2023 International Conference on Innovative Computing, Intelligent Communication and Smart Electrical Systems (ICSES) | 979-8-3503-1920-0/23/$31.

00 ©2023 IEEE | DOI: 10.1109/ICSES60034.2023.10465331

A Hybrid Deep Learning Neural Network Model

and Data-Driven Methods for Smart Grid Theft
Detection
S.Sudha Merlin1, R. Yugha2, Hubert Mary. L3, Varalakshmi. P4, Dinesh Mavaluru5
1
Assistant Professor, Department of EEE, DMI College of Engineering, Palanchur, Chennai.
2
Assistant Professor, School of Computing, Sathyabama Institute of Science and Technology, Chennai.
3
Assistant Professor, Department of ECE, Jeppiaar Institute of Technology, Chennai
4
Assistant Professor, Department of AI & DS, Dhanalakshmi College of Engineering, Chennai
5
Adjunct Professor, Department of VLSI Microelectronics, Saveetha School of Engineering, Saveetha Institute of Medical and
Technical Sciences (SIMATS), Chennai

E-mail :[email protected], [email protected], [email protected], [email protected],

[email protected]

Abstract-Smart grids are harmed by electricity theft, and use. In order to keep the energy supply and demand in
electric companies lose a lot of money as a result. Recent equilibrium, AMI facilitates two-way communication
research has looked at ways to use a mix of deep learning between users and the grid. Despite the difficulty in
(DL), machine learning (ML), and statistical methods to spot managing energy use, there are two forms of energy loss
anomalies and illegal tendencies in electrical consumption
(EC) data collected by smart meters. In this research, we
that plague the smart grid. Technical losses (TLs) are the
combine the best features of the convolutional neural first category, while nontechnical losses (NTLs) are the
network with the gated recurrent unit to create a unique second. The latter is caused by heat dissipation in the
hybrid model that we call the CNNGRU Model. Accurate electrical grid's transmission and distribution equipment.
power theft detection is one of the model's primary goals, The latter is what happens when consumers cause
along with resolving problems with previous attempts at problems for a company by not doing things like
this. Overfitting and an insensitivity to data imbalance are registering their networks, paying their bills, not leaving
two of the problems. Despite its processing complexity, the their smart meters alone, etc. There are two primary types
proposed strategy exceeds its rivals in terms of accuracy. of energy loss: non-technical losses (NTL) and technical
Scientists may dedicate their focus to finding ways that
increase performance in the presence of enormous volumes
losses. To combat technological losses, many approaches
of data as computers become more powerful. The efficacy of have been used or proposed.
the hybrid DL model is studied using a variety of
performance measures, including such as F1-score, Electricity theft is one of the NTLs affecting demand.
precision, recall, accuracy, and false positive rate. Power businesses have a lot to lose when electricity is
stolen or misused, since it disrupts their operations,
Index Terms—Smart Grid Theft Detection, Convolutional increases their generating load, and drives up the price of
Neural Network, Gated Recurrent Unit, Hybrid Deep Learning electricity for honest customers. Loss of electricity also
Model.
has negative effects on economic development and the
reliability of power grids. Typically, power theft and
I.INTRODUCTION fraud account for the majority of business losses, or NTL.
Electricity theft and fraud continue to be significant
Since the advent of the industrial revolution, electricity sources of revenue loss for power utilities. The theft of
has been an indispensable resource. The term "power electricity puts a strain on the power grid and can cause
grid" refers to the infrastructure built to transmit dangerous fires that put lives at risk. In addition, they
electricity to these areas. There are two halves to the result in lost money for utilities that produce electricity.
electricity grid: the supply side and the demand side. The Taking on the issue of stolen power is difficult [2].
demand side's requirements determine whether or not Double-tapping assaults, modifying meter readings via
electricity production is raised or lowered. While communication links, and the use of shunt devices are all
generating, transmitting, and delivering power, part of it methods that may be used to steal electricity. It is
is unfortunately wasted. In order to meet the rising common knowledge that a nation's power consumption
demand for energy, the grid must be able to efficiently has a direct bearing on its progress and is, thus, a crucial
produce and distribute electricity [1]. To this end, smart indicator that determines the basis of industrialization.
grid incorporates advanced metering infrastructure (AMI) Theft of electricity has reached an all-time high in
to track how often and how much electricity customers response to the ever-increasing demand for it. The

