Abstract
Non-Technical Losses (NTLs), such as electricity theft or metering irregularities, remain a persistent challenge for power distribution systems, leading to substantial financial losses and operational inefficiencies. In this study, we extend previous work on NTL detection by incorporating explainable machine learning (ML) techniques and evaluating emerging model architectures. Using a real-world dataset from the Hellenic Electricity Distribution Network Operator (HEDNO), we perform an in-depth analysis of consumption data collected from traditional meters. Beyond conventional models, we evaluate the potential of modern transformer-based architectures in capturing complex sequential trends in consumption. A key novelty of this work lies in its focus on explainability: we integrate Shapley values and Permutation Importance to interpret model predictions, offering actionable insights to utility providers and domain experts. Our experimental results demonstrate that ensemble models, particularly CatBoost, achieve strong performance in detecting NTLs, with confirmed NTL cases accounting for over 60% of the top-ranked 10% of users. Furthermore, we show that comparable predictive performance can be maintained using a reduced set of carefully selected features, enhancing both efficiency and explainability.