Detection of Non-Technical Losses using MLP-GRU based Neural Network to Secure Smart Grids

Abstract

In this paper, a data driven based solution is proposed to detect Non-Technical Losses (NTLs) in the smart grids. In the real world, the number of theft samples are less as compared to the benign samples, which leads to data imbalance issue. To resolve the issue, diverse theft attacks are applied on the benign samples to generate synthetic theft samples for data balancing and to mimic real-world theft patterns. Furthermore, several non-malicious factors influence the users' energy usage patterns such as consumers' behavior during weekends, seasonal change and family structure, etc. The factors adversely affect the model's performance resulting in data misclassification. So, non-malicious factors along with smart meters' data need to be considered to enhance the theft detection accuracy. Keeping this in view, a hybrid Multi-Layer Perceptron and Gated Recurrent Unit (MLP-GRU) based Deep Neural Network (DNN) is proposed to detect electricity theft. The MLP model takes auxiliary data such as geographical information as input while the dataset of smart meters is provided as an input to the GRU model. Due to the improved generalization capability of MLP with reduced overfitting and effective gated configuration of multi-layered GRU, the proposed model proves to be an ideal solution in terms of prediction accuracy and computational time. Furthermore, the proposed model is compared with the existing MLP-LSTM model and the simulations are performed. The results show that MLP-GRU achieves 0.87 and 0.89 score for Area under the Receiver Operating Characteristic Curve and Area under the Precision-Recall Curve (PR-AUC), respectively as compared to 0.72 and 0.47 for MLP-LSTM.

Full text

Detection of Non-Technical Losses using
MLP-GRU based Neural Network to Secure Smart
Grids
Benish Kabir1, Pamir1, Ashraf Ullah1, Shoaib Munawar2, Muhammad Asif1, Nadeem Javaid1,∗
1Department of Computer Science, COMSATS University Islamabad, Islamabad 44000, Pakistan
2Department of Electrical Engineering, International Islamic University, Islamabad 44000, Pakistan
Email: beniii.khan20016@gmail.com, pamirshams2011@yahoo.com, ashrafullahmarwat12@gmail.com,
shoaibmunawar26@yahoo.com, muhammad.asif.comsat@gmail.com,
∗Corresponding author: nadeemjavaidqau@gmail.com; www.njavaid.com
Abstract—In this paper, a data driven based solution is
proposed to detect Non-Technical Losses (NTLs) in the smart
grids. In the real world, the number of theft samples are less as
compared to the benign samples, which leads to data imbalance
issue. To resolve the issue, diverse theft attacks are applied on
the benign samples to generate synthetic theft samples for data
balancing and to mimic real-world theft patterns. Furthermore,
several non-malicious factors influence the users’ energy usage
patterns such as consumers’ behavior during weekends, seasonal
change and family structure, etc. The factors adversely affect
the model’s performance resulting in data misclassification. So,
non-malicious factors along with smart meters’ data need to be
considered to enhance the theft detection accuracy. Keeping this
in view, a hybrid Multi-Layer Perceptron and Gated Recurrent
Unit (MLP-GRU) based Deep Neural Network (DNN) is proposed
to detect electricity theft. The MLP model takes auxiliary data
such as geographical information as input while the dataset of
smart meters is provided as an input to the GRU model. Due
to the improved generalization capability of MLP with reduced
overfitting and effective gated configuration of multi-layered
GRU, the proposed model proves to be an ideal solution in terms
of prediction accuracy and computational time. Furthermore, the
proposed model is compared with the existing MLP-LSTM model
and the simulations are performed. The results show that MLP-
GRU achieves 0.87 and 0.89 score for Area under the Receiver
Operating Characterstic Curve and Area under the Precision-
Recall Curve (PR-AUC), respectively as compared to 0.72 and
0.47 for MLP-LSTM.
