Development of an Integrated AI Model Based on CNN-SVM for
Electricity Theft Detection
NENCHIN EMMANUEL, ADEMOH A. ISAH
Department of Electrical and Electronics Engineering,
Federal University of Technology Minna,
Minna, Niger State,
NIGERIA
Abstract: - This research presents the development and implementation of an integrated artificial intelligence
model for electricity theft detection, combining Convolutional Neural Networks (CNN) and Support Vector
Machines (SVM). The primary objective was to create a more accurate, efficient, and scalable method for
identifying fraudulent electricity consumption patterns. Our CNN-SVM hybrid model leverages CNNs for
automatic feature extraction from complex consumption data and SVMs for effective classification. This synergy
allows for superior performance in detecting subtle anomalies indicative of electricity theft. The methodology
involved pre-processing a large dataset of electricity consumption records, training the CNN to extract relevant
features, and optimising the SVM classifier to distinguish between normal and fraudulent patterns. We evaluated
the model's performance using metrics including accuracy, precision, recall, F1-score, and ROC AUC. Results
demonstrated that our integrated CNN-SVM model significantly outperformed conventional machine learning
techniques and standalone models in electricity theft detection. The model achieved an accuracy of 96.6%, with
a precision of 97.2% and a recall of 96.1%. Comparative analysis against other state-of-the-art algorithms
revealed consistently superior performance across all evaluation metrics. To enhance practical applicability, we
developed and deployed a web application that implements the model, allowing for user-friendly interaction and
real-time theft detection. This addition bridges the gap between research and real-world implementation,
providing utility companies with an accessible tool for fraud detection. The study also explored the model's
potential for real-time application and scalability to large-scale utility operations. Our findings suggest that the
computational efficiency of the CNN-SVM model, coupled with the web application, offers utility companies a
powerful and accessible tool for rapid response to potential fraud. This research contributes to the field of
electricity theft detection by introducing a novel, high-performance AI model with a practical web-based
implementation. The proposed approach not only improves detection accuracy but also offers potential for
immediate real-world application, paving the way for more effective fraud prevention in the utility sector.
Key-Words: - Electricity Theft Detection, Artificial Intelligence, Machine Learning, Deep Learning, CNN, SVM
Received: April 16, 2024. Revised: September 28, 2024. Accepted: November 9, 2024. Published: December 10, 2024.
1 Introduction
Electrical energy theft or non-technical loss (NTL) is
the unlawful usage of energy from the grid and is
performed without accurate metering and payment,
[1]. This poses a major problem for electricity
distribution companies who have significant losses
and unstable networks, with losses deprioritising
investments in upgrading the grid infrastructure, [2],
[3]. According to Onat in 2008, 14.4% of Turkey’s
electricity was used illegally, costing approximately
$895.3 million, [4]. The value of electricity stolen
annually in relatively poorer developing countries
accounts for billions of dollars in losses annually, [5],
as said by Messinis and Hatziargyriou.
Electrical energy theft has a significant effect on
Nigeria's power sector sustainability as described by
Tsado and Abel in 2022, and acts as a bottleneck to
nation building strides in Nigeria, [6]. Now, this
rampant theft not only robs genuine consumers of
quality power supply but also disincentivizes capital
inflows to upgrade and expand the grid, ensuring this
cycle to go on indefinitely, [7].
Manual metre reading and physical inspections are
the primary ways in which electricity theft is
traditionally detected. These approaches are slow,
manpower eating and not always successful in
finding high-level theft methods implemented by
thieves. Moreover, some new methods presented for
bypass detection problems are metre tampering,
Engineering World
DOI:10.37394/232025.2024.6.27
Nenchin Emmanuel, Ademoh A. Isah
E-ISSN: 2692-5079
246
Volume 6, 2024
unauthorised load manipulation and bypass
connections, [8], [9]. However, that abridgment is too
good to be true and manual inspections may miss
important regions throughout a distribution network
detecting the majority of cases of theft.
The development of smart grids has allowed utilities
and customers to communicate in both directions,
completely changing the way power is distributed.
Compared to standard metres, smart metres as an
essential part of Advanced Metering Infrastructure
(AMI) offer a more precise and detailed picture of
energy usage by providing real-time data on patterns
of electricity consumption, [3]. This data can be used
to train machine learning (ML) and deep learning
(DL) algorithms to precisely identify electricity theft
in power grids, [10].
The limitations of traditional detection methods
highlight the critical need for more sophisticated and
automated approaches to identify electricity theft. In
this regard, AI algorithms have shown promise as a
potential remedy. It can analyse vast amounts of data
from smart metres and identify minute irregularities
that could be signs of theft [10]. Artificial
intelligence methods, including deep learning
algorithms, which have the ability to identify intricate
patterns in past consumption data and apply them to
categorise current readings as either normal or
suspect. When compared to manual procedures, this
automated approach can greatly increase detection
efficiency and accuracy. While various ML and DL
models have been explored for electricity theft
detection, several limitations remain. Traditional ML
methods often rely on one-dimensional (1D)
consumption data, failing to capture the periodicity
inherent in electricity usage patterns, [11].
Additionally, imbalanced datasets, where legitimate
users significantly outnumber electricity thieves, can
hinder model performance in accurately classifying
the minority class (theft), [12].