Authorized licensed use limited to: Kwame Nkrumah Univ of Science and Technology. Downloaded on March 07,2025 at 22:25:34 UTC from IEEE Xplore. Restrictions apply.
 production of electricity from the combustion of fossil learning and deep learning models. Not only that, but the
fuels accounts for 70% of greenhouse gas emissions. FPRs for common ML models like k-nearest neighbors
Theft of power overshadows efforts to minimize and naive Bayes are rather good. Due to the scarcity of
greenhouse gas emissions in underdeveloped nations. manpower and equipment for on-site inspection, it is
Energy theft reduces a country's ability to generate widely acknowledged in the literature that power
electricity since it wastes resources. Blackouts and load- companies cannot abide by low rates of identification and
shedding are a result of electricity theft, which leads high FPRs. Rather of relying solely on traditional GRU
people to seek out other ways to meet their energy needs, and convolutional neural network (CNN) to handle the
such as using gas- or diesel-powered generators and overfitting problem, the suggested PSO-GRU approach
contributing to greenhouse gas emissions. uses a metaheuristic model called CNNGRU to fine-tune
the variables and increase the learning rate.
Energy theft has been largely ignored in climate
negotiations in favor of focusing on ways to lower Electricity Generation

greenhouse gas emissions. The capacity of smart meters

(SM) to remotely track and isolate high-risk regions of the
electrical grid has led to their consideration as a strategy
for combating energy theft. It is recommended that SMs
be installed on all transformers, distribution poles, and Power Flow

customer homes [3]. The readings are then sent across a

communications network to the distribution company's
database, where they are analyzed; if problem areas are
identified, electricity is switched off automatically. By
Power Flow
enhancing effectiveness with this technology, GHG
emissions may be reduced instantly and the distribution System Operator

company can get higher total returns. It would also

increase openness in the metering procedure.
Power Flow
In order to meet the rising demand for energy, the grid Advanced Metering
Infrastructure Network
must be able to efficiently produce and distribute
electricity. To this end, smart grid incorporates advanced ...
metering infrastructure (AMI) to track how often and how
Home Area Network
much electricity customers use. In order to keep the Building Area Network

energy supply and demand in equilibrium, AMI facilitates Figure 1. Smart Grid Theft Detection
two-way communication among users and the grid. In
Figure 1, we see how the smart grid may prevent theft. II.RELATED STUDY
Despite the difficulty in managing energy use, there are
two forms of energy loss that plague the smart grid. As a result of power theft, problems arise for both utilities
Technical losses (TLs) are the first category, while and their customers. Consumers feel the high cost of
nontechnical losses (NTLs) are the second [4]. The latter power as utility companies' economic growth is stifled
is caused by heat dissipation in the electrical grid's and new electric dangers are introduced. A combination
transmission and distribution equipment. Theft of power of machine learning and deep learning techniques applied
due to dishonest customers is also a key contributor to to the massive volumes of data generated by smart grids,
NTLs. such as information on consumer use, might be used to
detect cases of energy theft. A deep neural network-based
The development of intelligent electricity networks system of categorization that makes full use of data in the
creates fresh chances for ETD. SM, detectors, and temporal and frequency domains is presented in this study
computer equipment with self-healing processes and as a means of identifying theft. We address shortcomings
electronic communication make up the updated in datasets, such missing data and class imbalance
components of what is known as a "smart grid," which is problems [6], by interpolating existing data and
an improvement over the traditional electricity system. generating new data synthetically. We then conduct tests
EC, electricity rates, and grid health are only few of the in the combined and reduced feature space using principal
metrics that may be gathered by the smart meters and component analysis, and so on, to determine which
detectors. The acquired EC data is used to train the data- features are most important before applying a minimum
driven methods [5] to identify legitimate and malicious redundancy maximum relevance strategy to validate the
samples. Researchers has focused heavily on these best candidates. Through the use of a Bayesian optimizer
methods, however they suffer from issues including the as well as an adaptive moment estimate optimizer,
curse of the number of dimensions, unbalanced classes, research are conducted with varying values of important
and low rate of identification when used alone in machine variables to determine which settings result in the highest