Index Terms—smart grid, electricity theft, deep learning,
neural network
I. INTRODUCTION
The emergence of Advanced Metering Infrastructure (AMI)
is one of the core innovations of the smart grids. It helps
the power utilities to alleviate the possibility of energy theft
through its tracking capability and fine-grained calculations
[1]. However, using the smart metering system increases the
risk of electricity theft, which leads to the loss of electricity
and is one of the most apparent problems that negatively
affects the performance of the power grids. Electricity losses
are broadly classified into two categories: Non-Technical
Losses (NTLs) and Technical Losses (TLs) [2]. Transformer
and transmission line faults caused by internal power system
components are the most common causes of TLs. The NTLs,
on the other hand, can be calculated as the difference between
complete loss and TLs. Due to these NTLs, Pakistan is losing
0.89 billion rupees per year and an annual loss of 4.8 billion
rupees is faced in India [3].
One of the major NTLs is the stealing of electricity, which
normally involves manipulating the meter reading, bypassing
the electrical meter, etc. Electricity theft has an adverse effect
on the safety and quality of the power source. Different
users exhibit different Electricity Consumption (EC) behavior.
Though, it is a difficult task to recognize NTL patterns among
all regular patterns of EC. Various methods are used to identify
and address the NTLs. These methods are divided into three
fundamental categories: data-driven methods, network based
methods and hybrid based methods. In recent years, data-
driven approaches have gained the attention of academia and
researchers for performing ETD.
Data-driven approaches consist of different deep learning and
machine learning based solutions [3]. These solutions are
used to analyze and detect irregular patterns of consumers’
electricity consumption. Deep learning based methods for the
detection of electrical theft are used in [4], [9]. The authors
present a study of various deep learning models, including
Long-Short-Term Memory (LSTM), Multi-Layer Perceptron
(MLP), Convolutional Neural Networks (CNN), Gated Re-
current Unit (GRU), etc. However, these models have poor
generalization due to inappropriate tuning of hyperparameters.
Despite recent advancements in deep learning and its growing
success, relatively little work has been done in the literature on
the class imbalance issue. The authors in [4] used LSTM-MLP
model for data classification. However, imbalanced dataset
issue is not addressed, which leads to poor Area under the
Precision-Recall Curve (PR-AUC) score of 54.4%. Further-
more, ensemble models fail to detect diverse theft attacks due
to the imbalanced nature of the dataset [5], [6], which leads
to a high False Positive Rate (FPR) while detecting different
cases of theft attacks. Through the analysis of schemes used in

the literature, a hybrid MLP-GRU based Deep Neural Network
(DNN) is proposed in this paper to detect electricity theft using
real smart meters’ data along with auxiliary information.
The rest of the article is organized as follows: Section 2
provides the literature on the detection of energy theft in
smart grids. The proposed technique is discussed in Section 3
while Section 4 presents the performance evaluation metrics.
Simulation results are discussed in section 5 and the paper is
concluded in Section 6.
A. List of Contributions
The contributions of this study are as follows:
•Due to the availability of limited electricity theft samples,
the data augmentation approach is exploited to generate
the fake samples.
•Six theft attacks are used to generate synthesized theft
patterns. To balance the number of synthetic samples gen-
erated and to remove data imbalance problem, Adaptive
Synthetic (ADASYN) sampling approach is used.
•A hybrid model, known as MLP-GRU, is used that
integrates both auxiliary and smart meters’ data for NTL
detection. The proposed model classifies and detects elec-
tricity theft efficiently as compared to existing models.
II. LITERATURE REVIEW
The current hybrid-oriented NTL detection solutions are
based upon machine learning and deep learning approaches.
A complex anomaly detection task is the detection of NTL.
However, it is not accurate to rely on outlier detection meth-
ods (e.g., k-means clustering and local outlier factor) alone
[1]. Traditional approaches cannot work with sequential (EC
history) and non-sequential (auxiliary information) data [1].
Similarly, recent approaches such as CNN and MLP cannot
work with sequential data. In [4], the authors use MLP
combined with LSTM models for NTL detection. However,
the imbalanced dataset issue is not resolved.