DL algorithms are very good at finding patterns in
large, complicated data sets, which makes them ideal
for assessing the vast quantities of data produced by
smart metres, [8], [12]. By processing consumption
data across different time intervals, these algorithms
can capture temporal patterns, consumption spikes,
and other irregularities indicative of theft activities
[11]. Also, the requirement for labour-and resource-
intensive human feature engineering is eliminated by
deep learning models' ability to automatically extract
relevant features from the data [10].
The effective development and deployment of AI-
driven systems for detecting electricity theft present
various notable advantages for both power
distribution companies and consumers:
Decreased Revenue Losses: Through precise
identification of electricity theft, power distribution
companies can reduce monetary losses and enhance
their revenue flow. This financial gain can then be
channelled into grid enhancements and growth,
ultimately resulting in a more dependable and
effective power provision for all consumers [2], [7].
Improved Grid Stability: Early detection of theft
activities can help prevent overloading of the grid and
ensure stable power supply for legitimate consumers.
Unaddressed theft can lead to power outages and
disruptions, impacting businesses, homes, and
critical infrastructure [6].
Enhanced Efficiency: AI-based systems can
automate the detection process, freeing up manpower
for other critical tasks within the power distribution
company. Previously, manual metre reading and
inspection required a significant investment of
human resources. Automating theft detection allows
utilities to deploy personnel more effectively for
maintenance, customer service, and grid
infrastructure improvement projects [3].
Data-driven Decision Making: The insights gleaned
from the model's predictions can inform more
targeted and effective strategies for curbing
electricity theft [13]. Utilities can use the data to
identify areas with high theft prevalence and deploy
targeted intervention efforts. In addition, the data can
also be utilised to create consumer education
programmes that emphasise the negative effects of
electricity theft to the grid [13].
Despite the advancements in AI-based electricity
theft detection, several challenges remain
unaddressed;
Data Availability: Obtaining sufficient labelled data
for training complex AI models can be a significant
hurdle. Labelled data refers to data points where the
consumption pattern has been categorised as either
normal or indicative of theft. Due to privacy issues,
utilities might be reluctant to disclose customer data.
Data augmentation techniques and transfer learning
can be looked into to overcome this problem, [14].
Data Privacy: Data privacy concerns are paramount
when dealing with customer consumption data.
Sensitive information about a consumer's daily
activities can be inferred from their electricity usage
patterns, [15]. Anonymization and privacy-
preserving strategies must be used to guarantee
adherence to applicable laws, [11].
Class Imbalance: Electricity theft data might be
imbalanced, with a significant majority of readings
representing normal consumption patterns and a
smaller portion reflecting theft activities. This
imbalance can hinder the training process of AI
models [12]. This issue could be solved by
Engineering World
DOI:10.37394/232025.2024.6.27
Nenchin Emmanuel, Ademoh A. Isah
E-ISSN: 2692-5079
247
Volume 6, 2024
employing strategies like oversampling or under-
sampling the data.
Evolving Theft Techniques: As detection methods
improve, perpetrators may adopt more sophisticated
theft techniques to circumvent them [11]. The AI
model needs to be continuously updated and adapted
to identify new and emerging theft patterns.
Several machine learning and deep learning
approaches have been explored for electricity theft
detection, with varying degrees of success.
2 Related Works
The detection of electricity theft has evolved
significantly over the past two decades, progressing
from simple meter-based detection to sophisticated
AI-driven approaches. Early pioneering work by
Hernandes Jr et al., in 2001 established the
foundation for non-invasive theft detection by
developing an electronic Ah meter device for
comparing customer consumption patterns, [16].
Their comprehensive statistical study, involving over
80,000 customers in São Paulo, Brazil, demonstrated
the effectiveness of consumption pattern analysis in
identifying tampered or defective meters, achieving
significant improvement in inspection efficiency.
Building on this statistical approach, Nagi et al.,
introduced one of the first applications of Support
Vector Machines (SVM) for electricity theft
detection in 2008. Their work with Tenaga Nasional
Berhad in Malaysia demonstrated how machine
learning could effectively analyse customer load
profiles to expose abnormal behaviour correlated
with Non-Technical Loss (NTL) activities, [17]. This
marked a crucial shift towards automated
classification approaches, achieving superior results
compared to traditional inspection methods. Stajić et
al., further contributed to this evolution by
developing an interoperable smart grid platform in
Serbia, emphasizing the importance of
comprehensive monitoring systems in loss detection
and establishing the groundwork for modern AI-
based approaches, [18].
More recent developments have focused on deep
learning approaches. Zheng et al., introduced a wide
and deep CNN model for electricity theft detection in
smart grids, [19]. Their model comprises two
components: a deep CNN component for capturing
nonperiodic theft patterns and a wide component for
extracting global features from electricity
consumption data. By leveraging 2-D data
representation, their approach demonstrated superior
performance in detecting theft activities compared to
existing methods. Hasan et al., proposed a CNN-
LSTM hybrid model tailored for smart grid data
classification, [20]. LSTM architecture addresses the
time-series nature of power consumption data, while
CNN automates feature extraction and classification
processes. Their work emphasises the importance of
data pre-processing, including interpolation and
outlier handling, to enhance model accuracy.
