −z = a2i b2i + 2ai aj 2bi bj ;
i=1 i=1 i=2 j=1
ð11Þ
X
m pffiffiffiffiffi pffiffiffiffiffi 
+ 2zai 2zbi + z2
i=1

D pffiffiffi pffiffiffi pffiffiffi

φðaÞ = a2m ,…, a21 , 2am am − 1 , …, 2am a1 , 2am − 1 am − 2
pffiffiffi pffiffiffi pffiffiffiffiffi pffiffiffiffiffi E ð12Þ
…, 2am − 1 a1 ,…, 2a2 a1 , 2zam , …, 2za1 , z,

Radial basis function kernel(RBF): RBF Kernel is a common kernel function that is predominantly used in
numerous kernelized machine learning algorithms. RBF kernel is widely used for SVM classification tasks. The RBF
0
kernel for two separate sample vectors a and aa , in decision boundary, is shown in Equation 13.
!
0 ka−a0 k2
F ða, a Þ = exp − ð13Þ
2ρ2
0
ka − a k2 is considered to be the Euclidean distance between two vectors and ρ is considered as a variable where
ξ = 2ρ1 2 .

F ða, a0 Þ = exp −ξka −a0 k
2
ð14Þ

The RBF kernel value decreases with respect to the euclidean distance ranging from 0 to ∞. When ρ = 1 the equa-
tion is expressed as in Equation:

1 X∞
ða T a 0 Þ 1
j
1
exp − ka− a0 k = exp − k ak2 exp − ka0 k ;
2 2
2 j=0
j! 2 2
m m ð15Þ
∞ X
X 1 1  af
am 1 f
1 a0 1m1  a0 f f
exp − k ak2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp − ka0 k pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
=
j = 0 Σmi = j
2 m1 ! mf ! 2 m1 !mf !
12 of 23 BASHIR ET AL.

Sigmoid Kernel Sigmoid Kernel is often referred to as Hyperbolic Tangent Kernel which is originated from the
neural network research area. In most cases, sigmoid function has been used as an activation function for neural net-
works. The mathematical representation of the sigmoid function is shown in the Equation 16.
 
F ða, bÞ = tanh αaT b + z ð16Þ

where α is chosen to be a slope, z is constant, α is considered as M1 where M is data dimension.

3.3 | Logistic regression

Logistic regression is one of the most popular machine learning algorithms.52 The goal of the algorithm is to seek a sim-
ilarity between both the likeliness of desired outcome and attributes. Logistic Regression Equation is as follows:

p ða Þ
log = ζl + A1a ð17Þ
1 − pðaÞ

eX 0 + ζ1a
pðaÞ = ð18Þ
1 + k ζ0 + ζ1a

Equation 18 is considered as a sigmoid function, generating an S-like curve. The probability value is in the range
of 0<p<1.

3.4 | Naive Bayes

Naive Bayes (NB)53 is a classification technique that assumes independence between predictors. NB consists of two
parts namely Naive and Bayes. The NB classifier assumes that the presence of a particular feature in a class is indepen-
dent of the presence of any other feature. All the features independently contribute to the probability of a variable
belonging to the specific target class or otherwise. NB is quite trivial to develop and is particularly useful for very large
dataset. In probability theory and statistics, this is alternatively known as Bayes Law. The conditional probability of NB
is shown in the Equation 19.
   
  p Z f p ajZ f
p Z f ja = ð19Þ
p ða Þ

The Bayes classifier of the probability model is shown in the Equation 20.

 Y m  
^b = argmax p Z f p ai jZ f ð20Þ
f f1, …,F g i=1

3.5 | K-Nearest neighbour (k-NN)

k-NN is one of the simplest ways to classify data and is primarily used for regression and classification. It is used to
identify data points using the closest training examples in the feature space. It involves instance-based learning where
the function is locally approximated and all performance is deferred until classification. The F value is considered as
the number of closest neighbors in a vector classification and the selection of the most appropriate F value is essential
BASHIR ET AL. 13 of 23

for attaining superior quality results. In the proposed approach, we considered F = 5, leaf-size = 30 and Minkowski
metric weights are uniform. Equations for k-NN are as follows:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u F
uX
Eculidean equation = t ðAi −biÞ2 ð21Þ
i=1