In existing supervised learning algorithms, the usage of SVM
and Logistic Regression (LR) has become an active area of
research in ETD. However, they require manual feature ex-
traction that relies on expert knowledge and does not perform
data preprocessing [2], [3], [5], [26], [30]. In [3], the authors
propose a wide and deep model to analyze electricity theft
data. However, the model cannot predict a descent in the
EC that happened before the period of analysis. Whereas, in
[5], the authors propose a Machine Learning (ML) algorithm
based on a boosting technique, termed as Gradient Boosting
Theft Detector (GBTD), which performs efficient feature
engineering based preprocessing to enhance theft detection
performance.
Traditionally, ML techniques are widely used to analyze the
irregular electricity consumption behavior of users to detect
electricity theft [6]. Though, most of these approaches have
poor accuracy due to a lack of generalization. Furthermore,
the high dimensional data degrades the performance of a
neural network with shallow architecture [7]. So, the authors
in [6] introduce a boosting technique, known as XGBoost, as a
supervised learning method for the classification of malicious
users. However, the proposed method has high FPR due to the
extremely imbalanced dataset and on-site inspections, which
are tedious and time-consuming.
Traditional classification techniques have some issues such as
imbalanced dataset problem, high FPR due to non-malicious
factors, and how to detect Zero-day attacks that cannot be
obtained from historic data [8]. A Consumption Pattern-based
Electricity Theft Detection (CPBETD) algorithm is proposed
for the detection of diverse electricity theft attacks against
AMI. However, due to the SVM’s misclassification rate, high
FPR is reported [8].
In the literature, hybrid deep learning techniques are mostly
used for ETD, in which CNN, LSTM, and Random Forest
(RF) models are of vital importance [9], [10], [26], [29].
Moreover, raw datasets are used as inputs for training and
testing, which degrades the models’ classification perfor-
mance. Moreover, we have less number of theft samples
as compared to benign samples in the real world, which
causes data imbalance problem. Balancing the dataset using
Synthetic Minority Oversampling Technique (SMOTE) can
lead to overfitting due to the generation of synthetic samples.
Furthermore, during back-propagation in the CNN network,
its generalization performance degrades due to overtraining in
the softmax classifier layer.
The existing supervised learning based data analytics meth-
ods require labeled data for training as well as additional
information is needed for the detection of energy thefts [12],
[19]. Although, conventional classification based techniques
are used for NTL detection in Power Distribution Companies
(PDCs), they have poor detection rate as well as high FPR
resulting in higher inspection cost, which is a time consuming
and tedious task as well. In article [13], the authors propose
an ensemble bagged tree classification algorithm that uses the
dataset of Multan Electric Power Company (MEPCO) to detect
electricity theft. However, the suggested method requires a
significant amount of time for training the model.
Conventional data driven techniques face issues when classify-
ing daily and weekly energy consumption data. Thus, if these
techniques are extended to hourly or more granular electricity
usage data, their accuracy will be minimum. Since, they fail
to express the trend of intraday electricity usage. In [15],
Text Convolutional Neural Network (Text-CNN) is proposed
to classify the two-dimensional time-series data. However, the
proposed models’ accuracy degrades. Numerous data driven
approaches focus on boosting techniques and ignore bagging
methods, i.e., RF and Extra Trees (ET) as ensemble learning.
In [17], a thorough analysis is performed on an ensemble ML
classifier based on bagging and boosting.
Unsupervised learning approaches have gained a lot of cover-
age of detecting electricity theft. However, on large data, these
methods have a lack of generalization. In [18], the authors
use a Stacked Sparse Denoising Auto-Encoder (SSDAE) for
extracting abstract features of large data in an unsupervised
manner. However, auto-encoders consume more processing
time due to excessive hyperparameters’ tuning. Furthermore,

in [20], the authors propose an unsupervised learning based
anomaly pattern detection approach that requires only normal
users’ consumption data for the model’s training to detect
electricity theft. However, patterns that are detected as outliers
by a classifier may be patterns of more energy usage during
holidays and weekdays.
Likewise, in [22], the authors perform a detailed analysis of
three ML classifiers such as SVM, RF, and k-Nearest Neigh-
bors (KNN), using a dataset of the electric supply company
of Pakistan to predict the existence of NTL. However, there is
a lack of reliable performance metrics for model evaluation.
