Table 1.

Literature overview on energy theft detection based on consumption data.

Ref.	Year	Platform	Proposed Model	Dataset	Accuracy	Presented Main Contribution
[31]	2019	N/A	CNN-LSTM based	SGCC	89	The irregular and abnormal consumption patterns of consumers are analyzed
[34]	2018	N/A	Clustering based	ISSDA	74 (AUC)	Malicious examples are not needed to train the method for future detection
[28]	2020	Python 3.x	CNN-GRU-PSO	SGCC	89	Preprocessing steps, feature selection, feature extraction, and classification are performed using a lot of techniques and the proposed model outperforms imbalancing issue
[27]	2020	N/A	MP-ANN	ISSDA	93.4 (DR)	Self-organizing is used for clustering the consumers according to similar consumption patterns, i.e., classification as honest or malicious. The number of transformers that have suspect consumers is reduced without the need to install measurement units on all transformers
[32]	2020	Python 3.x	CNN-LSTM	SGCC	87.9	An efficient solution to overcome imbalanced data, overfitting, and high-dimensional data limitations is introduced
[24]	2016	N/A	DT-SVM	OpenEnergy	92.5	The newly proposed system exhibits the capability to accurately identify instances of energy theft in real time across all stages of power transmission and distribution
[29]	2018	Python 3.x	GRU (RNN-based)	ISSDA	92.5 (DR)	Temporal patterns are utilized in energy consumption, and a GRU-based RNN enhances detection performance, optimizing hyperparameters through a random search analysis in the learning phase
[23]	2016	N/A	SVM-based	ISSDA	94 (DR)	Six different attack vectors are designed to obtain manipulated consumption data
[25]	2021	Python 3.x	Ensemble ML	ISSDA	90 (AUC)	Data pre-processing is used to address imbalanced data with SMOTE and Near-miss techniques, achieving optimal detection rates through bagging-type ensemble ML demonstrated with diverse consumer samples
[30]	2018	N/A	Wide and Deep CNN	SGCC	80 (AUC)	Unlike existing methods tailored for one-dimensional data, wide and deep CNN handles detecting electricity theft by effectively capturing both periodic and non-periodic consumption patterns in two-dimensional data
[33]	2020	N/A	LSTM-based	SGCC	93.6	A new technique is devised to streamline data, enhancing usability and facilitating the extraction of meaningful insights from the dataset
[26]	2021	Python 3.x	Neural Network	Grid LabD Tool	93	A novel method is introduced for detecting electricity theft, focusing on “balance attacks” with prosumers manipulating readings for total aggregated balance. A cluster-based detection model is introduced as a middle-ground approach, bridging the gap between using a single model for all users and individual models for each user
[36]	2021	Matlab2019	CNN-WeightedRF	Mathpower Tool	95.71	An FDI intrusion-detection model combining CNN and weighted RF is able to detect the spurious data more accurately compared with other detection models
[37]	2015	N/A	SVM	ISSDA	75.8	The classification models simplify a demand-side management study, analyze tariff methods, and offer insights for policymakers
[38]	2010	VisualBasic	SVM	Tenaga Nasional	60	This work aims to aid Tenaga Nasional Berhad Distribution in Malaysia to reduce NTLs within the distribution sector caused by electricity theft
[39]	2018	Python 3.x	DNN-based	ISSDA	92.6 (DR)	This work proposes a DNN-based customer-specific detector that can mitigate electricity theft cyber-attacks
[40]	2017	N/A	Density-based clustering	ISSDA	93.2	This work exhibits superior performance compared to alternative methods across nearly all categories of theft
[41]	2022	Python 3.x	Attention LSTM Inception	SGCC	95	This work addresses the elevated FPR issue arising from widespread misclassification, leading to financial burdens
[42]	2022	Python 3.x	KTBoost Classifier	SGCC	93.38	Taking into account all minority sample regions in the dataset, the robust-SMOTE technique generates minority class samples with reduced susceptibility to overfitting and the generation of noisy samples
[43]	2023	Python 3.7	Deep-CNN	Researcher-generated	95	The proposed theft detection method, utilizing the SMOTE technique to generate minority class samples with reduced susceptibility to overfitting and noise, attains the highest accuracy compared to all other studied methods
Our work	2024	Python 3.10	CNN-based	ISSDA	95.34	CNN-based architecture is combined with traditional ML methods. A detector that provides high success in detecting all attack vectors has been designed. The imbalanced data problem was solved using GAN.