Authorized licensed use limited to: Kwame Nkrumah Univ of Science and Technology. Downloaded on March 07,2025 at 22:25:34 UTC from IEEE Xplore. Restrictions apply.
 accuracy for energy theft recognition and, thus, increase accuracy in detecting when compared with previous
the efficiency of the system. In the end, we show that the techniques. Based on the results of the simulations, the
strategy holds its own against rival approaches that have proposed ETD-ConvLSTM strategy is either more
been evaluated using the same dataset. Our AUC from effective than or as effective as the state-of-the-art
validation was 97%, which was 1% higher than the best methods [8].
AUC from previous research, and our accuracy was
91.8%, making us second best compared to the gold Neither utilities as well as their consumers suffer when
standard [6]. power is stolen. As the economic expansion of utility
businesses is hindered and new electric risks emerge,
Digital analysis is becoming increasingly crucial in consumers experience the high cost of electricity. Smart
China's power business as the country undergoes its grids provide vast amounts of data, including data about
digital revolution [7]. Digital analysis has several customer consumption that might be used in conjunction
important applications, one of which is the detection of with machine learning and deep learning methods to
power theft in massive volumes of consumer data. For the identify and prevent the theft of electricity. The
smart grid to reap its benefits, reliable methods of researchers of this article [9] offer a method to recognize
identifying and stopping power theft are essential [7]. theft that makes use of a deep neural network-based
However, because of customers' stochastic behavior, classification strategy, based on extensive data in the
power consumption data is a complex signal. The spatial and frequency domains. Data interpolation and the
difficulty of detection may increase with the volume of generation of synthetic data can be used to compensate
consumption data. This challenge renders the status quo for shortcomings in the dataset including missing data and
of data analysis techniques inadequate. Deep learning has class imbalance problems. In order to validate the most
repeatedly been brought up in the academic literature promising choices, we first run experiments in the merged
when this topic is discussed. But there needs to be and reduced feature space using principal component
accurate models for detection. We provide a novel analysis, etc. to establish which characteristics are most
approach to deep learning optimization in this paper. The crucial. To determine the ideal settings that generate the
proposed approach utilizes a data pre-processing highest accuracy for energy theft identification and hence
technique and a finely-tuned neural network to achieve enhance the efficacy of the system, studies are done with
desirable outcomes. To evaluate the efficacy of the different values of important variables using a Bayesian
proposed method [7], an extensive dataset of real power algorithm and an adaptive moment estimation optimizer.
consumers is employed. Several well-known deep Finally, we demonstrate that the methodology is
learning approaches are outperformed by the proposed competitive with other methods that have been tested with
model [11]. the same dataset. We were able to attain an AUC of 97%,
which is 1% higher than the best AUC in previous
In order to create a "smart grid," many Internet of Things studies, and a validation accuracy of 91.8%, which is the
(IoT)-based devices are installed across power networks second highest on the benchmark [9].
[8]. However, many IoT devices include vulnerabilities
that may be exploited by hackers to gain control in a While machine learning approaches for data-driven ETD
number of digital and physical ways. As much as $96 in smart grids allowed by manufacturing informatics have
billion was lost to energy theft in 2017. Most current shown promising results [10], the high costs of manually
methods for identifying power theft either need a lot of labeling massive sensor information during learning data
resources to execute or have a low level of detection preparation have not been taken into account in any of the
accuracy. In light of these issues, we present a new existing studies. This paper provides a low-cost data-
Electricity Theft Detector (ETD-ConvLSTM) that makes driven ETD approach to addressing this issue by
use of a Convolutional Long Short Term Memory neural decreasing the time-consuming and resource-intensive
network. It is feasible to identify areas where malicious data labeling procedure without sacrificing ETD
users have taken root by establishing a central observer precision. In specifically, a clever deep active learning
meter [8]. In order to convert the time series of electricity (DAL) approach is employed to practically implement the
expenditures in these neighborhoods, which exhibit method in a methodical manner. The DAL approach
temporal associations, into spatio-temporal time sequence efficiently identifies the most relevant cases for training
information, a two-dimensional matrix of consumptions ETD predictions by integrating CNN learning with Monte
and use disparities spanning several neighboring days is Carlo dropout-based Bayesian active query. The
constructed. To create a ConvLSTM network, several suggested method might be used to construct a
stacked ConvLSTM nodes are employed, with every level trustworthy CNN-based ETD model at much reduced
formed by a sequence of temporally concatenated labeling expense [10]. The total effectiveness of the
ConvLSTM nodes fed through a series of sub-matrices. suggested method is demonstrated by practical test results
By combining global and local information, the ETD- on a real-world ETD dataset given by the State Grid
ConvLSTM technique is able to better identify periodicity Corporation of China [10].
in consumers' buying habits, leading to much higher