Additionally, they employed synthetic data
generation to mitigate class imbalance issues,
achieving satisfactory results in identifying theft
users. However, LSTMs can be computationally
expensive to train and may require large datasets for
optimal performance. Fang et al., introduced a LSTM
and a modified CNN to predict electricity usage
patterns and detect abnormalities, [21]. The authors
extract relevant features that affect meter error, such
as voltage and current readings at different intervals.
They included polynomial fitting to model the error
values in electricity measurements. By comparing
different polynomial degrees, the authors identify the
best fit for the data, which helps in understanding the
underlying patterns and trends. An LSTM model with
40 neurons in the first hidden layer and using the root
mean square error (RMSE) as the loss function was
used. The model was trained for 1000 epochs. the
authors use a sliding window approach to identify
days with significant deviations between predicted
and actual values. This method not only helps in
detecting anomalies but also helps in identifying
faulty meters that need replacement. The authors also
compare different models and found that the time
series recurrence plot CNN (TS-RP CNN) performs
best in detecting anomalies, achieving an accuracy of
about 82%. Yao et al., introduced a hybrid method,
AdaBoost-CNN, combining adaptive boosting
(AdaBoost) and CNN for electricity theft detection,
[22]. Multiple CNN-based classifiers were trained to
extract diverse features from consumption data,
which were then aggregated by AdaBoost to enhance
classification performance. Their experimental
results, based on real smart energy data,
demonstrated the superiority of the hybrid classifier
over conventional methods in detecting theft
activities. Singh and Venkaiah in 2023, addressed the
challenge of data imbalance in theft detection by
proposing a multi-layer model classifier, [23]. Their
approach involves data preparation, including
interpolation and outlier handling, followed by data
balancing techniques such as AdaSys. A two-layer
model, comprising heterogeneous machine learning
models and an ANN, demonstrated improved
performance in identifying theft users on real
consumption datasets. Mazid et al., proposed a hybrid
approach combining principal component analysis
(PCA) and CNN for power theft detection, [24].
Their method involves feature selection, extraction,
Engineering World
DOI:10.37394/232025.2024.6.27
Nenchin Emmanuel, Ademoh A. Isah
E-ISSN: 2692-5079
248
Volume 6, 2024
and classification stages applied to smart metre data.
By leveraging optimised hyperparameters and CNN-
PCA methodology, their approach achieved high
accuracy rates, outperforming previous methods.
Zhou et al., addressed the challenge of sparse and
imbalanced data in low-voltage networks by
proposing a CNN and data augmentation method for
theft detection, [25]. Their approach utilises kernel
density estimator (KDE) and Monte Carlo method for
data expansion, followed by CNN classification.
Experimental results confirmed the effectiveness of
their method in achieving high performance metrics.
Ibrahim et al., presented a CNN-based approach for
electricity theft detection in smart grids, [26]. Their
work focused on feature reduction using the Blue
Monkey (BM) algorithm to enhance classifier
performance. By designing high-performance signal
classifiers, their approach demonstrated promising
results in identifying theft activities. Dimf et al.,
proposed a bi-LSTM and CNN hybrid model for theft
detection, incorporating various techniques such as
data pre-processing, synthetic data generation, and
feature selection, [27]. Their approach achieved high-
quality results comparable to existing methods,
highlighting the effectiveness of combining CNN and
bi-LSTM architectures. Khan et al., addressed
challenges in electricity theft detection using
supervised learning techniques on smart metre data,
[15]. Their proposed model combines Adasyn
algorithm for class imbalance, VGG-16 module for
feature extraction, and FA-XGBoost for
classification. Simulation results demonstrated
superior performance in handling large time series
data and accurate theft detection. Abel et al.,
proposed a matrix converter-based solution to
mitigate electricity theft at low distribution voltages,
[28]. By focusing on frequency variation and Total
Harmonic Distortion (THD), their approach aimed to
eliminate metre bypassing theft, presenting a novel
solution to complement smart metering systems.
Ullah et al., introduced a hybrid deep neural network
model combining CNN, particle swarm optimization,
and gated recurrent unit for electricity theft detection,
[29]. Their approach addressed issues of overfitting
and data imbalance, achieving robustness, accuracy,
and generalisation in theft detection tasks.
The evolution of electricity theft detection techniques
reveals a clear trend toward increasingly
sophisticated machine learning approaches, with
particular emphasis on improving classification
accuracy through hybrid models and advanced data
pre-processing techniques. While early methods
relied on simple statistical comparisons, modern
approaches leverage deep learning architectures to
capture complex patterns in consumption data.
However, challenges remain in balancing
computational efficiency with classification
accuracy, particularly when dealing with imbalanced
datasets and real-time detection requirements. These
challenges present opportunities for further research
in developing more efficient and accurate detection
methods.
3 Methodology
This research proposes an integrated AI model that
combines the strengths of CNNs and SVMs to
address the limitations of existing techniques. This
chapter discusses a breakdown of the proposed
method. The proposed method is divided into three
phases; (1) data pre-processing, (2) feature
extraction, and (3) classification phase. Figure 1
shows the workflow of the proposed method.