X
F
Manhattan equation = j Ai −Bi j ð22Þ
i=1

!1=r
X
F
r
Minkowski = ð jA−Bij Þ ð23Þ
i=1

3.6 | Decision tree (DT)

DT belongs to one of the few classification models where we can understand the exact reasoning behind the classifier,
making a particular decision.54 DT provides a graphical representation of all possible decision solutions based on cer-
tain conditions. It begins with a root and then branches off to a number of possible solutions, just like a tree. The root
node initially gets added to the tree, which receives the trained data set, and then each node asks a true and false ques-
tion to one of the features. Henceforth, the dataset gets divided into two different subsets. These subsets then act as an
input to the child node. The aim is to produce the purest possible distribution of the labels at each node. This iteration
process continues till no further questions are asked, and it finally reaches the leaf. The equation for Entropy is given in
Equation 24.

X
n
E : I ðp1 , p2 , …pn Þ = ðpi logð1=pi Þ ð24Þ
i=1

(p1, p2, …pn) denotes the probabilities of class labels.

In the proposed model the splitting was performed using Gini-index, max-depth = none, min-samples-split = 2,
min-samples-leaf = 1, class-weights = none, random-state = none, min-impurity-decrease = 0.0 and min-impurity-
split = none.

3.7 | Neural networks

The neural network maps the input units to its appropriate output unit after performing some mathematical calcula-
tions.55,56 The neural network consists of input layer where features are given as input, hidden layers are placed
between input and output layers. The role of the hidden layer is to multiply weights to the input layer and then pass
the resultant values to activation functions. The predictions are performed at the last layer, called as an output layer.
The general architecture of neural networks is depicted in Figure 4.
!
X
k
H n = ϕ1 + W mn + θn ð25Þ
i=1

Equation 25 shows the working of hidden layer when features fi are given as input. Weight between nth input and
mth hidden layer is calculated by Wmn where θn is the value of bias factor.
14 of 23 BASHIR ET AL.

FIGURE 4 Neural network architecture

FIGURE 5 Neural network results

!
X
k
output = ϕ2 + W MO + θO ð26Þ
i=1

The mapping inputs to the outputs is an iterative operation, in which weights Wmn are changed in each iteration. Back
propagation (BP) algorithm is one of the widely used algorithms and the equation for back propagation is given below:

∂Ef
W nm ðt + 1Þ = W nm ðt Þ− e ð27Þ
W nm

The error between the calculated and target output is used for weighing updates.

4 | R ES U L T S A N D D I S C U S S I O N

In this work, the experimentation is carried out using “Google Colab”, Google's online Graphical Processing Unit (GPU).
Python 3.7 is used as the programming language. The dataset used for the experimentation is collected from publicly avail-
able UCI machine learning repository57 which consists of 10 000 instances with 14 attributes. The attributes in the dataset
are electricity producer values, nominal power consumed/produced, coefficient related to price elasticity, the maximum
value of the equation root and the stability of the system (class label, whether the system is stable or not).
The performance metrics considered in the study are accuracy, recall, F1 measurement, and detection rate
(DR) which helps to evaluate the proposed solution. The above performance measurements are based on True Positive
(TP), False Positive (FP), False Negative (FN) and True Negative (TN).
BASHIR ET AL. 15 of 23

In the present study, 70% of the SG dataset is used for training and remaining 30% of the SG dataset is used for test-
ing and validation purposes. The machine learning models are evaluated based on the parameters mentioned above in
association with the Receiver Operating Characteristic (ROC) curve which helps in justifying the results.

4.1 | Neural network results

Figure 5A represents the confusion metric (CM) and Figure 5B represents classification report (CR) for neural network classifier.
Using neural network classifier we achieved an accuracy of 97.50%, for stable class and 98.50% accuracy for fault class. Also, 2.5%
false positive rate (FPR) and 1.5% false negative rate (FNR) is achieved for stable and fault class respectively, depicted in
Figure 5A. The result highlight the fact that 97.60% precision, 99.50% recall, and 98.50% F1-Measure is achieved for fault class.
Similarly for stable class precision, recall and F1-Measure scores are 99.00%, 95.70% and 97.30% respectively as represented in
Figure 5B. Figure 6 represents the ROC curve for neural network wherein the area under the curve score is 97.96%.