Authorized licensed use limited to: Kwame Nkrumah Univ of Science and Technology. Downloaded on March 07,2025 at 22:25:34 UTC from IEEE Xplore. Restrictions apply.
 III.METHODOLOGY Dataset
This study makes use of user-provided real-time SGCC
This section describes the dataset employed in this work EC data. Smart meters with built-in antennas collect the
and the suggested model. In addition, several methods information directly from the consumers. There are 1,035
explored in this paper. Figure 2 depicts the suggested features in the data collection. A sample of 3000
system model. customers' records is drawn from the larger dataset; of
these, 2480 are considered "normal" customers while the
Following is an explanation of each component of the remaining 520 are suspected of being fraudulent.
proposed CNNGRU-based model. The suggested system Evidence suggests that ETD is significantly impacted by
model depicts a series of steps beginning with data the dataset's obvious imbalance. In this study, data is
collection from smart meters equipped with intelligent normalized by subtracting the number of legitimate
antennas at the consumer end (bottom box). Each smart customers from the total number of customers served by
house is equipped with a smart meter and a smart antenna SMOTE. Moreover, 75% and 25% of the dataset are used
to aid in the collection of EC data. SGCC stores and for training and testing, respectively.
makes available the collected information. After that, the
dataset undergoes preprocessing to standardize the data Data Preprocessing
and eliminate any extraneous or unnecessary information. Removing outliers and missing data, standardizing the
Additionally, the data is refined by excluding the outliers. data's numerical values, extracting relevant features, and
Using a feature engineering technique, we can extract the dealing with class imbalance are all part of the data
most useful characteristics. Features are selected and preparation process. In the SGCC dataset, 'NAN' denotes
extracted throughout this procedure. Next, we separate missing data or a value that is not a number.
legitimate buyers from scammers.
The outcomes of ML and DL techniques suffer if there
are missing values in the data. The issue of under-fitting
Start might arise if the records with incomplete values are
omitted since this may result in the loss of crucial
information. Linear imputation is used to fill in the
missing variables and prevent under-fitting. In addition,
there are some extreme values in the data. In data
Data Acquisition
preparation, eliminating or handling outliers is a crucial
step. Experiments published in the literature reveal how
vulnerable ML and DL models are to data splitting and
erroneous outcomes. In this analysis, the outliers are
Pre Processing Data handled using the TSR (three-sigma rule).

Normalizing the Data Values

The data is then normalized using a min-max technique
SMOTE once the preceding stages have been completed. This is
because ML and DL struggle when presented with a wide
variety of data.

Feature Extraction (CNN) Class Imbalance Problem

If there are disproportionately more samples from one
class than another, a problem known as "class imbalance"
or "uneven distribution of class samples," would arise as a
major difficulty in ETD. Machine learning and deep
Model Training - GRU
learning that have been trained on an unbalanced samples
provide biased, high FPR outcomes. High FPRs are
intolerable for power production businesses due to a lack
of personnel and equipment for regular on-site
Classify Data inspections, as has been previously discussed in the
literature. In the literature, class imbalance issues are
often addressed using either under-sampling or over-
sampling. Rather of removing samples to achieve a
End balance, as in the later method, minority class duplicates
are created in the former. The problems with both
Figure 2. Proposed System Model. approaches are, however, overfitting, redundant data, and
data loss.

