Electric energy is a fundamental requirement for daily life activities and processes in the modern world. Using energy resources is essential for the economic development and growth of every country worldwide (Lowitzsch et al., 2020). However, crises can arise when energy consumption exceeds production, leading to a shortfall and interruption in energy supplies (Ren et al., 2021). Many underdeveloped and economically unstable countries, including the India, Indonesia, Malaysia, Pakistan, Nigeria, Ethopia and China, are currently facing energy crises (Østergaard et al., 2021). Electricity is one of the primary forms of energy used globally, and the demand for it is rapidly increasing. However, many developed and developing countries face electricity crises due to Technical Losses (TL) and Non-Technical Losses (NTLs) (Kumar et al., 2019). NTLs are caused by fraud, electricity theft, tampering with the recording process, and non-billing of electricity (Rahman et al., 2020). To mitigate NTLs, modern methods for detecting fraud and electricity theft are required. NTLs are the primary cause of revenue loss in smart grids (Stracqualursi et al., 2023). Recent studies have revealed that NTLs result in global annual losses of USD 89.3 billion in the utility sector of electrical energy (de Souza et al., 2020). The problem of NTLs is still relevant in both developed and developing countries. Normally, energy losses in developed countries range between 0.5% and 3% yearly in revenue collection (Park and Kim, 2020; Kumar et al., 2020). In developing countries, the losses are approximately 4.5 billion USD annually, accounting for about 50% of electricity produced (Quasim et al., 2023). Developed countries like the United Kingdom (UK) and the United States of America (United States of America) face annual losses ranging from 1 to 6 billion USD (Duarte Soares et al., 2022). Pakistan also faces NTLs of 0.89 billion USD per annum due to non-billing and electricity theft (Rehan et al., 2023). NTLs can be intentional or unintentional, and many power supply companies try to detect and reduce them efficiently (Zhang et al., 2020). Due to the rise in electricity fraud, several methods have been adopted to automatically detect electricity theft, like assessing electricity consumption records (Kocaman and Tümen, 2020). Hardware and data-driven solutions are implemented to mitigate NTLs. Hardware solutions commonly use grid system variables, including power, voltage, and current, while data-driven solutions analyze and mine consumers’ load profiling and other information to detect NTLs (Muzumdar et al., 2022). However, it faces many obstacles in ETD and fraudulent consumers due to its technical theft strategies, including line tapping, meter tempering, etc., and needs extra devices for implementation (Chandrasekhar et al., 2020). Hence, it is costly, Furthermore, to tackle the NTLs and ETD in smart grids, Advanced Metering Infrastructure (AMI) is better than old mechanical metering (Saxena et al., 2021).

Due to the increase in the number of electricity thieves, the electric utilities are facing problems in providing electricity to their consumers in an efficient way (Xie, 2023). An accurate ETD is quite challenging due to the inaccurate classification on the imbalance electricity consumption data, the overfitting issues and the high false positive rate of the existing techniques (Blazakis et al., 2020). Therefore, intensified research is needed to accurately detect the electricity thieves and to recover a huge revenue loss for utility companies. To address the above limitations, this paper presents a new model, which is based on the supervised machine learning techniques and real electricity consumption data.

In this study, we propose a new hybrid system based on deep learning models that accurately detect electricity theft in smart grids while also being efficient. The first step involves preprocessing the data and replacing the missing values using a simple imputer method. Next, we use the standard-scalar approach to execute a min-max operation for data normalization. After data preparation is finished, we obtain samples for typical users. We then balance the data using k-means SMOTE to create samples for fraudulent users by altering honest samples with current theft attacks. In the third step, the balanced data from the previous stage is used for classification purposes. The MLP and GRU modules were created in Python^® using balanced smart meter data and supplementary data as input for prediction. We apply efficient performance criteria in the final phase to investigate the results. To validate the proposed model’s performance using various performance measures, including accuracy, F1-score, precision, and recall. In addition, we also test the consistency and efficiency of the trained model on new samples in the second phase to identify whether the new sample belongs to the honest class or the malicious. The main applications of this paper are: the proposed approach provides the solution for the problem present in the power sector, such as to wastage of electrical power due to electricity theft. This model can efficiently be applied by the utility companies using the real electricity consumption data to identify the electricity thieves and reduce the energy wastage and finally, the proposed approach can be used against the all types of consumers who steal the electricity.