4.2 | Support vector machine results

4.2.1 | SVM polynomial kernel

Figure 7A and Figure 7B presents the SVM Polynomial Kernel results. As shown in the figure, 91.20% average predica-
tion accuracy is achieved in case of SVM Polynomial Kernel wherein 977 records are classified accurately as a stable
class achieving an accuracy of 90.10%. On the other hand, 1759 out of 1916 records are detected as fault class with the
accuracy of 91.80% using SVM with Polynomial kernel as shown in Figure 7A. In Figure 7B SVM polynomial kernel

FIGURE 6 NN roc curve

FIGURE 7 SVM polynomial kernel results

16 of 23 BASHIR ET AL.

FIGURE 8 SVM polynomial kernel ROC curve

FIGURE 9 SVM RBF kernel results

FIGURE 10 SVM RBF kernel ROC curve

precision score for fault and stable grids are 94.10% and 83.00%, respectively. Similarly recall for both fault and stable
classes are 89.60% and 90.10% respectively. F1-Measure for fault class is 97.80% and for stable class F1-Measure score is
86.40%, respectively. The Area under the curve score is 90.21% as shown in Figure 8.
BASHIR ET AL. 17 of 23

4.2.2 | SVM RBF kernel

Figure 9A and Figure 9B presents the SVM RBF Kernel results wherein 87.46% average predication accuracy is achieved in case of
the SVM RBF Kernel. A total number of 858 records are classified accurately as stable class generating an accuracy of 79.20%, while
1766 out of 1916 records are detected as fault class yielding an accuracy of 92.20% when SVM with RBF kernel is deployed as shown
in Figure 9A. In Figure 9B, SVM RBF kernel precision score for fault and stable grids are 86.80% and 83.10%, respectively. Similarly
recall for both fault and stable classes are 91.30% and 75.50%, respectively. F1-Measure for the fault class is 89.00% and for the stable
class, the F1-Measure score is 79.10%, respectively. The Area under the curve score is 86.89% as shown in Figure 10.

4.2.3 | SVM sigmoid kernel

It can be observed from Figure 11A that 157 records are correctly classified as stable and 1023 records are predicated as
fault class when SVM sigmoid kernel is implemented. Similarly Figure 11B depicts the classification report for SVM sig-
moid kernel. Precision, recall and F1-Measure scores for fault class are 52.30%, 53.10%, and 52.70%, respectively. Simi-
larly precision for stable class is 14.70%, recall score is 14.30% and F1-Measure score is 14.50%, respectively. Figure 12
represents the area under the curve (AUC) for SVM sigmoid kernel and AUC score for SVM sigmoid kernel is 33.71%.

4.3 | Decision tree

A total of 1084 records are classified correctly as stable, with an accuracy of 100%, while 1915 records are detected as
fault with an accuracy of 99.90%, respectively using decision tree as depicted in Figure 13A. The precision, recall and

FIGURE 11 SVM sigmoid kernel results

FIGURE 12 SVM sigmoid kernel ROC curve

18 of 23 BASHIR ET AL.

FIGURE 13 Decision tree results

FIGURE 14 Decision tree ROC curve

FIGURE 15 KNN ROC curve

F1-Measure scores are 100%, 99.90%,100% respectively for fault class. Similarly for stable class precision, recall and F1-Measure
scores are 99.90%, 100%,100% respectively as depicted in Figure 13B. Figure 14 presents the decision tree AUC score of 99.95%.

4.4 | K nearest neighbour (KNN) results

Figure 15 depicts the RoC for KNN. Figure 16A represents the confusion matrix while Figure 16B represents classifica-
tion report for KNN. KNN classifier detects 701 stable class instances correctly with an accuracy of 64.70%. Similarly,
BASHIR ET AL. 19 of 23

FIGURE 16 KNN results

FIGURE 17 Naive Bayes results

the fault grid prediction accuracy for KNN is observed to be 85.60%, which refers 1641 instances to be detected accu-
rately out of 1916. The results show that 81.10% precision, 85.60% recall and 83.30% F1-Measure score is achieved in
case of fault class. For the stable class, precision, recall and F1-Measure scores are 71.80%, 64.74%, 68.10%, respectively.
Figure 15 represents the ROC curve of KNN classifier and AUC score for KNN classifier, which is 76.45%.

