p26 / Journal
Explainable machine learning for non-technical loss detection in power systems
Sustainable Energy, Grids and Networks 47, 102479 · 2026
In plain language
Explainable learning helps utilities identify non-technical losses, including electricity theft and metering irregularities.
Research context
Ranking customers with transparent explanations can support targeted inspection and more efficient utility analytics.
Approach
Traditional meter records from HEDNO are analyzed with conventional and transformer models. Shapley values and permutation importance explain predictions.
Findings
CatBoost performs strongly: confirmed losses comprise over 60% of the highest-ranked 10% of users. Carefully reducing the feature set preserves comparable predictive performance.
Summary based on abstract.