Multiclass Non-Technical Loss Detection with Reduced Synthetic Data via a Two-Stage Hierarchical 1D-CNN

Abstract

Non-Technical Loss (NTL) detection is a critical challenge in power distribution systems, particularly in regions dominated by conventional mechanical meters and severe class imbalance, as NTL causes substantial revenue loss and undermines system reliability. This paper proposes a two-stage hierarchical 1D-CNN framework for multiclass NTL classification using real-world monthly consumption data from the Provincial Electricity Authority (PEA) in Khon Kaen, Thailand, where over 90% of meters are manually read and NTL rates are the highest in Northeast Thailand. The first stage performs three-class classification (Normal, Energy Theft, Defective Meter) directly on raw imbalanced data, achieving 100% recall for defective meters without synthetic augmentation. The second stage applies Synthetic Minority Over-sampling Technique (SMOTE) selectively to the filtered theft class, reducing synthetic data volume by approximately 50% - thus lowering computational cost -while preserving 99.95% theft recall. Evaluated across four experiments against SVM, Random Forest, XGBoost, and LSTM baselines, the proposed method eliminates manual triage by enabling direct operational routing: repair dispatch for defective meters and legal investigation for energy theft. By minimizing artifact bias and overhead, this work bridges the gap between academic binary models and real-world utility requirements, delivering a robust, deployable solution for NTL management in resource-constrained environments.