4.5 | Naive Bayes results

A total of 1051 records are classified correctly as stable with an accuracy of 97%, while 1875 records are detected as fault
with an accuracy of 97.90%, respectively, using Naive Bayes as depicted in Figure 17A. Precision, recall, and
F1-Measure scores are 98.30%, 97.90%, 98.10%, respectively, for fault class. Similarly for stable class, precision, recall
and F1-Measure scores are 96.20%, 97%, 96.60%, respectively, depicted in Figure 17B. AUC for Naive Bayes is 97.26% as
depicted in Figure 18.

4.6 | Logistic regression results

It is observed that 877 records are classified correctly as stable with an accuracy of 80.90%, while 1770 records are
detected as fault with an accuracy of 92.40%, respectively, using logistic regression as depicted in Figure 19A. Precision,
recall, and F1-Measure scores are 89.50%, 92.40, 90.90%, respectively, for the fault class. Similarly for stable class,
20 of 23 BASHIR ET AL.

FIGURE 18 Naive Bayes ROC curve

FIGURE 19 Logistic regression results

FIGURE 20 Logistic regression ROC curve

precision, recall and F1-Measure scores are 85.70%, 80.90%,83.20%, respectively, as depicted in Figure 19B. AUC for
logistic regression is 87.63% which is represented in Figure 20.
The predication accuracy for the decision tree model is 99.96%, which is highest in comparison to other classifiers
used in this work as depicted in Table 2. Since Decision tree is a probability based algorithm, it outperforms other algo-
rithms used in this work in terms of prediction accuracy, precision, recall, and F1-Measure, respectively.
The authors in25 achieved 99.07% accuracy, 97% precision, 100% recall, and 99.00% F1-Measure for the stable class.
Similarly for unstable class they achieved 100% precision, and 99.00% recall and F1-Measure, respectively. In this
research, decision tree achieved 99.96% predication accuracy, 100% precision, and F1-Measure, respectively, for fault
BASHIR ET AL. 21 of 23

T A B L E 2 Testing accuracy
Classifier Testing accuracy
comparison of ML algorithms
SVM(Polynomial Kernel) 91.20
SVM (RBF Kernel) 87.46
SVM (Sigmoid Kernel) 39.33
KNN 78.06
Logistic Regression 88.23
Naive Bayes 97.53
Decision Tree 99.96
Neural Network 98.13

TABLE 3 Comparison with

Method Accuracy
existing works
25
MLSTM 99.07
58
AdaBoost 97.87
KNN59 95.90
Proposed model 99.96

class and 99.90% recall. For stable class, decision tree classifier achieved 99.90% precision, 100% recall, and 100%
F1-Measure score, respectively. In25 authors also achieved 99.27% AUC using multidimensional long short term mem-
ory, while decision tree classifier achieved 99.95% AUC. Table 3 depicts the comparison of the current work with recent
works.
From the results obtained the following can be concluded:

• Since the size of the dataset is not huge, ML algorithms are apt for classification of the SG dataset when compared to
deep learning models.
• DT classifier outperforms the other ML algorithms considered as the number of attributes are relatively less.

5 | C ONCLUSIONS A ND F UTURE SC OPE

Machine learning algorithms play a vital role in maintaining the stability of SG due to its ability of predicting the elec-
tricity demands of the customers. With the emergence of various machine learning algorithms, the ultimate challenge
is to find the most appropriate algorithm to predict the stability of the SG. In order to achieve this, a comprehensive sur-
vey on the state-of-the-art machine learning algorithms have been performed in order to predict the stability of SGs.
The dataset used in this work is collected from the publicly available UCI machine learning repository. The experimen-
tal results proved that the decision tree classification algorithm outperforms SVM, KNN, Naïve Bayes, Logistic Regres-
sion and Neural Network. The limitations of the current work indicate that the size of the dataset is not quite huge. But
in real-time, SGs generate massive data. In order to address this problem, as part of future work, effective feature
engineering-based models could be applied on real-time SG data.

