A Review of Machine Learning-Based Theft Detection for Smart Energy Meters

Abstract

The implementation of Advanced Metering Infrastructure (AMI) in smart grid environments has led to a paradigm shift in energy consumption management by means of high-throughput bi-directional communication. Unfortunately, the adoption of AMI is associated with growing Non-Technical Losses (NTL) such as electricity theft, estimated at global losses exceeding $96 billion a year, threatening the stability and resilience of electricity grids. As a response to the emerging challenge, traditional NTL detection, based on periodic field audits and threshold-based inspections, has been proven to be ineffective in addressing stochastic and increasingly intelligent cyber-physical manipulations. Hence, this paper provides a literature review of Machine Learning applications in the domain of Electricity Theft Detection (ETD) that is grounded in 71 scholarly studies produced between 2021 and 2026 and selected according to criteria of machine learning application in the process of theft/NTL detection and accompanying empirical validation. Out of those, 44 primary research works have been organized into tabulation and contrasted with regard to eight distinct categories of methods including classical ML, deep learning, hybrid and reinforcement and federated learning techniques. The comparative evidence reveals that the hybrid and stacked architectures provide the best detection performance whereas federated designs typically sacrifice their maximum accuracy in favor of better privacy and deployability. Challenges in assessing the comparative gains are discussed, and it is revealed that benchmarked datasets are scarce and inconsistent as well as only 9 of 44 studies included in the table reported their false positive rates that determine audit expenses.