A Model-Agnostic Approach to Reveal Detection Basis and Occurrence Time of Electricity Theft

In recent years, a large number of machine learning models have been widely used in electricity theft detection tasks, but most of them belong to black-boxes lacking interpretability, which poses challenges for users to trust their detections. To address this issue, this article aims to provide interpretations for machine learning models in electricity theft detection. In particular, a theoretical analysis is performed to explain the necessity for interpretability in machine learning models for electricity theft detection, highlighting the motivations from a measurement science perspective. Then, a model-agnostic explainable artificial intelligence (XAI) approach is proposed to reveal the possible occurrence time of electricity theft events and the detection basis of machine learning models. Simulation results show that the proposed XAI approach is able to accurately capture the detection basis and occurrence time of electricity theft events, which enables suspects to comprehend why they are identified as potential thieves and reduces unnecessary investigations due to false alarms. Furthermore, when compared to existing XAI approaches (e.g., local interpretable model-agnostic explanation (LIME) and permutation feature importance (PFI) technique), the interpretations provided by the proposed XAI approach are more trustworthy across numerous machine learning models and datasets. This contributes to the advancement of the field of instrumentation and measurement by providing a robust framework for the interpretation of machine learning-based measurements and detections of electricity theft, promoting better integration of artificial intelligence into measurement systems.