Two-stage generalizable approach for electricity theft detection in new regions

As the current mainstream method for electricity theft detection, the data-driven detection model's performance noticeably deteriorates in new regions due to variances in electricity consumption characteristics. Moreover, owing to the lack of theft labels in new regions without on-site inspections, training a high-accuracy theft detection model becomes challenging. To address these issues, this study proposes a novel two-stage generalizable approach for electricity theft detection in new regions. In this approach, Stage 1 presents an adaptation method based on Deep Subdomain Adaptation Network (DSAN). This stage precisely aligns electricity consumption features between different regions. It transfers the characteristics of electricity consumption curves from the established region to the unlabeled new region, successfully adapting the model to the latter. Next, in Stage 2, an enhancement method based on Core-Set Active Learning (CSAL) is proposed to select the representative electricity consumption samples in the new region for fine-tuning. At the early stage of on-site inspections, Stage 2 rapidly and stably enhances the model's performance within limited costs, providing reliable guidance for regular on-site inspections. Stage 1 innovatively transfers a lot of learned features for the theft detection model, and Stage 2 further learns new regions' representative samples combined with inspections for the first time. They complement each other in terms of rich information on electricity consumption behaviors and new regions' electricity consumption characteristics. Finally, experiments on real-world datasets in China demonstrate the superiority of our proposed approach. The results show its high detection performance and rapid enhancement in unlabeled new regions, providing strong support for anti-electricity theft work.