The key contributions of this paper: A comprehensive data pre-processing is performed using interpolation, three sigma rule, and normalization methods to deal with missing values and outliers in the dataset. The data pre-processing step gives the refined input, which improves the performance of the classifier. A class balancing technique, K-means SMOOTHE, is proposed to address the problem of imbalance data. MLP is applied to predict final misclassification, which improves the performance along with MLP, GRU technique is utilized for efficient parameter optimization of the classifier. The complete procedure of the proposed hybrid system (MLP-GRU) for ETD is: first, preprocess the data; second, balance the data; third, apply the GRU model to the extracted purified data; fourth, apply the MLP model to the extracted purified data; and finally, evaluate the performance of the proposed system using different performance measures like a graphical analysis and a statistical test. We conduct extensive simulations on real electricity consumption data set and for comparative analysis, precision, recall, F1-score, Matthews Correlation Coefficient (MCC), Receiving Operating Characteristics Area Under Curve (ROC-AUC), and Precision Recall Area Under Curve (PR-AUC) are used as performance metrics.

The rest of the article is organized as follows: Section 2 contains the existing literature and Section 3 consists of a proposed hybrid system. Section 4 discusses the results and compared the results with the best studies available in literature. Finally, Section 5 presents a conclusion and directions for future research work.

The problem of ETD and NTLs is rising quickly at global scale, and researchers are developing techniques to tackle this problem by applying statistical, machine learning, and deep learning models (Jaiswal et al., 2020). The machine learning models, including Random Forest (RF), Decision trees (DT), Bagging Ensemble (BE), Artificial Neural Networks (ANN), and K-Nearest Neighbors (KNN), were comparatively evaluated for automated ETD in smart grid environments and found RF yields 10% more improved accuracy in ETD compared to other used methods (Zidi et al., 2023). The rapid growth in NTLs and electricity thefts are the major challenges for distribution network operators. In (Fei et al., 2022) neural network model based on Neural Architecture Search (NAS) is developed to analyze and detect electricity theft in missing value scenarios through density-based spatial clustering of application and noise clustering technique and achieved an excellent result of AUC of 0.926 in the NTLs and ETD. The authors in (Banga et al., 2022) used different deep learning models for NTLs and ETD, including GRU, Long Short-Term Memory (LSTM) models, MLP, and Convolutional Neural Networks (CNN). However, these models lack efficient hyper parameter tuning, which leads to poor generalization for tackling these issues.

The researchers in (Asif et al., 2022) proposed hybrid deep learn ing models based on Bidirectional Long Short-Term Memory (Bi-LSTM) networks and Two-Dimensional Convolutional Neural Networks (2D-CNN) to detect NTLs in smart meters, and they outperformed other methods with ROC 0.97 and AUC 0.98 in smart meters data. In detecting NTLs and ETD, the problem of class imbalance, the curse of dimensionality reduction, and inappropriate tuning of hyper parameters arise in commonly used machine learning and deep learning models. Therefore, to cope with these problems, the authors in (Ullah et al., 2022) proposed a hybrid deep learning method based on Alexnet and Adaboost for ETD in smart grids. They achieved the best performance results compared to other methods used, and the problem of class imbalance and the curse of dimensionality is being tackled by under-sampling techniques and tuning hyper parameters by the Artificial bee Colony (ABC) optimization algorithm. Furthermore, the authors in (Kumar et al., 2022) proposed a hybrid method based on CNN and RF to predict ETD in power grids accurately. The RF is employed for classification, while CNN efficiently extracts the potential features. A deep learning-based hybrid model is designed by (Hasan et al., 2019), which uses the pros of both CNN and LSTM models and efficiently extracted the hidden patterns and temporal correlation in ETD of consumers in smart grid systems, respectively. The researchers in (Gupta et al., 2022) proposed a Deep Neural Network (DNN) model, first they resolved the dimensionality problem, and then important features were selected for the detection of fraud in the electricity consumption of the smart grid. The ETD results show the proposed method’s best performance over the other used models. The problem of diverse theft patterns in electricity consumption due to a significant class imbalance in data leads to higher false positive rates, and ensemble models fail to detect NTLs. Hence, the authors in (Alameady et al., 2022) tried to overcome this issue by proposing hybrid neural networks named MLP-GRU for detecting electricity thefts in smart meter data by analyzing the auxiliary information of the consumers.