III. PROP OS ED SY ST EM MO DE L
The proposed solution for NTL detection is shown in Fig.
1. The proposed hybrid deep learning based electricity theft
detector has two phases: training and testing. These two
phases typically consist of five main steps: (1) The data
preprocessing is performed at the first stage of the training
phase. In the data preprocessing, a simple imputer method is
used to replace missing values from the dataset. Afterwards,
a min-max operation is performed for data normalization
using the standard-scalar method. After completing the data
preprocessing, normal users’ samples are obtained. (2) The
normalized and cleaned data is then passed to the next step
in which data augmentation is performed. Fraudulent users’
samples are generated by altering honest samples according
to existing theft attacks [5]. (3) The ADASYN method is
applied on benign samples. (4) Balanced data that is received
from the previous stage is then passed to the next phase
for classification purpose. Balanced smart meters’ data and
auxiliary data are passed to create the MLP module and GRU
module as input for prediction. (5) At the last stage, results are
evaluated by using effective performance measures. Numerous
performance measures are used for comparative analysis, such
as Area under the Receiver Operating Characterstic Curve
(ROC-AUC), F1-score, PR-AUC and accuracy for validating
the proposed model’s efficiency. In the second phase, testing
is performed on new samples to evaluate the trained model’s
performance to determine whether the new sample belongs to
an honest class or a malicious class.
A. Data Preprocessing
The actual electricity consumption data recorded by smart
meters often contain missing values that may arise due to vari-
ous reasons including the short circuit transmission equipment,
bad connection errors, etc., that degrade the performance of
many ML models. The missing data in the dataset provokes
the classifier to classify fraudulent customers incorrectly. Fur-
thermore, when data is dispersed on a large scale, interpreta-
tion becomes complicated and execution time increases. In
the proposed method, we exploit the interpolation method
to recoup the missing data [4]. Simple Imputer is used as
an interpolation method to impute the missing values using
mean, median and the most expected values. Afterwards, the
standard-scalar method is applied for data normalization that is
used for scaling inconsistent data to a common scale between
0 and 1 for better prediction.
B. Data Augmentation using Six Theft Attacks
The number of malicious samples is substantially smaller
than the normal samples in the real world. In this scenario, if
we train ML and deep learning models on the imbalanced data,
the models will be biased regarding the majority class, and in
some cases, it will completely disregard the minority group
that leads to performance degradation. This data mismatch is
a big issue in ETD that needs to be addressed. Numerous
resampling techniques are used in the literature to tackle
this problem [1], [16], [19]. However, the undersampling
techniques lead to important information loss issues. On the
other hand, the oversampling techniques duplicate the minority
class samples that are prone to overfitting.
In consideration of the strong disparity of the massive en-
ergy consumption datasets and the drawbacks of existing
approaches, we generate synthetic samples of theft in te
proposed work by modifying honest samples. Thus, we exploit
the existing six theft attacks to generate different malignant
patterns from normal ones to train ML models with diverse
types of theft patterns [5]. It is an important aspect to generate
different malicious patterns of theft, which introduces variabil-
ity in the dataset. Also, data augmentation helps to analyze the
diversity in the consumption behavior of consumers as well as
reduce overfitting by synthetic data generation.
After generating malicious samples, the minority class (nor-
mal) is oversampled to balance the malicious and non-
malicious data points using ADASYN, which is a variant of
SMOTE.
1) Six Theft Attacks: To generate theft samples, we use
the existing theft cases of diverse types of attacks to alter
the smart meters’ data [5]. If the consumers’ actual use is
denoted by xtwhere xt= [x1,x2, ..., x365], then these theft
cases are used for modifying the real usage of patterns :
Theft attack (a1): xt=xt*r,r= random (0.1, 0.9);
Theft attack (a2): xt=xt*rt,rt= random(0.1, 1.0);
Theft attack (a3): xt=xt*rt,rt= random[0, 1];
Theft attack (a4): xt= avg (x) * rt,rt= random (0.1, 1.0);
Theft attack (a5): xt= avg (x);
Theft attack (a6): xt=xT-t.