Authorized licensed use limited to: Kwame Nkrumah Univ of Science and Technology. Downloaded on March 07,2025 at 22:25:34 UTC from IEEE Xplore. Restrictions apply.
 Feature Extraction forgotten. This information is meaningless when seen in
It is only when the data has been cleaned and normalized the broader context of things. These GRU layers have the
that the feature engineering process may begin. There are potential to extract temporally-related patterns when they
two stages to this procedure: selecting features and are given EC data.
extracting those features. In the first instance, the most
important characteristics are selected from the whole IV.RESULTS AND DISCUSSIONS
dataset to minimize overfitting as well as training time
while increasing accuracy, while in the latter, they are The performance of the hybrid model is examined
used to achieve decreased dimensionality and eradicate through a series of simulations. The simulation outcomes,
redundant data. Better model performance is achieved efficacy metrics, and similarities to baseline models are
through feature engineering. discussed. Here, we apply a number of metrics for
effectiveness to the dataset to evaluate the effectiveness of
Proposed Model - CNNGRU the proposed model. We also look at how the suggested
Convolution, pooling, and fully linked layers are only model stacks up against several industry standards in
some of the usual components of a CNN's design. In the terms of output.
first convolution layer, multiple convolution filters, also
known as kernels, carry out the actual mapping. Performance Metrics.
Immediately following the convolution layer is the Since there are more samples from the "normal" category
pooling layer in a CNN architecture. It is the job of this than the "malignant" category, unequal distribution of
layer to down-sample each feature map in order to class samples is a major issue in ETD. For this reason,
accomplish dimensionality reduction and to extract the training an ML or DL model on this kind of data often
relevant features. Moreover, the network's processing leads to inaccurate findings or false alarms since the
time is shortened. When it comes to feature maps, there model favors examples from the majority class and
are two types of pooling functions: maximum pooling, disregards those from the minority class. According to the
which determines the maximum number of ideal features, research, electric companies cannot afford to respond to
and average pooling, which determines the average false alarms because of their limited on-site testing
number of features. The final categorization in a CNN is capabilities. The test data are not balanced, despite the
carried out by the third completely linked layer. Data in fact that the training data were sampled using the
ETD is divided into two groups: legitimate businesses and recommended method. In order to compare the results of
dishonest ones. the reference and suggested models, we require reliable
effectiveness metrics. In this study, we evaluate our
Whereas the CNN is responsible for retrieving more model's performance using the aforementioned accuracy,
concrete or overt patterns, the GRU is responsible for F1 score, recall, ROC-AUC, and PR-AUC metrics. We
extracting time-related patterns. The suggested model is employ a confusion matrix, which is a table with columns
made up of the GRU and the CNN. The GRU module is for true negative (TN), true positive (TP), false negative
fed one-dimensional data, while the CNN is fed two- (FN), and false positive (FP) outcomes, to derive the
dimensional data, such that both may learn abstract aforementioned metrics.
characteristics. The input characteristics from both
modules are used by the hybrid module, which then uses All of these metrics, including loss and accuracy, are
the combined features to identify dangerous and benign taken into account during the simulations used to assess
patterns. the quality of the suggested and reference models.
Additionally, the models are combined for analysis of
To put it another way, GRU may be seen of as an results. The SVM and LR both get an ACU score of 0.86.
upgraded RNN. The phenomenon of vanishing gradients With the kernel approach for handling nonlinear data,
is a severe defect in RNNs that prevents the model from SVM achieves a better AUC than LR. In comparison,
learning any more information and forces sequential DL LR's AUC is just 0.84% because its single hidden layer
algorithms into local optima. This issue also prevents struggled with high dimensional data and became trapped
RNNs from being used in machine learning. In response in the region of minimum. You may find SVM and LR
to the problem described above, the GRU technique was performance data in Tables 1 and 2, respectively. Using
created. The manner in which the network learns temporal its convolutional and pooling layers, the CNN is trained to
patterns from EC data is influenced by both the update identify and pull out abstract and latent properties from
gate and the reset gate that are included in the EC data. LSTM, on the other hand, can identify temporal
construction of the GRU. Update gates are basically patterns within consumer data and categorize them as
decision-making processes that determine which bits of either normal or abnormal.
data should be transmitted to following layers or units.
Update gates may be found in many computer systems. In
the event that this is not the case, the reset gate will
determine the extent to which the past should be

