Temporal Convolutional Networks for Electricity Theft Detection With Limited Data

Electricity theft remains a critical challenge for utilities globally, with developing nations like India reporting significant annual losses due to non-technical losses (NTLs). Conventional detection methods-often reliant on manual meter audits or seasonal usage thresholds-fail to address the adaptive tactics of modern theft schemes, which exploit gaps in lowresolution consumption data. Our study tackles this gap by integrating a Temporal Convolutional Network (TCN) with domain-specific preprocessing: Gaussian noise injection mitigates sensor-level volatility in rural grids, while Synthetic Minority Oversampling Technique (SMOTE) counteracts the class imbalance. Deployed on a 4.2 million-meter EU benchmark dataset, the framework achieved 97 % accuracy and 96 % F1-score, reducing false positives by 22 % compared to LSTM-based benchmarks. A 2023 pilot in São Paulo's high-theft zones demonstrated actionable outcomes: <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 9 \%}$</tex> reduction in NTLs within 6 months, alongside improved dynamic load management for 41,000 households. This approach equips utilities with scalable, context-aware tools to curb revenue leakage while preserving grid stability.