A C K N O WL E D G E M E N T
Researchers Supporting Project number (RSP-2020/250), King Saud University, Riyadh, Saudi Arabia.

CONFLICT OF INTEREST
The authors declare no potential conflict of interests.

P EE R R EV IE W
The peer review history for this article is available at https://publons.com/publon/10.1002/2050-7038.12706.
22 of 23 BASHIR ET AL.

DATA AVAILABILITY STATEMENT

The data used in this work is from publicly available UCI machine learning repository.

ORCID
Ali Kashif Bashir https://orcid.org/0000-0003-2601-9327
Thippa Reddy Gadekallu https://orcid.org/0000-0003-0097-801X

R EF E RE N C E S
1. Zhao B, Zeng L, Li B, et al. Collaborative control of thermostatically controlled appliances for balancing renewable generation in smart
grid. IEEJ Trans Elect Electron Eng. 2020;15(3):460-468.
2. Gungor VC, Sahin D, Kocak T, et al. Smart grid technologies: communication technologies and standards. IEEE Trans Indus Informat.
2011;7(4):529-539.
3. Alazab M, Huda S, Abawajy J, et al. A hybrid wrapper-filter approach for malware detection. J Networks. 2014;9(11):2878-2891.
4. Alazab M, Layton R, Broadhurst R, Bouhours B. Malicious spam emails developments and authorship attribution. IEEE. 2013;1:58-68.
5. Alazab M, Broadhurst R. An Analysis of the Nature of Spam as Cybercrime. Cham, Switzerland: Springer; 2017:251-266.
6. Desai SK, Dua A, Kumar N, Das AK, Rodrigues JJ. Demand response management using lattice-based cryptography in smart grids.
IEEE. 2018;1:1-6.
7. Deepa N, Pham QV, Nguyen DC, et al. A survey on blockchain for big data: approaches, opportunities, and future directions arXiv Pre-
print arXiv:2009.00858 2020.
8. Numan M, Subhan F, Khan WZ, et al. A systematic review on clone node detection in static wireless sensor networks. IEEE Access.
2020;8:65450-65461.
9. Dileep G. A survey on smart grid technologies and applications. Renew Energy. 2020;146:2589-2625.
10. Jindal A, Aujla GS, Kumar N. SURVIVOR: a blockchain based edge-as-a-service framework for secure energy trading in SDN-enabled
vehicle-to-grid environment. Comput Networks. 2019;153:36-48.
11. Kaneriya S, Tanwar S, Nayyar A, et al. Data consumption-aware load forecasting scheme for smart grid systems. IEEE. 2018;1:1-6.
12. Qureshi NMF, Bashir AK, Siddiqui IF, Abbas A, Choi K, Shin DR. A knowledge-based path optimization technique for cognitive nodes
in smart grid. IEEE. 2018;1:1-6.
13. Bayindir R, Colak I, Fulli G, Demirtas K. Smart grid technologies and applications. Renew Sustain Energy Rev. 2016;66:499-516.
14. Reddy GT, Sudheer K, Rajesh K, Lakshmanna K. Employing data mining on highly secured private clouds for implementing a security-