With the advent of smart meters, different types of electricity theft techniques have been adopted, and their detection is very difficult using conventional methods. The researchers in (Li et al., 2019) designed a statistical and machine learning-based Internet of Things (IoT) system to identify and notify electricity consumers about electricity thefts. Many studies based on data-driven techniques have been used for NTLs identification in the literature. Most studies have focused on boosting approaches, and less attention is given to bagging approaches like Extra Trees (ET) and RF (Siu et al., 2022). Furthermore, commonly used machine learning models such as Support Vector Machine (SVM) and neural networks yield a higher false positive rate and a lower detection rate in ETD (Ahir and Chakraborty, 2022). The authors in (Gong et al., 2020) have explored a Conditional Variation Auto-Encoder (CVAE) combined with CNN for extracting relevant features from high-dimensional data and presented a solution to the problem of data augmentation. Furthermore, many studies and experiments have been conducted on ETD in AMI through machine learning techniques (Aziz et al., 2020). The researchers in (Jokar et al., 2015) implemented a pattern recognition technique based on unsupervised learning models for ETD in the data set of smart meters.

The deployment of advanced sensors has strengthened the monitoring capability of power plants. In the context of the cogeneration process, the plant cooling is performed by the cooling towers using the condensation process on exhaust steam. However, the computer networks and industrial control systems built on this sensor-based digital layer may become vulnerable to cyber attacks. This may eventually raise a concern on the performance and security of these energy utilities. To resolve this issue, an inoculated subobserver-based fusion filter is proposed. It improves the resilience against malicious attacks in combined cycle power plants with desalination units, which are usually functioning in a closed-loop environment and infected with injected attacks. A time-delay-based state representation is considered for the system (Khalid et al., 2019).

IoT is a developing technology that provides the simplicity and benefits of exchanging data with other devices using the cloud or wireless networks. However, the changes and developments in the Internet of Things (IoT) environment are making IoT systems susceptible to cyber attacks which could possibly lead to malicious intrusions. The impacts of these intrusions could lead to physical and economical damages. This article primarily focuses on the IoT system/framework, the IoT, learning-based methods, and the difficulties faced by the IoT devices or systems after the occurrence of an attack. Learning-based methods are reviewed using different types of cyber attacks, such as denial-of-service (DoS), distributed denial-of-service (DDoS), probing, user-to-root (U2R), remote-to-local (R2L), botnet attack, spoofing, and man-in-the-middle (MITM) attacks (Inayat et al., 2022).

Synchrophasor-based wide-area monitoring system (WAMS) applications are vital for acquiring the real-time grid information under ambient and nonlinear conditions. The high dependence on sensor data and signal-processing software for daily grid operation is becoming a concern in an era prone to cyberattacks. To resolve this issue, a mixture density-based maximum likelihood (MDML) estimation was proposed to detect attack vectors. The algorithm was deployed at each monitoring node using a track-level fusion (TLF)-based architecture. A parallelized message passing interface (MPI)-based computing was processed to reduce its computational burden. This work adopted a mature application known as oscillation detection as an example of a monitoring candidate to demonstrate the proposed method (Khalid et al., 2023).

This paper introduces the theft detection method which uses comprehensive features in time and frequency domains in a deep neural network-based classification approach. We address dataset weaknesses such as missing data and class imbalance problems through data interpolation and synthetic data generation processes. We analyze and compare the contribution of features from both time and frequency domains, run experiments in combined and reduced feature space using principal component analysis and finally incorporate minimum redundancy maximum relevance scheme for validating the most important features. We improve the electricity theft detection performance by optimizing hyper-parameters using a Bayesian optimizer and we employ an adaptive moment estimation optimizer to carry out experiments using different values of key parameters to determine the optimal settings that achieve the best accuracy (Lepolesa et al., 2022).