Theft attack 1 generates the malicious consumption patterns
by multiplying the benign class of electricity consumption with
the randomly generated values between 0.1 and 0.9. Theft
attack 2 introduces a similar theft case scenario in which
meter readings of normal consumers are multiplied by different
random numbers lying between 0.1 and 1.0 that shows a
discontinuity in manipulated values and tracking of theft. In
theft attack 3, the normal samples are multiplied by 1 at a
given time interval t, and at t+1, the samples are multiplied
by 0. This implies that the consumers either send the actual
readings at a given random timestamp or merely send zero
energy usage at a subsequent time. Moreover, in the theft
attack 4 scenario, an average value of the total consumption is

Smart Meter Data
Theft 1
Malicious
Data
Benign Data
Benign Data
ADAS YN
Balanced Dataset
Data Balancing by Data Augmentation
Theft 2 Malicious
Data
Theft 3
Theft 4
Theft 5
Theft 6
Imbalanced Data
Data Interpolation
Data Normalization
Preprocessing Data
Reliable Performance
F1-Score
ROC-AUC
PR-AUC
Accuracy
Measures
Reliable Performance
MLP
Auxiliary Data
GRU
Feature Extraction and prediction with Parameter Optimization
Weighted Aggregation
Malacious samples
Benign samples
L.1 Imbalanced data set
Limitations Addressed
L.2 Missing values and
outliers
L.3 Miss Classification
due to non-Malacious
factors
S.2
S.1
S.3
Fig. 1: MLP-GRU model architecture
TABLE I: Mapping between identified limitations and proposed solutions
Limitations Identified Solutions Proposed Validations Done
L1: Imbalanced dataset problem S1: Apply six theft attacks on benign samples V1: Comparison with oversampling tech-
niques
L2: Misclassification due to non-malicious factors S2: Incorporate auxilliary data to reduce high FPR V2: Performance comparison with existing
models
multiplied by a random state between (0.1, 1.0) to under-report
the consumed energy. Whereas, theft attack 5 takes the mean
value of the total energy consumption by reporting a consistent
consumption throughout a day. Theft attack 6 occurs when the
malicious users reverse or shift the order of readings from on-
peak hour to off-peak hour [6]. Fig. 2 (a) and (b) depict a
daily energy consumption trend as well as six different types
of malicious attacks.
2) Classification and prediction with MLP and GRU: In the
proposed work, the hybrid neural network of GRU and MLP is
introduced. The proposed GRU-MLP network uses electricity
consumption data as input. The proposed methodology is
inspired by work done in [4] for detecting electricity theft.
The work in [4] developed a hybrid neural network classifier,
known as LSTM-MLP. The preprocessed energy consumption
data of the smart meter is fed into the GRU module with 100
neurons. The GRU layer has twice as many neurons as com-
pared to the MLP model. With relatively fewer cells, the GRU
layer generalizes the embedding at a lower computational cost.
Auxiliary data is passed to the MLP module as input with 20
neurons, as the data has low dimensional features. Until the
data is sent to the final dense layer, the data is normalized
using the batch normalization technique. There is only one
neuron with an activation mechanism of sigmoid in the final
layer.
IV. PERFORMANCE EVALUATION METRICS
In this section, a detailed analysis is performed to compare
the performance of the proposed MLP-GRU network with the
baseline MLP-LSTM model. The performance metrics used to
validate the performance of the above schemes are accuracy,
ROC-AUC, F1-score and PR-AUC [27]. These are derived
from the confusion matrix parameters, which are True Positive
(TP), False Positive (FP), True Negative (TN), False Negative
(FN) that reflect the number of consumers that are correctly
classified as fair consumers, incorrectly classified as normal,
correctly classified as fraudulent and incorrectly classified as
fraudulent users, respectively. Accuracy is one of the most
commonly used metrics that indicate the percentage of correct
predictions by the model. Equation 1 demonstrates accuracy
mathematically. While, the F1 score is another metric that

(a) Theft attack 1,2 and 5
(b) Theft attack 3,4 and 6
Fig. 2: Analysis of different theft attacks
indicates the balance between precision and recall, which is
defined in Equation 2.