asa-service framework. J Theor Appl Inf Technol. 2014;59(2):317-326.
15. RM SP, Bhattacharya S, Maddikunta PKR, et al. Load balancing of energy cloud using wind driven and firefly algorithms in internet of
everything. J Parallel Distributed Comput. 2020;142:16-26.
16. Bera S, Misra S, Rodrigues JJ. Cloud computing applications for smart grid: a survey. IEEE Trans Parallel Distribut Syst. 2014;26(5):1477-
1494.
17. Zhang Q, Sun Y, Cui Z. Application and analysis of ZigBee technology for smart grid. IEEE. 2010;1:171-174.
18. Maddikunta PKR, Gadekallu TR, Kaluri R, Srivastava G, Parizi RM, Khan MS. Green communication in IoT networks using a hybrid
optimization algorithm. Comput Commun. 2020;159:97-107.
19. Kappagantu R, Daniel SA. Challenges and issues of smart grid implementation: a case of Indian scenario. J Elect Syst Informat Technol.
2018;5(3):453-467.
20. Mittal M, Iwendi C, Khan S, Rehman Javed A. Analysis of security and energy efficiency for shortest route discovery in low-energy adap-
tive clustering hierarchy protocol using Levenberg-Marquardt neural network and gated recurrent unit for intrusion detection system.
Trans Emerg Telecommunicat Technol. 2020;1:e3997.
21. Deepa N, Prabadev B, Kumar Reddy P, et al. An AI-based intelligent system for healthcare analysis using ridge–Adaline stochastic gradi-
ent descent classifier. J Supercomput. 2020;1:1998-2017.
22. Gadekallu TR, Khare N, Bhattacharya S, Singh S, Maddikunta PKR, Srivastava G. Deep neural networks to predict diabetic retinopathy.
J Ambient Intelligence Humanized Comput. 2020;1:1–14.
23. Patel H, Singh Rajput D, Thippa Reddy G, Iwendi C, Kashif Bashir A, Jo O. A review on classification of imbalanced data for wireless
sensor networks. Int J Distributed Sensor Networks. 2020;16(4):1550147720916404.
24. Tian WA. Review of smart grids and their future challenges. In:. 173. EDP. Sciences. 2018;174:02025.
25. Alazab M, Khan S, Krishnan SSR, Pham QV, Reddy MPK, Gadekallu TR. A multidirectional LSTM model for predicting the stability of
a smart grid. IEEE Access. 2020;8:85454-85463.
26. Venayagamoorthy GK. Dynamic, stochastic, computational, and scalable technologies for smart grids. IEEE Comput Intelligen Magaz.
2011;6(3):22-35.
27. Bouktif S, Fiaz A, Ouni A, Serhani MA. Optimal deep learning lstm model for electric load forecasting using feature selection and
genetic algorithm: comparison with machine learning approaches. Energies. 2018;11(7):1636.
28. Javed F, Arshad N, Wallin F, Vassileva I, Dahlquist E. Forecasting for demand response in smart grids: an analysis on use of anthropo-
logic and structural data and short term multiple loads forecasting. Appl Energy. 2012;96:150-160.
BASHIR ET AL. 23 of 23