Fig. 1: Workflow of the proposed method
3.1 Data Pre-processing
The main aim of pre-processing is to verify the
quality of the data to be used and to transform the data
into usable format, [30], [31]. The pre-processing
steps involve data cleaning (outlier removal,
duplicate removal and filling of missing values) and
data normalisation. Outlier removal was done by
applying the IQR technique. The filling of missing
values is done using the SimpleImputer method and
the mean strategy. Oversampling involves replicating
data points from the minority class (theft) to create a
more balanced dataset. Data normalisation was done
by scaling the features of the dataset to a standard
range using Sklearn’s StandardScaler method to
ensure uniformity and prevent dominance by certain
features. Data pre-processing is very important
because the model’s efficiency is also dependent on
the quality of the data [32].
3.2 Feature Extraction
Engineering World
DOI:10.37394/232025.2024.6.27
Nenchin Emmanuel, Ademoh A. Isah
E-ISSN: 2692-5079
249
Volume 6, 2024
Extracting relevant features from raw data helps in
improving the model’s performance and accuracy. In
this project feature extraction was done using CNN.
3.2.1 Detailed Feature Extraction Architecture of
CNN
The feature extraction architecture of a
Convolutional Neural Network (CNN) comprises
several key components, including convolutional
layers, pooling layers, activation functions, and fully
connected layers. Each component plays a crucial
role in the network's ability to learn hierarchical
features from input data.
The convolutional layers consist of filters that
convolve across the input data to extract features.
Each filter detects patterns or features in the input
data through element-wise multiplications followed
by summation operations. Mathematically, the output
of a convolutional layer can be expressed as:
(1)
Where * denotes the convolution operation, f is the
activation function (ReLU), and b represents the bias
term, [33].
Activation functions introduce non-linearity into the
network, enabling it to learn complex patterns. The
Rectified Linear Unit (ReLU) is commonly used due
to its computational efficiency and ability to mitigate
the vanishing gradient problem:
ReLU(x) = max(0, x) (2)
Pooling layers down-sample the feature maps
obtained from the convolutional layers, reducing
their spatial dimensions while retaining important
features, [34]. This helps in reducing computational
complexity, providing a form of translation
invariance, controlling, and overfitting, [34].
Common pooling operations include max pooling
and average pooling.
After passing through the convolutional and pooling
layers, the extracted features reside in a multi-
dimensional format. The flatten layer transforms this
data into a single 1D vector suitable for feeding into
the next layer, [35].
The fully connected layers connect every neuron in
one layer to every neuron in the next layer, enabling
high-level feature learning and classification. The
output of a fully connected layer can be expressed as:
y = f(Wx + b) (3)
Where W is the weight matrix, x is the input vector,
b is the bias vector, and f is the activation function,
[33].
The output of the last fully connected layer is fed into
a sigmoid activation function for binary classification
into different classes, [35]:
(5)
For multi-class classification, a softmax function
may be employed, [35]. The overall CNN
architecture integrates these components in a
sequential manner, allowing for end-to-end learning
of features and classification. Figure 2 illustrates a
typical CNN architecture, showcasing the
arrangement of these layers.
Fig. 2: Typical CNN architecture
3.3 Classification
When presented with a new data point (representing
an unseen consumption sequence), the CNN extracts
features and presents them to the SVM as shown in
figure 3. Based on the hyperplane and the support
vectors learned during training, the SVM classifies
the new data point as either normal consumption or
potential theft.
Fig. 3: Framework of proposed model
3.3.1 Detailed Architecture of SVM
Support Vector Machines (SVMs) are supervised
learning models commonly used for classification
and regression. The primary objective of SVM is to
find an optimal hyperplane that maximizes the
margin between two classes in a dataset, [35].
In SVM, the goal is to maximize the margin, or
distance, between the separating hyperplane and the
closest data points from each class. For linearly
separable data, this margin maximization is
formulated as a convex optimization problem, [36].
The hyperplane can be defined as:
f(x)=wx+b (4)
Engineering World
DOI:10.37394/232025.2024.6.27
Nenchin Emmanuel, Ademoh A. Isah
E-ISSN: 2692-5079
250
Volume 6, 2024
where w is the weight vector, and b is the bias term.
The optimization goal for a hard-margin SVM is to
minimize the norm w2, which directly maximizes
the margin. This leads to the primal optimization
problem:
Solving SVMs in their dual formulation often
simplifies computations, especially with high-
dimensional data. The dual formulation uses
Lagrange multipliers to reformulate the objective
in terms of dot products between input vectors, which
enables the kernel trick. The kernel trick allows
SVMs to implicitly compute the dot product in the
higher-dimensional feature space without explicitly
transforming the data, [37]. This makes SVMs
computationally efficient for high-dimensional data.
The dual problem is given by:
subject to:
This dual form allows us to apply kernel functions,
which compute dot products in a high-dimensional
feature space without explicitly transforming the
data, reducing computational complexity.
Kernel functions are used by SVMs to transform
input data into higher-dimensional feature spaces,
where linear separation becomes possible [35].
Common kernel functions include:
Linear kernels: Suitable for linearly separable
data and high-dimensional spaces.
Polynomial kernels: Useful for data with
complex relationships.
Radial basis function (RBF) kernels: Effective
for non-linearly separable data.
Sigmoid kernels: It is used as an alternative to the
RBF kernel, often used in neural networks; less
commonly applied in SVMs.