Authorized licensed use limited to: Kwame Nkrumah Univ of Science and Technology. Downloaded on March 07,2025 at 22:25:34 UTC from IEEE Xplore. Restrictions apply.
 Table 1. SVM Metrics for Performance Analysis Modeling using Genetic Algorithms. The GA is an
Models Results example of a metaheuristic approach. Its effectiveness is
AUC 86% measured in terms of the indicators listed in Table 5.
F1-Score 86.23% Based on the findings, GA outperforms the standard
Accuracy 85.48% SVM, LR, LSTM, and GRU models. As a result of its
Precision 87.91% superior ability to learn, GA is also found to be more
Recall 87.45% accurate and robust than the standards.
Table 5. Measures of CNN-GRU's Effectiveness
Formula for Logistic Regression. LR is a well-liked Models Results
classifier that serves dual purposes in both categorization AUC 96.32%
and regression. It is also known as an ETD in the F1-Score 94.55%
scholarly community. In this study, we evaluate LR's Accuracy 96.16%
effectiveness with respect to the aforementioned Precision 95.72%
parameters. As can be seen in Table 2, LR does poorly in Recall 96.04%
comparison to the reference methods. The problems with
overfitting and LR's inability to deal with skewed ETD Accuracy comparisons of several models are displayed in
data are to blame. Figure 3. Figures 4 and 5 show the accuracy as well as
loss values of the merged CNNGRU model on the
Table 2. LR's Key Performance Indicators training as well as testing datasets, correspondingly.
Models Results Optimal features are extracted using a CNN, and the GRU
AUC 84% model handles the classification process. Extracted high
F1-Score 83.97% variance features are processed using CNN to find the
Accuracy 84.92% most important data, and the GRU model's reset and
Precision 82.71% update gates help to filter out noise and superfluous data.
Recall 84.65% Thus, CNN-GRU outperforms CNN-LSTM after
undergoing this procedure. There is an overfitting issue
The Model of Long-Short Term Memory. In ETD, the with the CNN-GRU model, as evidenced by the fact that
DNN model long short-term memory (LSTM) is the accuracy and loss curves for the training and testing
commonly utilized for feature extraction and datasets are different by 4% and 6.97%, respectively. If
categorization. The effectiveness of LSTM is evaluated the hyperparameters are not properly tuned, the model
using the aforementioned standards. Table 3 displays the will perform well on observed data but poorly on unseen
findings, which reveal that LSTM outperforms SVM and data, a phenomenon known as overfitting.
LR but is inferior to the suggested model. Since the
suggested model does not experience the performance
degradation caused by the overfitting problem, it
outperforms LSTM.

Table 3. Performance Evaluation Metrics of LSTM

Models Results
AUC 91.27%
F1-Score 90.65%
Accuracy 92.97%
Precision 91.85%
Recall 91.72%
Figure 3. Accuracy Comparison over Various Model

The Model of Gated Recurrent Neurons. A DNN model,

like GRU. It's a step up from the LSTM algorithm. Table
4 displays the findings, which are superior to the
standards. This indicates that GRU can deal with
unbalanced data without resorting to overfitting.
Table 4. Metrics for Assessing GRU Performance
Models Results
AUC 94.51%
F1-Score 93.48%
Accuracy 93.78%
Precision 94.92%
Recall 94.23% Figure 4. Model Validation Accuracy