This work proposes two novel methods to resolve the above-mentioned issues: Tomek Link Borderline Synthetic Minority Oversampling Technique with Support Vector Machine (TBSSVM) and Temporal Convolutional Network with Enhanced Multi-Layer Perceptron (TCN-EMLP). The former resamples the data by balancing the majority and minority class instances. Whereas, the latter classifies normal and fraudulent consumers. Moreover, deep learning models suffer from high variance in their final results due to the assignment of different weights. Therefore, an averaging ensemble strategy is applied in this work to reduce the high variance (Arif et al., 2022).

In previous literature, most research focused on non-malicious electricity consumption patterns and showed low detection rates for NTLs. In (Ding et al., 2019), the authors proposed a hybrid approach based on the enhanced internal structure of the LSTM model with a combination of the Gaussian Mixture Model (GMM). However, it only applies to low-dimensional data and is not robust to outliers. In further studies (Jindal et al., 2016), a hybrid method based on SVM has been proposed for detecting fraudulent consumers but has not shown an effective performance in the overall technical evaluation. The authors (Kabir et al., 2022; Kumari et al., 2022) proposed a hybrid deep learning model based on Multi-Layer Perceptron (MLP) and Gated Recurrent Unit (GRU) for the detection of electricity thefts and NTLs in smart meter data. The MLP network is used for analyzing non-malicious factors on auxiliary information in the daily consumption of electricity data, while the GRU network is used for analyzing smart meter data. Furthermore, a random search algorithm turns hyper-parameters and performs better than other methods.

In this section, we discuss in detail the complete procedure of the proposed hybrid system for ETD. To do this, first, preprocess the data; second, balance the data; third, apply the GRU model to the extracted purified data; fourth, apply the MLP model to the extracted purified data; and finally, evaluate the performance of the proposed system using different performance measures like a graphical analysis and a statistical test. The framework of proposed method is given in Figure 1.

Real-time consumer data for residential customers makes up the exploited data. The customers are divided into two groups based on similar indexing patterns and appropriately designated consumption categories. Each consumer’s consumption pattern is represented by a staging numeric binary. Label 0 denotes an honest consumer, while label 1 denotes a dishonest one. Each consumer’s tracked and reordered patterns are collected every 24 h. For each of the theft versions, benign class data is modified to create harmful data. Data from both types is later combined. Due to the model’s deviation towards the majority class, a data balancing strategy is necessary to minimize the class bias issue. The data is balanced using K-means SMOTE. The hyper-parameters and their appropriate values obtained during the tuning of the proposed MLP-GRU model are shown in Table 3. Due to their lengthy computation, we investigate fewer hyper-parameters.

The proposed model’s training process is handled by the epoch variable. We accomplish 15 iterations, or epochs, of our model. The results show that the MLP-GRU’s training accuracy (accuracy on seen data) gradually improves with each iteration, reaching a maximum of 86% in the last iteration. Whereas the MLP-GRU accuracy gradually rises as well, reaching 88% at the final iteration when utilizing the testing data (accuracy using unseen data). The CNGC dataset has some zero values, which makes it difficult for the proposed classifier to learn it correctly in the early iterations. As a result, during the first three epochs, training accuracy is higher than testing accuracy, indicating that overfitting has taken place. The suggested model successfully learns the zero values after the third iteration and the overfitting problem is fixed. At various iterations, the MLP-GRU loss is also calculated and recorded. The training loss is decreasing with each iteration, until it achieves a minimum of 0.108 at the 10th iteration. The testing loss also decreases until the final iteration, when it hits 0.080. Due to the zero values included in the dataset, the model overfits during the first three iterations. After the third iteration, the model has learned both the dataset and the zero values, which eliminates the overfitting problem. Finally, the proposed model’s training and testing accuracy leads to the conclusion that it generalizes effectively and avoids overfitting.