To handle complex data distributions, SVM employs
a non-linear mapping that transforms the input data
into a high-dimensional feature space. In this feature
space, a linear decision boundary can be identified,
which corresponds to a non-linear boundary in the
original input space, [37]. By using kernel functions,
SVM avoids the computational cost of explicitly
mapping data, making it both efficient and powerful
for non-linear data [35].
In summary, SVMs are effective in high-dimensional
and non-linear settings due to the combination of
kernel functions, dual formulation, and support
vector optimization, providing an adaptable solution
for tasks like electricity theft detection. This
adaptability enables SVMs to classify data accurately
by maximizing the margin between classes, thus
ensuring robust performance across various
applications. Figure 4 presents the SVM
classification architecture.
Fig. 4: SVM classification architecture
3.4 Model Evaluation
After training, the integrated model underwent
evaluation to assess its performance in detecting
electricity theft.
Various performance metrics such as accuracy,
precision, recall, F1-score, and ROC-AUC were used
to measure the model's effectiveness in
distinguishing between normal and theft activities.
The equation of accuracy, precision, recall and f1-
score are given below.
A confusion matrix was generated to visualise the
model's classification results and identify any
Engineering World
DOI:10.37394/232025.2024.6.27
Nenchin Emmanuel, Ademoh A. Isah
E-ISSN: 2692-5079
251
Volume 6, 2024
misclassifications or errors. Confusion matrix
divides the dataset into four basic segments: true
positive (TP), false positive (FP), true negative (TN)
and false negative (FN), [38]. TP and TN show the
correctly predicted positive and negative samples
whilst FP and FN show the falsely classified negative
samples as positive and positive samples as negative,
respectively, [38]. ROC curve which analyses the
trade-off between true positive rate and false positive
rate was also plotted. The true positive rate and false
positive rate are expressed below.
3.5 Deployment
The model was deployed to the cloud using a Python
web framework, Streamlit, after satisfactory
evaluation and hyperparameter tuning. Streamlit is an
open source Python for building interactive web
applications and for easy deployment of machine
learning models.
4 Result
The model's performance is demonstrated by the
experimental results. We evaluated the performance
with different performance metrics. The model’s
performance was compared to the performance of
other machine learning models using the same
dataset.
4.1 Experimental Results
Using evaluation measures like accuracy, recall,
precision, F1-score, and ROC AUC score, we
assessed the effectiveness of the proposed approach.
The percentage of all subjects that were correctly
classified is referred to as accuracy. Recall is the
percentage of those who test positive and actually
have the condition. The number of the subjects
accurately classified as positive out of all those
classified as positive is known as precision. A
harmonic mean of recall and precision is the F1-
score. ROC AUC score shows how well the classifier
distinguishes positive and negative classes. Table 1
shows the performance score of the model for
different metrics.
Table 1. Performance of Proposed model
Table 1 shows the performance score of our proposed
model for different metrics. As shown, the model
achieves high scores across all metrics, with an
accuracy of 0.966 and a ROC-AUC of 0.976,
indicating strong overall performance and
discriminative ability.
Fig. 5: Confusion Matrix of test set prediction result
of proposed model
The classification results of the model proposed from
the figure 5, which is the confusion matrix are as
follows: the number of TPs, FNs, FPs, and TNs is
9679, 283, 394, and 9644, respectively. Figure 6
shows the roc curve.
Fig. 6: ROC curve of proposed model
Metric
Score
Accuracy
0.966
Recall
0.961
Precision
0.972
F1
0.966
ROC AUC
0.976
Engineering World
DOI:10.37394/232025.2024.6.27
Nenchin Emmanuel, Ademoh A. Isah
E-ISSN: 2692-5079
252
Volume 6, 2024
4.2 Comparative Analysis
To evaluate the effectiveness of our proposed hybrid
CNN-SVM model, we conducted a comprehensive
comparative analysis against several state-of-the-art
machine learning and deep learning models. Table 2
presents the ROC AUC scores for different methods,
while Figure 7 illustrates the performance metrics
across all comparative experiments.
Table 2. Performance comparison with other models
As evident from Table 2, our proposed CNN-SVM
model achieves the highest ROC AUC score of
0.976, outperforming all other models. The next best
performer is XGB (Extreme Gradient Boosting) with
a ROC-AUC of 0.964, followed closely by LR
(Logistic Regression) at 0.963 and RF (Random
Forest) at 0.962. The standalone CNN and SVM
models show lower performance with ROC AUC
scores of 0.94 and 0.89, respectively.
Figure 7 provides a more detailed comparison across
multiple performance metrics
Fig. 7: Comparison with other models
The superior performance of the proposed CNN-
SVM model is evident across all metrics. It achieves
the highest scores in accuracy (0.966), precision
(0.972), F1 score (0.966), and ROC AUC (0.976).
The model's recall (0.961) is slightly lower than the
standalone CNN (0.98), but this is compensated by
its significantly higher precision, resulting in a better
overall F1 score.
The effectiveness of our hybrid approach is further
emphasised when comparing it to the standalone
CNN and SVM models. The CNN-SVM model
outperforms both in all metrics except recall, where
the standalone CNN shows a marginally higher score.
This suggests that the hybrid model successfully
leverages the strengths of both techniques while
mitigating their individual weaknesses.