Authorized licensed use limited to: Kwame Nkrumah Univ of Science and Technology. Downloaded on March 07,2025 at 22:25:34 UTC from IEEE Xplore. Restrictions apply.
 REFERENCES
[1] Noor Mahmoud Ibrahim, et al., "Electricity-theft detection in
smart grids based on deep learning", Bulletin of Electrical
Engineering and Informatics, DOI: 10.11591/eei.v10i4.2875,
2021.
[2] Yuefei Sun, et al., "Smart Grid Theft Detection Based on
Hybrid Multi-Time Scale Neural Network", Appl. Sci., DOI:
10.3390/app13095710, 2023.
[3] Md.NazmulHasan, et al., "Electricity Theft Detection in Smart
Grid Systems: A CNN-LSTM Based Approach", Energies,
DOI: 10.3390/en12173310, 2019.
[4] A. G, V. Mohanavel, M. Tamilselvi, G. Ramkumar and R. T.
Prabu, "An Intelligent LoRa based Women Protection and
Safety Enhancement using Internet of Things," 2022 Sixth
Figure 5. Model Validation Loss International Conference on I-SMAC (IoT in Social, Mobile,
Analytics and Cloud) (I-SMAC), Dharan, Nepal, 2022, pp. 43-
48, doi: 10.1109/I-SMAC55078.2022.9987425.
V.CONCLUSION AND FUTURE SCOPE [5] G. R and M.M, "Uncompressed digital video watermarking
using stationary wavelet transform," 2014 IEEE International
Power losses due to electricity theft affect both developed Conference on Advanced Communications, Control and
and developing nations. Because of this, electricity utility Computing Technologies, Ramanathapuram, India, 2014, pp.
1252-1258, doi: 10.1109/ICACCCT.2014.7019299.
firms experience significant disruptions in their [6] LelokoJ.Lepolesa, ShaminAchari and Ling Cheng, "Electricity
operations, which results in a drop in income. In addition, Theft Detection in Smart Grids Based on Deep Neural
the economy and the reliability of the power grid both Network", IEEE Access, DOI:
suffer when electricity is wasted. In this research, we 10.1109/ACCESS.2022.3166146, 2022.
[7] 1.M. T. Ahammed et al., "Applied Informatics in Renewable
introduce CNNGRU, a DL model that combines a GRU Energy Grid with Dynamic Reactive Power Configuration
with a CNN. The suggested model involves a few Based on a Probability Assessment of Transient," 2022
different procedures. The preprocessing phase involves International Conference on Advances in Computing,
removing outliers and filling in missing numbers in Communication and Applied Informatics (ACCAI), 2022, pp. 1-
6, doi: 10.1109/ACCAI53970.2022.9752520.
addition to standardizing the raw data. The local average [8] Xiaofang Xia, et al., "ETD-ConvLSTM: A Deep Learning
strategy and the min-max normalization procedure do the Approach for Electricity Theft Detection in Smart Grids", IEEE
preprocessing. After that, convolutional neural networks Transactions on Information Forensics and Security, DOI:
(CNN) are used for feature engineering. After collecting 10.1109/TIFS.2023.3265884, 2023.
[9] Leloko.J.Lepolesa, et al., "Electricity Theft Detection in Smart
and normalizing the best data possible, categorization is Grids Based on Deep Neural Network", IEEE Access, DOI:
completed with the help of a GRU integrated CNN. Here, 10.1109/ACCESS.2022.3166146, 2022.
we separate legitimate buyers from those trying to pull a [10] Lipeng Zhu, et al., "Deep Active Learning-Enabled Cost-
fast one. Accuracy, recall, precision, area under the curve Effective Electricity Theft Detection in Smart Grids", IEEE
Transactions on Industrial Informatics, DOI:
(AUC), and F1-score are only few of the performance 10.1109/TII.2023.3249212, 2023.
indicators utilized to verify the validity of the suggested [11] D. Rose, K. Vijayakumar, D. Kirubakaran, R. Pugalenthi and G.
model. The findings of the comparisons reveal that the Balayaswantasaichowdary, "Neural Machine Translation Using
proposed hybrid model is superior to the existing methods Attention," 2023 International Conference on Artificial
Intelligence and Knowledge Discovery in Concurrent
in terms of ETD efficiency, accuracy, robustness, and Engineering (ICECONF), Chennai, India, 2023, pp. 1-7, doi:
generalizability. Regardless, our approach is superior to 10.1109/ICECONF57129.2023.10083569.
existing methods and will lead to more effective ETD.
Because the modules of the proposed model are integrated
sequentially (CNNGRU), it has a little greater
computational cost. Hyperparameter tuning for DL
models will be optimized in the future with the help of
meta-heuristic approaches.

Authorized licensed use limited to: Kwame Nkrumah Univ of Science and Technology. Downloaded on March 07,2025 at 22:25:34 UTC from IEEE Xplore. Restrictions apply.