Detecting Non-Technical Losses in Smart Grids Using Statistical Distances of Forecasting Residuals

Abstract

Energy theft poses a significant challenge to the sustainability of smart grids, affecting the financial stability of electrical utilities and the overall management of resources. In this thesis, we present a novel load forecasting residuals based framework for detecting electricity theft using forecasting models and statistical distances. Our approach involves forecasting daily energy consumption and generating residuals, which are then compared to residuals from normal days using a statistical distance. Aday with a statistical distance exceeding a predetermined threshold is flagged as a potential energy theft incident. The proposed framework can be seamlessly integrated into existing forecasting models with minimal computational overhead for statistical distance calculation. Additionally, the framework offers high explainability, substantially reducing the costs associated with false positives. We evaluated the performance of our approach using two publicly available datasets, testing its ability to detect twelve energy theft attack models and faulty meters. Our framework demonstrated better performance than the state of the art on two different forecasting models across two different datasets.