Accuracy = (T P +T N )/(T N +T P +F N +F P )(1)
F1−score = 2∗(P recision ∗Recall)/(P recision +Recall)
(2)
The primary objective of ETD is to increase fraud Detection
Rate (DR) or True Psotive Rate (TPR) and low FPR [28]. The
ROC-AUC is an appropriate measure for binary classification
to detect NTLs. It is constructed by plotting TPR also known
as Recall, against FPR while changing the decision thresholds.
The score varies from 0 to 1. It is a more reliable measure in
case of a class imbalance problem. Though, TPR and FPR are
useful indicators for measuring a model’s efficiency for NTL
detection, they do not take into account the models’ precision.
Hence, to evaluate the precision of the model, PR-AUC is a
useful metric that is also a suitable measure for imbalanced
datasets. Therefore, we use PR-AUC, which takes into account
the classifier’s accuracy as well as shows the expense of on-
site utility inspections.
V. SIMULATION AND RESU LTS
In this section, the simulation results are discussed. The pro-
posed model is evaluated on Pakistan Residential Electricity
Consumption (PRECON) dataset.
A. Evaluation Results
The classification accuracy of the proposed hybrid model
is shown in Fig. 3 (a). Since the proposed MLP-GRU model
outperforms the GRU model in terms of ROC-AUC, obtaining
an AUC score of 0.87 on the test data. By considering non-
sequential or auxiliary features like the contracted power,
permanent residents and property area, etc., the NTL detection
performance is significantly increased. Whereas, MLP-LSTM
has poor performance with an AUC score of 0.72 on the test
dataset due to the limited generalization ability of the LSTM
model.
Fig. 3 (a) demonstrates a comparative analysis of the
developed hybrid model. TPR and FPR are plotted on X-
axis and Y-axis, respectively. TPR indicates the correctly
classified samples of the total available samples whilst FPR is
an expensive parameter. Initially, the proposed hybrid model
classifies the binary distribution with high accuracy and low
FPR. After reaching a peak TPR of 0.7, a slight change is
observed with an increasing FPR. However, the observed FPR
of our model is much lower than MLP-LSTM. Afterwards,
a periodic regain of our proposed models’ ROC is observed
exponentially. The periodic increase in TPR reduces FPR that
enhances the model’s stability and accuracy. This reduction
in FPR by our designed model reduces on-site inspections,
which is an expensive parameter for the utility providers due
to the deployment of experts for on-site inspections to verify
the cause. Similarly, Fig. 3 (b) shows the PR-AUC curve. The
PR-AUC of the proposed model is significantly higher than
MLP-LSTM with a PR-AUC value of 0.89 on the test dataset.
In Table 2, it is observed that the proposed MLP-GRU model
outperforms the MLP-LSTM model in terms of AUC, accuracy
and F1-score. The reason is that the computational complexity
of the proposed MLP-GRU model is low because few gates
are used in GRU as compared to the LSTM model. Also, it
performs better due to the usage of a small dataset.
TABLE II: Comparison Results
Models AUC score Accuracy F1 score
MLP-GRU 0.87 0.78 0.82
MLP-LSTM 0.72 0.51 0.62
VI. CONCLUSION
In this paper, we propose a hybrid model, known as
MLP-GRU, for detecting NTLs using smart meters’ data and

(a) ROC-AUC Curve
(b) PR-AUC Curve
Fig. 3: Evaluation Metric of MLP-GRU and MLP-LSTM
Models
auxiliary data. We incorporate sequential data as input to
the MLP module and auxiliary data is passed to the GRU
module. Furthermore, since the electricity consumption data
contains a limited number of malicious users, it makes the
classification model biased towards the majority class. To
address this problem, we generate synthetic theft patterns
using six theft attacks on benign samples for data balancing
and to incorporate diversity in theft patterns. Afterwards, we
evaluate the performance of our hybrid model against non-
malicious changes in electricity consumption patterns of users
and diverse theft attacks. Simulations are conducted using
the PRECON dataset along with theft attacks. The results
show that our proposed hybrid model outperforms the baseline
MLP-LSTM model. It is observed that by integrating auxiliary
information along with smart meters’ data, the model’s per-
formance is significantly improved in terms of PR-AUC and
ROC-AUC score with 0.89 and 0.87, respectively. In fact, the
efficiency of baseline model MLP-LSTM and the proposed
hybrid MLP-GRU network is quite low in terms of training
accuracy, which implies the importance of an optimization
algorithm for tuning hyperparameters of models to achieve
optimal results.