Using the same dataset, we trained the Alexnet, BGRU, and RNN models to compare our proposed hybrid model (MLP-GRU) to benchmark methods. Hence, Table 4 displays the performance measures and PR-AUC for the MLP-GRU, Alexnet, BIGRU, and RNN models in three training and testing sets cases. In the first case of 50% training and 50% testing, the proposed MLP-GRU produced recall 90, precision 87, F1-Score 89, accuracy 87.80, test loss 32, AUC 91, ROC 100, and MCC 73, respectively. Meanwhile, Alexnet produces recall 85, precision 96, F1-Score 90, Accuracy 87.63, test loss 34, AUC 94, ROC 85, MCC 73, and GRU poorly performed with recall 75, precision 86, F1-Score 80, Accuracy 74.98, test loss 49, AUC 85, ROC 75, and MCC 46, respectively. Moreover, in the second case of 75% training and 25% testing, the proposed model yields recall 95, precision 95, F1-Score 91, accuracy 92.12, test loss 90, AUC 93, ROC 100, and MCC 76, respectively, and Alexnet yields the second-best results with recall, precision, F1-Score, Accuracy, Test loss, AUC, ROC, and MCC 94, 91.01, 20, 90, 97, and 82 accordingly. However, the results produced by other benchmarks were not up to par in the second case either. Similarly, the proposed models show the best performance in the third case of 90% training and 10% testing set with recall 95, precision 94, F1-Score 94, accuracy 93.33, test loss 20, AUC 95, ROC 100, and MCC 85, followed by Alexnet with recall 94, precision 86, F1-Score 90, accuracy 89.12, test loss 22, AUC 94, ROC 80, and MCC 64. Meanwhile, GRU again performed poorly with the metrics of recall 88, precision 86, F1-Score 87, accuracy 83.86, test loss 37, AUC 89, ROC 80, and MCC 64.

It is evident from the results that the proposed model outperformed other models, and Alexnet was found to be the second-best performer in all three cases because it uses a k-means smote sampling technique to balance the data. However, GRU, performs the least well among the classifiers, with an accuracy of 74.98% at a 50% training ratio. This is because GRU does not capture long-term dependencies from the huge time series data since it is based on the probability notion and employs neural network theory. Furthermore, the training using the majority of class samples makes it biased when identifying genuine incidents of electricity theft. As a result, GRU is unable to classify the vastly unbalanced dataset accurately. By achieving 87% accuracy, the Alexnet, in contrast, performs marginally better than the BGRU. Alexnet is a deep learning model that extracts hidden patterns from data on power use to detect electricity thieves. It has numerous stacks of hidden layers. However, because of the thick layers, it suffers from overfitting problems. It is unable to perform well across the board. Furthermore, we have compared our proposed model concerning execution time, and it is revealed in Table 5 that the execution time of the proposed hybrid MLP-GRU is 106.837 s in the first case of 50% training, 147.163 s for 75% training, and 167.221 s in 90% training set. This indicates that the proposed model is executed in less time than other benchmark models, including GRU, BGRU, and RNN, except for Alexnet. The Alexnet model here performed best in terms of less execution time due to the issues of imbalanced binary classification; this method works better in the least possible time.

Finally, the DMT results (test statistic values) are tabulated in Table 6. This table confirms that the prediction of the proposed hybrid system demonstrated that the GRU-MLP model significantly outperformed Alexnet, GRU, BGRU, and RNN. Moving forward, the corresponding test statistical values were negative in all cases at the 5% significance level using the loss square function. In the end, the performance of the proposed MLP-GRU versus the baseline models is plotted in Figure 6 for all three scenarios of training and testing, such as 50%, 50%; 75%, 25%; and 90%, 10%. The first scenario (50%, 50%) of performance matrices such as recall, precision, F1-Score, accuracy, test loss, AUC, ROC, and MCC is displayed in Figure 6A. The MLP-GRU is represented by a blueviolet bar, Alexnet by a periwinkle blue bar, BGRU by a green bar, a pink bar for RNN, and GRU by a yellow bar for the first case of training and testing, respectively. Therefore, it is evident in the bar plot that the proposed MLP-GRU performed best, followed by Alexnet, and GRU showed poor performance for ET prediction in smart grids. Moreover, Figures 6B, C explore the performance of these considered models for the second (75%, 25%) and third (90%, 10%) cases of the training and testing sets. After thoroughly evaluating the displayed plots, it was concluded that the proposed MLP-GRU model outperformed their competitors in the consistent prediction of electricity thefts in smart grids. However, the Alexnet model showed the second-best results. Additionally, we have plotted level plots of DMT p-values in Figure 7 for the proposed MLP-GRU, Alexnt, BGRU, RNN, and GRU to check their significance in predicting electricity thefts. However, Figure 7A explores the p-values of the test for the first case of 50% training and 50% testing, Figure 7B displays the second of 75% training and 25% testing, and the third one is plotted in Figure 7C for 90% training and 10% testing set. It is confirmed from the plotted level plot that the proposed MLP-GRU significantly outperformed other used deep learning models in electricity theft prediction.

