Machine Learning Techniques for Electricity Theft Detection in Smart Grids: A Comprehensive Review

Abstract

Electricity theft remains a critical threat to power distribution infrastructure globally, with annual losses exceeding USD 89 billion and non-technical loss rates reaching 40% in developing economies. While machine learning has emerged as the dominant analytical approach for automated theft detection in smart grid environments, the field lacks a unifying framework that connects algorithm selection to the operational realities of Distribution System Operators (DSOs). Existing reviews catalogue methods and report benchmark metrics without addressing how detection paradigm selection should be aligned with data maturity, regulatory requirements, computational constraints, and institutional capacity. This review addresses that gap by systematically analysing 90 peer-reviewed studies published between 2015 and 2025, identified through structured multi-database searches, screened against explicit eligibility criteria, and graded with a formal five-criterion quality rubric, through a unified adversarial time-series formulation that provides a consistent analytical lens across all major learning paradigms. The analysis covers supervised ensemble methods, unsupervised and semi-supervised anomaly detection, deep learning architectures, including convolutional neural networks, long short-term memory networks and Transformer models, graph neural networks, federated learning, and explainable artificial intelligence. Key findings reveal that no single paradigm achieves optimality across all deployment dimensions simultaneously, that gradient boosting methods deliver near state-of-the-art performance with significantly lower computational overhead than deep learning, and that hybrid architectures achieve AUC-ROC scores of 0.95 to 0.98 on benchmark datasets but require complementary governance mechanisms to satisfy regulatory defensibility requirements. A lifecycle-aligned deployment framework and a layered detection architecture are proposed, offering practitioners a structured pathway from early AMI rollout through to advanced smart grid deployment. The principal outcomes of the review are a formal characterisation of which component of the detection problem each learning paradigm estimates, quality-graded and harmonised benchmark performance ranges, and a quantified illustrative analysis indicating that the proposed layered architecture can improve inspection productivity by roughly an order of magnitude at a fixed field budget. Four priority research challenges are identified: real-time edge detection, continual learning, multi-modal data fusion, and standardised benchmarking.