Among the traditional machine learning models,
XGB and LR show competitive performance,
particularly in terms of accuracy and ROC AUC.
However, they fall short of the CNN-SVM model
across all metrics. The Decision Tree (DT) model
shows the lowest performance among the compared
models, indicating its limitations in capturing the
complex patterns inherent in electricity theft data.
4.3 Web Application
An intuitive interface for non-technical users in
utility companies was developed to allow them to
easily interpret and act on the model's outputs. Figure
8, shows the home page which contains the fields
where the user is expected to input. After the user
puts in the relevant inputs and clicks the predict
button the result is shown which is either figure 9, for
theft detected, and figure 10, for a normal user.
Fig. 8: Home page
Fig. 9: Theft Detected
Model
ROC AUC
score
CNN-SVM
0.976
CNN
0.94
SVM
0.89
XGB
0.964
RF
0.962
DT
0.951
LR
0.963
Engineering World
DOI:10.37394/232025.2024.6.27
Nenchin Emmanuel, Ademoh A. Isah
E-ISSN: 2692-5079
253
Volume 6, 2024
Fig. 10: Normal user
4.4 Discussion
Accuracy, recall, precision, f1-score, ROC-AUC
score, and confusion matrix, which summarise the
prediction outcomes on the test data, were used to
assess the performance of the proposed model. On
the test data, the suggested model outperformed other
machine learning models like CNN-SVM, CNN,
SVM, XGB, RF, DT, and LR, achieving an accuracy
of 0.966. In comparison to these models, the
proposed model had higher recall, precision, f1-
score, and ROC-AUC score.
The results of this study indicate that the proposed
CNN-SVM hybrid model is an efficient approach for
electricity theft detection, offering improvements in
both accuracy and reliability.
The web application’s interface makes it easy for
non-technical users in utility companies to easily
interpret and act on the model's outputs.
5 Conclusion and Future Work
The proposed CNN-SVM hybrid model addresses the
significant energy and financial losses caused by
electricity theft and consumer misuse. This approach
not only promises to reduce non-technical losses for
utility companies but also encourages more efficient
electricity usage among consumers.
Our integrated CNN-SVM model demonstrated
superior performance compared to traditional
machine learning approaches in detecting electricity
theft. By synergizing the feature extraction
capabilities of CNNs with the robust classification
strength of SVMs, we achieved higher accuracy,
precision, and recall. The CNN component proved
particularly effective in automatically extracting
relevant features from raw consumption data, while
the SVM classifier excelled in discriminating
between legitimate consumption patterns and
fraudulent activities. This resulted in a lower false
positive rate, crucial for practical implementation in
real-world scenarios.
The model can be integrated into smart grid systems
for real-time monitoring and detection of anomalies
in electricity consumption patterns. Utility
companies can use this system to identify potential
theft cases, thereby protecting their revenue streams.
The model's insights can help in understanding and
predicting consumer behaviour, leading to improved
energy management strategies. The system can assist
in ensuring compliance with energy regulations by
detecting unusual consumption patterns that may
indicate non-compliance.
This research demonstrates the effectiveness of
combining deep learning (CNN) with traditional
machine learning (SVM) for complex pattern
recognition tasks. The success of our model in
extracting features from electricity consumption data
advances the field of automated feature learning in
time series analysis. Our approach shows how AI can
be scaled to handle large-scale, real-world problems
in the energy sector.
The CNN-SVM model can be further enhanced by
integrating it with other AI technologies.
Incorporating Reinforcement Learning algorithms
could allow the model to adapt and improve its
detection capabilities over time based on feedback
from real-world implementations. Implementing
explainable AI techniques could make the model's
decisions more interpretable, increasing trust and
adoption among stakeholders. Also, federated
Learning could enable multiple utility companies to
collaboratively train the model without sharing
sensitive data, enhancing its generalisation
capabilities.
Future research directions for this project are varied
and promising. To further enhance the model's
capabilities, incorporating data from smart metres,
weather patterns, and socio-economic indicators
could provide valuable contextual insights, leading to
improved detection accuracy. Another critical area of
focus is ensuring the model's resilience against
adversarial attacks, which is crucial for high-stakes
applications where reliability is paramount.
Additionally, exploring the model's adaptability to
different geographical regions and energy
consumption patterns through transfer learning
techniques could significantly broaden its
applicability. Investigating the feasibility of
deploying lightweight versions of the model on edge
devices would enable real-time, on-site detection,
making the system even more efficient. Lastly,
addressing potential biases in the model and ensuring
fair treatment across different consumer
demographics is essential for maintaining a just and
equitable solution. By pursuing these research
Engineering World
DOI:10.37394/232025.2024.6.27
Nenchin Emmanuel, Ademoh A. Isah
E-ISSN: 2692-5079
254
Volume 6, 2024
directions, the model can become even more
effective, robust, and widely applicable, ultimately
driving progress in electricity theft detection and
contributing to a more sustainable energy future.
In conclusion, while our CNN-SVM hybrid model
shows significant promise in addressing the critical
challenge of electricity theft, continued research and
development are necessary. Future work should
focus on enhancing the model's adaptability,
interpretability, and ethical implementation. As we
advance, the integration of this technology with
broader smart grid initiatives and energy
management systems could lead to more efficient,
secure, and sustainable energy ecosystems. The
potential impact extends beyond theft detection,
potentially revolutionising how we understand and
manage energy consumption in an increasingly
connected world.