REFERENCES
[1] Buzau, M. M., Tejedor-Aguilera, J., Cruz-Romero, P., & G´
omez-Exp´
osito,
A. (2018). Detection of non-technical losses using smart meter data and
supervised learning. IEEE Transactions on Smart Grid, 10(3), 2661-2670.
[2] Kong, X., Zhao, X., Liu, C., Li, Q., Dong, D., & Li, Y. (2021). Electricity
theft detection in low-voltage stations based on similarity measure and
DT-KSVM. International Journal of Electrical Power & Energy Systems,
125, 106544. DOI: 10.1016/j.ijepes.2020.106544.
[3] 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.
[4] Buzau, M. M., Tejedor-Aguilera, J., Cruz-Romero, P., & G´
omez-Exp´
osito,
A. (2019). Hybrid deep neural networks for detection of non-technical
losses in electricity smart meters. IEEE Transactions on Power Systems,
35(2), 1254-1263.
[5] Punmiya, R., & Choe, S. (2019). Energy theft detection using gradient
boosting theft detector with feature engineering-based preprocessing.
IEEE Transactions on Smart Grid, 10(2), 2326-2329.
[6] Yan, Z., Wen, H. (2021). Electricity theft detection base on extreme
gradient boosting in AMI. IEEE Transactions on Instrumentation and
Measurement, 70, 1-9.
[7] Avila, N. F., Figueroa, G.Chu, C. C. (2018). NTL detection in electric
distribution systems using the maximal overlap discrete wavelet-packet
transform and random undersampling boosting. IEEE Transactions on
Power Systems, 33(6), 7171-7180.
[8] Jokar, P., Arianpoo, N., & Leung, V. C. (2015). Electricity theft detection
in AMI using customers’ consumption patterns. IEEE Transactions on
Smart Grid, 7(1), 216-226.
[9] Li, S., Han, Y., Yao, X., Yingchen, S., Wang, J., Zhao, Q. (2019).
Electricity theft detection in power grids with deep learning and random
forests. Journal of Electrical and Computer Engineering, 2019. DOI:
10.1155/2019/4136874.
[10] Hasan, M., Toma, R. N., Nahid, A. A., Islam, M. M., Kim, J. M. (2019).
Electricity theft detection in smart grid systems: A CNN-LSTM based
approach. Energies, 12(17), 3310. DOI: 10.3390/en12173310.
[11] Fenza, G., Gallo, M., Loia, V. (2019). Drift-aware methodology for
anomaly detection in smart grid. IEEE Access, 7, 9645-9657.
[12] Zheng, K., Chen, Q., Wang, Y., Kang, C.,Xia, Q. (2018). A novel
combined data-driven approach for electricity theft detection. IEEE Trans-
actions on Industrial Informatics, 15(3), 1809-1819.
[13] Saeed, M. S., Mustafa, M. W., Sheikh, U. U., Jumani, T. A., Mirjat,
N. H. (2019). Ensemble bagged tree based classification for reducing
non-technical losses in multan electric power company of Pakistan.
Electronics, 8(8), 860. DOI: 10.3390/electronics8080860.
[14] Li, W., Logenthiran, T., Phan, V. T., Woo, W. L. (2019). A novel smart
energy theft system (SETS) for IoT-based smart home. IEEE Internet of
Things Journal, 6(3), 5531-5539.
