LLM-Based Electricity Theft Detection

Electricity theft remains a persistent challenge for distribution companies, leading to substantial revenue losses and jeopardizing grid stability. Leveraging Artificial Intelligence (AI), especially Large Language Models (LLMs), this study rigorously compares nine state-of-the-art instruction-tuned lLMs, fine-tuned with Quantized Low-Rank Adaptation (QLoRA), on a binary electricity-theft classification task. Among all evaluated models, falcon-7b-instruct emerges as the best performer. Despite training only 0.9343% of its total parameters, it achieves an accuracy of 0.8396, precision of 0.8864, recall of 0.7712, and an F1-score of 0.8248 on the held-out test set. The proposed framework is optimized for operational efficiency, making it suitable for near-real-time monitoring of millions of meters across large-scale distribution networks. This enables faster anomaly isolation, a reduced load on downstream validation systems, and lower manual investigation costs.