MSDM: Multi-Scale Differencing Modeling for Cross-Scenario Electricity Theft Detection

Abstract

Electricity theft emerges as a critical concern within non-technical losses in power systems, resulting in substantial economic losses and posing a substantial threat to the power grid’s reliability. Traditionally, characteristics utilized in electricity theft detection (ETD) models are tailored to specific application scenarios. However, consumption behavior’s intricate and variable nature across diverse regions, seasons, and user types introduces notable scenario-dependent variations. Consequently, existing methods experience substantial performance degradation when the application scenario changes. We propose an innovative framework for multi-scale differencing modeling (MSDM) to address this challenge. Our approach introduces a multi-scale differencing mechanism to alleviate the impact of scenarios, effectively decoupling electricity consumption behavior from specific scenarios. Additionally, we incorporate an adaptive quantization module to suppress scenario-related noise in various sequences. Concurrently, a multi-scale convolution network is designed to capture intra- and inter-period features, ultimately enhancing the generalization and adaptability of the ETD model across different scenarios. Extensive experiments conducted on a series of cross-scenario and real-world benchmarks highlight the advantages of our method over state-of-the-art techniques. The average AUC reached 0.979 and 0.911, respectively. The demonstrated improvements in performance underscore the effectiveness of our MSDM in addressing the challenges posed by varying application scenarios and real-world complexities.