[15] Feng, X., Hui, H., Liang, Z., Guo, W., Que, H., Feng, H., Ding,
Y. (2020). A Novel Electricity Theft Detection Scheme Based on
Text Convolutional Neural Networks. Energies, 13(21), 5758. DOI:
10.3390/en13215758.
[16] Qu, Z., Li, H., Wang, Y., Zhang, J., Abu-Siada, A.,Yao, Y. (2020).
Detection of electricity theft behavior based on improved synthetic
minority oversampling technique and random forest classifier. Energies,
13(8), 2039. DOI: 10.3390/en13082039.
[17] Gunturi, S. K.Sarkar, D. (2020). Ensemble machine learning models for
the detection of energy theft. Electric Power Systems Research, 106904.
DOI: 10.1016/j.epsr.2020.106904.
[18] Huang, Y., & Xu, Q. (2021). Electricity theft detection based on stacked
sparse denoising autoencoder. International Journal of Electrical Power &
Energy Systems, 125, 106448. DOI: 10.1016/j.ijepes.2020.106448.
[19] Gong, X., Tang, B., Zhu, R., Liao, W., Song, L. (2020). Data Aug-
mentation for Electricity Theft Detection Using Conditional Variational
Auto-Encoder. Energies, 13(17), 4291. DOI: 10.3390/en13174291
[20] Park, C. H., & Kim, T. (2020). Energy Theft Detection in Advanced
Metering Infrastructure Based on Anomaly Pattern Detection. Energies,
13(15), 3832. DOI: 10.3390/en13153832.

[21] Adil, M., Javaid, N., Qasim, U., Ullah, I., Shafiq, M., Choi, J. G. (2020).
LSTM and bat-based RUSBoost approach for electricity theft detection.
Applied Sciences, 10(12), 4378. DOI: 10.3390/app10124378.
[22] Maamar, A., Benahmed, K. (2019). A hybrid model for anomalies
detection in AMI system combining K-means clustering and deep neural
network. Comput. Mater. Continua, 60(1), 15-39.
[23] Ding, N., Ma, H., Gao, H., Ma, Y.,Tan, G. (2019). Real-time anomaly
detection based on long short-Term memory and Gaussian Mixture
Model. Computers Electrical Engineering, 79, 106458.
[24] Jindal, A., Dua, A., Kaur, K., Singh, M., Kumar, N.,Mishra, S. (2016).
Decision tree and SVM-based data analytics for theft detection in smart
grid. IEEE Transactions on Industrial Informatics, 12(3), 1005-1016.
[25] Lu, X., Zhou, Y., Wang, Z., Yi, Y., Feng, L. ,Wang, F. (2019).
Knowledge embedded semi-supervised deep learning for detecting non-
technical losses in the smart grid. Energies, 12(18), 3452. DOI:
10.3390/en12183452.
[26] Arif, A., Javaid, N., Aldegheishem, A., Alrajeh, N. Big Data Analytics
for Identifying Electricity Theft using Machine Learning Approaches
in Micro Grids for Smart Communities. Concurrency and Computation
Practice and Experience, DOI: 10.1002/cpe.6316
[27] Ghori, K. M., Imran, M., Nawaz, A., Abbasi, R. A., Ullah, A., Szath-
mary, L. (2020). Performance analysis of machine learning classifiers
for non-technical loss detection. Journal of Ambient Intelligence and
Humanized Computing, 1-16.
[28] Aslam, Z., Ahmed, F., Almogren, A., Shafiq, M., Zuair, M., & Javaid,
N. (2020). An Attention Guided Semi-Supervised Learning Mechanism
to Detect Electricity Frauds in the Distribution Systems. IEEE Access, 8,
221767-221782.
[29] Aldegheishem, A., Anwar, M., Javaid, N., Alrajeh, N., Shafiq, M., &
Ahmed, H. (2021). Towards sustainable energy efficiency with intelligent
electricity theft detection in smart grids emphasising enhanced neural
networks. IEEE Access, 9, 25036-25061.