29. Hernández L, Baladrón C, Aguiar JM, Carro B, Sánchez-Esguevillas A, Lloret J. Artificial neural networks for short-term load forecast-
ing in microgrids environment. Energy. 2014;75:252-264.
30. Ahmad T, Chen H. Potential of three variant machine-learning models for forecasting district level medium-term and long-term energy
demand in smart grid environment. Energy. 2018;160:1008-1020.
31. Shayeghi H, Ghasemi A, Moradzadeh M, Nooshyar M. Simultaneous day-ahead forecasting of electricity price and load in smart grids.
Energ Conver Manage. 2015;95:371-384.
32. Khan AR, Mahmood A, Safdar A, Khan ZA, Khan NA. Load forecasting, dynamic pricing and DSM in smart grid: a review. Renew Sus-
tain Energy Rev. 2016;54:1311-1322.
33. Raza MQ, Nadarajah M, Ekanayake C. On recent advances in PV output power forecast. Solar Energy. 2016;136:125-144.
34. Raza MQ, Khosravi A. A review on artificial intelligence based load demand forecasting techniques for smart grid and buildings. Renew
Sustain Energy Rev. 2015;50:1352-1372.
35. Gupta S, Kambli R, Wagh S, Kazi F. Support-vector-machine-based proactive cascade prediction in smart grid using probabilistic frame-
work. IEEE Trans Indus Electron. 2014;62(4):2478-2486.
36. Pan X, Lee B. A comparison of support vector machines and artificial neural networks for mid-term load forecasting. IEEE. 2012;1:
95-101.
37. Mitchell G, Bahadoorsingh S, Ramsamooj N, Sharma C. A comparison of artificial neural networks and support vector machines for
short-term load forecasting using various load types. IEEE. 2017;1:1-4.
38. Ali AS, Azad S. Demand Forecasting in Smart Grid. London: Springer; 2013:135-150.
39. Khan MS, Coulibaly P. Application of support vector machine in lake water level prediction. J Hydrolog Eng. 2006;11(3):199-205.
40. Khemchandani R, Jeyadeva CS. Regularized least squares fuzzy support vector regression for financial time series forecasting. Expert Syst
Appl. 2009;36(1):132-138.
41. Li W, Yan N. Zhang Zg. Study on long-term load forecasting of MIX-SVM based on rough set theory. Power System Protection Control.
2010;38(13):31-34.
42. Tay FE, Cao L. Modified support vector machines in financial time series forecasting. Neurocomputing. 2002;48(1–4):847-861.
43. Cao L, Gu Q. Dynamic support vector machines for non-stationary time series forecasting. Intelligent Data Analysis. 2002;6(1):67-83.
44. Hong WC. A hybrid support vector machine regression for exchange rate prediction. Int J Informat Manage Sci. 2006;17(2):19-32.
45. Hong WC, Dong Y, Chen LY, Panigrahi BK, Wei SY. Support vector regression with chaotic hybrid algorithm in cyclic electric load forecast-
ing. Cham: Springer; 2012:833-846.
46. Mohandes MA, Halawani TO, Rehman S, Hussain AA. Support vector machines for wind speed prediction. Renew Energy. 2004;29(6):
939-947.
47. Tripathy B, Acharjya D, Cynthya V. A framework for intelligent medical diagnosis using rough set with formal concept analysis. arXiv
preprint arXiv:1301.6011 2013.
48. Reddy GT, Reddy MPK, Lakshmanna K, et al. Analysis of dimensionality reduction techniques on big data. IEEE Access. 2020;8:54776-
54788.
49. Tripathy B, Arun K. A new approach to soft sets, soft multisets and their properties. Int J Reason-Based Intellig Syst. 2015;7(3/4):244-253.
50. Gola J, Webel J, Britz D, et al. Objective microstructure classification by support vector machine (SVM) using a combination of morpho-
logical parameters and textural features for low carbon steels. Comput Mater Sci. 2019;160:186-196.
51. Gopi AP, Jyothi RNS, Narayana VL, Sandeep KS. Classification of tweets data based on polarity using improved RBF kernel of SVM. Int
J Informat Technol. 2020;1:1-16.
52. Xie Y, Liu J, Huang T, et al. Outdoor thermal sensation and logistic regression analysis of comfort range of meteorological parameters in
Hong Kong. Build Environ. 2019;155:175-186.
53. Kwon Y, Kwasinski A, Kwasinski A. Solar irradiance forecast using nave Bayes classifier based on publicly available weather forecasting
variables. Energies. 2019;12(8):1529.
54. Zhang Q, Yang Y, Ma H, Wu YN. Interpreting cnns via decision trees. In:; 2019: 6261–6270.
55. Javed AR, Sarwar MU, Khan S, Iwendi C, Mittal M, Kumar N. Analyzing the effectiveness and contribution of each axis of tri-axial
accelerometer sensor for accurate activity recognition. Sensors. 2020;20(8):2216.
56. Javed AR, Beg MO, Asim M, Baker T, Al-Bayatti AH. AlphaLogger: detecting motion-based side-channel attack using smartphone key-
strokes. J Ambient Intelligen Human Comput. 2020;1:1-14.
57. Arzamasov V, Böhm K, Jochem P. Towards concise models of grid stability. IEEE. 2018;1:1-6.
58. Pisica I, Eremia M. Making smart grids smarter by using machine learning. VDE. 2011: 1–5.
59. Hosseinzadeh J, Masoodzadeh F, Roshandel E. Fault detection and classification in smart grids using augmented K-NN algorithm. SN
Appl Sci. 2019;1(12):1627.

How to cite this article: Bashir AK, Khan S, Prabadevi B, et al. Comparative analysis of machine learning
algorithms for prediction of smart grid stability. Int Trans Electr Energ Syst. 2021;e12706. https://doi.org/10.
1002/2050-7038.12706