To evaluate the effectiveness of the proposed hybrid system for ETD by comparing its performance with other state-of-the-art methodologies reported in the literature. To achieve this, we have presented a comparison Table 7 that highlights the superiority of the proposed hybrid system compared to the best approaches reported in the literature. For instance, study (Khan et al., 2024), the proposed method used the current study dataset and obtained performance measures [accuracy (91%), precision (97.96%), and area under curve (91.68%)] that were comparatively higher than our proposed hybrid technique [accuracy (93.3%), precision (97.5%), and area under curve (95%)]. In another study, (Munawar et al., 2021), the best-proposed method used the current study dataset and computed the performance metrics [accuracy (89.9%), F-score (90.86%), and area under curve (78%)] that were also comparatively higher than our proposed system. On the other hand (Gul, 2020), reported the best-proposed approach for detecting electricity theft in the smart grid. For comparison, the authors obtained the evaluation measures for this approach, such as accuracy (91.29%), F1 score (90.96%), and area under the curve (0.87%), which were comparatively greater than the current proposed hybrid system. However, in the work (Bohani et al., 2021), the best-proposed method used the present study dataset and computed the performance indicators, that is, accuracy (91%), F1score (88.99%), and area under the curve (86%) that were also relatively higher than our proposed hybrid system. Also, the best-proposed model of (Razavi et al., 2019) was applied to this work’s dataset, and their performance measures were obtained. The best-proposed model of (Meidan et al., 2020) reported the performance measures values as the following: accuracy = 57.70%, F-score = 70.01%, and area under curve = 77.01%, respectively, which are remarkably greater than our performance measures values: accuracy = 93.3%, F-score = 94.96%, and area under the curve = 95%. In a previous study (Pamir et al., 2022), the authors used the same dataset as our proposed hybrid system and achieved higher performance measures. For example, the accuracy, F1-score, and area under curve achieved by their best method were 73.20%, 70.10%, and 69.50%, respectively, which were significantly better than our proposed measures. Another study (Qu et al., 2020) also used our dataset and achieved higher accuracy (87.90%), F-score (96.11%), and area under the curve (87.90%) than our proposed hybrid system. In summary, our proposed hybrid system achieved high accuracy but was outperformed by the best methodologies in the literature.

ETD is a significant issue in many countries, leading to a substantial financial loss for power utilities worldwide. However, conventional methods for ETD face challenges such as the curse of dimensionality and an imbalanced distribution of electricity consumption data. To overcome these problems, this study proposes a hybrid system named MLP-GRU that analyzes and solves electricity theft using data from the CNGC. The proposed hybrid system consists of several steps. Firstly, the data undergoes preprocessing. Secondly, the k-means SMOTE technique is used to balance the data. Thirdly, the GRU model is applied to the extracted, purified data. Fourthly, the MLP model is also applied to the extracted, purified data. Finally, the performance of the proposed system is evaluated using various performance measures like a graphical analysis and a statistical test. To verify the consistency of the proposed hybrid system, the dataset is trained and tested using three different ratios. The study’s results show that the proposed hybrid system for ETD is highly accurate and efficient compared to other models including Alexnet, GRU, BGRU and RNN. In our case, Hybrid MLP-GRU has solved complex nonlinear problem. It handles large amounts of input data and makes quick predictions after training. The same accuracy ratio can be achieved even with smaller samples.

The main advantage of this research is introducing an effective ETD model for power utilities, which enables them to minimize financial loss. Furthermore, the accurate and prompt detection of energy thieves decreases line losses in transformers and other grid components. The suggested model also has certain drawbacks. The model can only be trained using high-frequency data on electricity consumption, which restricts its ability to capture the minutest electricity consumption trends. Its accuracy was also reduced as a result, more cases of misclassification occur. Furthermore, there is no method for tweaking the hyper parameters, so it can take much computing time. In the future, we will improve its performance by using the minimum frequency dataset and decreasing the delay in identifying electricity theft in the CNGC dataset. Finally, the authors believe that the system proposed in this work can be extended to the ETD in other parts of the country and the world.