References:
[1] Shehzad, F., Javaid, N., Almogren, A., Ahmed,
A., Gulfam, S. M., & Radwan, A. A robust
hybrid deep learning model for detection of non-
technical losses to secure smart grids. IEEE
Access, 9, 128663-128678, 2021.
[2] Henriques, H. O., Corrêa, R. L. S., Fortes, M. Z.,
Borba, B. S. M. C., & Ferreira, V. H. (2020).
Monitoring technical losses to improve non-
technical losses estimation and detection in LV
distribution systems. Measurement, 161,
107840.
[3] Marques, L., Silva, N., Miranda, I., Rodriges, E.,
& Leite, H. (2016). Detection and localisation of
nontechnical losses in low voltage distribution
networks.
[4] Onat, N. E. V. Z. A. T. (2010). Techno-
economic analysis of illegal electricity usage in
Turkey and policy proposals. WSEAS
Transactions on Power Systems, 3(5), 213-222.
[5] Messinis, G. M., & Hatziargyriou, N. D. Review
of non-technical loss detection
methods. Electric Power Systems
Research, 158, 250-266, 2018.
[6] Tsado, J., & Abel, S. (2022). Electricity Theft
Mitigation at Low Voltage Distribution End
Using Indirect Matrix Converter.
[7] OLADOSU, D. (2024). Energy Theft Detector
(ETD): A Salvage Module from Meter
Bypassing and Illegal Tapping of Electricity.
[8] Nirmal, S., Patil, P., & Kumar, J. R. R. (2024).
CNN-AdaBoost based hybrid model for
electricity theft detection in smart grid. e-Prime-
Advances in Electrical Engineering, Electronics
and Energy, 7, 100452.
[9] Saeed, M. S., Mustafa, M. W., Hamadneh, N.
N., Alshammari, N. A., Sheikh, U. U., Jumani,
T. A., ... & Khan, I. (2020). Detection of non-
technical losses in power utilities—A
comprehensive systematic
review. Energies, 13(18), 4727.
[10] Siniosoglou, I., Radoglou-Grammatikis, P.,
Efstathopoulos, G., Fouliras, P., &
Sarigiannidis, P. (2021). A unified deep learning
anomaly detection and classification approach
for smart grid environments. IEEE Transactions
on Network and Service Management, 18(2),
1137-1151.
[11] Liu, X., Ding, Y., Tang, H., & Xiao, F. (2021).
A data mining-based framework for the
identification of daily electricity usage patterns
and anomaly detection in building electricity
consumption data. Energy and Buildings, 231,
110601.
[12] Promper, C., Engel, D., & Green, R. C. (2017,
November). Anomaly detection in smart grids
with imbalanced data methods. In 2017 IEEE
symposium series on computational intelligence
(SSCI) (pp. 1-8). IEEE.
[13] Kim, S., Sun, Y., Lee, S., Seon, J., Hwang, B.,
Kim, J., ... & Kim, J. (2024). Data-Driven
Approaches for Energy Theft Detection: A
Comprehensive Review. Energies, 17(12),
3057.
[14] Shorten, C., & Khoshgoftaar, T. M. A survey on
image data augmentation for deep
learning. Journal of big data, 6(1), 2019,1-48.
[15] Khan, Z. A., Adil, M., Javaid, N., Saqib, M. N.,
Shafiq, M., & Choi, J. G. (2020). Electricity
theft detection using supervised learning
techniques on smart meter
data. Sustainability, 12(19), 8023.
[16] Hernandes Jr, L. J., Duarte, L. C., Morais, F. O.,
Ferreira, E. C., & Siqueira, J. A. (2001).
Optimizing the inspection routine for the
detection of electrical energy theft in aes
eletropaulo in são paulo, brazil. WSEAS
Transactions on Power Systems, 7(2), 81-89.
[17] Nagi, J., Mohammad, A. M., Yap, K. S., Tiong,
S. K., & Ahmed, S. K. (2008, December). Non-
technical loss analysis for detection of electricity
theft using support vector machines. In 2008
IEEE 2nd International Power and Energy
Conference (pp. 907-912). IEEE.
[18] Stajić, Z., Janjić, A., & Simendić, Z. (2011,
July). Power quality and electrical energy losses
as a key drivers for smart grid platform
development. In Proceedings of the 15th
WSEAS International Conference on Systems,
Engineering World
DOI:10.37394/232025.2024.6.27
Nenchin Emmanuel, Ademoh A. Isah
E-ISSN: 2692-5079
255
Volume 6, 2024
“Recent Researches in System science”, Corfu
Island, Greece, July (pp. 14-16).
[19] Zheng, Z., Yang, Y., Niu, X., Dai, H. N., &
Zhou, Y. (2017). Wide and deep convolutional
neural networks for electricity-theft detection to
secure smart grids. IEEE Transactions on
Industrial Informatics, 14(4), 1606-1615.
[20] Hasan, M. N., Toma, R. N., Nahid, A. A., Islam,
M. M., & Kim, J. Electricity theft detection in
smart grid systems: A CNN-LSTM based
approach. Energies, 12(17), 3318, 2019.
