A Hybrid K-Means-LSTM-XGBoost Framework for Low False-Positive Electricity Theft Detection in Power Distribution Networks

Electricity theft causes approximately $96 billion in annual global losses, threatens grid stability, and creates public safety hazards. Existing detection methods suffer from unacceptably high false positive rates (10–15%), making them impractical for real-world deployment where wrongful accusations carry high operational and reputational costs. This paper proposes a novel hybrid framework, K-MeansLSTM-XGBoost (K-MLX), that integrates unsupervised clustering, deep learning-based temporal feature extraction, and ensemble classification to address this critical gap.