[21] Fang, J., Liu, F., Su, L., & Fang, X. (2024).
Research on Abnormity Detection based on Big
Data Analysis of Smart Meter. WSEAS
Transactions on Information Science and
Applications, 21, 348-360.
[22] Yao, Y., Hui, H., Liang, Z., Feng, X., & Guo,
W. AdaBoost-CNN: A hybrid method for
electricity theft detection. 2021 6th Asia
Conference on Power and Electrical
Engineering (ACPEE), 436-440, 2021.
[23] Singh, S., & Venkaiah, C. Multi-layer model
classifier for cyberattack detection in smart
electric grid. 2023 5th International Conference
on Energy, Power and Environment: Towards
Flexible Green Energy Technologies (ICEPE),
1-6, 2023.
[24] Mazid, A. A., Manaullah, M., & Kirmani, S. A
hybrid approach based on principal component
analysis and convolution neural network for
power theft detection. 2023 International
Conference on Recent Advances in Electrical,
Electronics & Digital Healthcare Technologies
(REEDCON), 313-317, 2023.
[25] Zhou, Y., Zhang, X., Tang, Y., Mu, Z., Shao, X.,
Li, Y., & Cai, Q. (2021, July). Convolutional
neural network and data augmentation method
for electricity theft detection. In 2021 IEEE/IAS
Industrial and Commercial Power System Asia
(I&CPS Asia) (pp. 1525-1530). IEEE.
[26] Ibrahim, N. M., Al-Janabi, S. T., & Al-Khateeb,
B. Electricity-theft detection in smart grid based
on deep learning. SSRN Electronic Journal,
202,. https://doi.org/10.2139/ssrn.3915286
[27] Dimf, G. P., Kumar, P., & Joshua, K. P. CNN
with BI-LSTM electricity theft detection based
on Modified Cheetah Optimization Algorithm in
deep learning, 2023.
[28] Abel, S., Tsado, J., & Tola, O. J. Mitigation of
electricity theft at low distribution voltage end
using matrix converter. In 2022 5th Information
Technology for Education and Development
(ITED) (pp. 1-5). IEEE, November, 2022.
[29] Ullah, A., Javaid, N., Yahaya, A. S., Sultana, T.,
Al-Zahrani, F. A., & Zaman, F. A hybrid deep
neural network for electricity theft detection
using intelligent antenna-based smart meters.
Wireless Communications and Mobile
Computing, 2021, 1-19.
https://doi.org/10.1155/2021/6612165
[30] Mhaske, D., Satam, R., Londhe, S., Kohad, T.,
& Kadam, S. An efficient electricity theft
detection using xg boost. International Journal
of Engineering Applied Sciences and
Technology, 282-287, 2022.
[31] Munawar, S., Asif, M., Kabir, B., Pamir, Ullah,
A., & Javaid, N. Electricity theft detection in
smart meters using a hybrid Bi-directional GRU
Bi-directional LSTM model. In Complex,
Intelligent and Software Intensive Systems:
Proceedings of the 15th International
Conference on Complex, Intelligent and
Software Intensive Systems (CISIS-2021) (pp.
297-308). Springer International Publishing,
2021.
[32] Petrlik, I., Lezama, P., Rodriguez, C., Inquilla,
R., Reyna-González, J. E., & Esparza, R.
Electricity Theft Detection using Machine
Learning. International Journal of Advanced
Computer Science and Applications, 13(12),
2022.
[33] Ahmad, I. S., Zhang, S., Saminu, S., Wang, L.,
Isselmou, A. E. K., Cai, Z., ... & Kulsum, U.
(2021). Deep learning based on CNN for
emotion recognition using EEG signal. WSEAS
Transactions on Signal Processing, 17, 28-40.
[34] Agrawal, V., Goswami, P. K., & Sarma, K. K.
(2021). Week-ahead Forecasting of Household
Energy Consumption Using CNN and
Multivariate Data. WSEAS Trans. Comput, 20,
182-188.
[35] Aurélien Géron. Hands-On Machine Learning
with Scikit-Learn, Keras, and TensorFlow
(Third Edition). O’Reilly Media, Inc, 2022.
[36] Serrat, A., & Benyettou, M. (2019). ATNN and
SVM for Autonomous Mobile Robot.
Énternational Journal of Electrical Engineering
and Computer Science (EEACS), 1, 84-89.
[37] SHIRLY, A. D., SUGANYADEVI, M.,
RAMYA, R., ADAIKALAM, I. A. D., &
MUTHUKUMAR, P. Computation of an
Effective Hybrid DFA-SVM Approach Aimed
at Adaptive PV Power Management.
[38] Bhardwaj, N., Sood, M., & Gill, S. (2024).
Design of Transfer Learning based Deep CNN
Paradigm for Brain Tumor Classification.
WSEAS Transactions on Biology and
Biomedicine, 21, 162-169.
Engineering World
DOI:10.37394/232025.2024.6.27
Nenchin Emmanuel, Ademoh A. Isah
E-ISSN: 2692-5079
256
Volume 6, 2024
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors have no conflicts of interest to declare.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
Engineering World
DOI:10.37394/232025.2024.6.27
Nenchin Emmanuel, Ademoh A. Isah
E-ISSN: 2692-5079
257
Volume 6, 2024
