Electricity theft (ET) poses significant economic and operational risks to smart grids. While deep learning approaches have advanced electricity theft detection (ETD), they often face fundamental architectural limitations. Specifically, fixed-scale feature extraction struggles to simultaneously capture fine-grained anomalies and long-term periodic trends. Furthermore, reliance on rigid positional information limits generalization across temporal shifts, while standard pooling operations frequently discard critical localized diagnostic signals. To overcome these challenges, this paper proposes the Multi-Scale Transformer (MST). This framework integrates three core innovations: (1) a Hierarchical Feature Representation module that utilizes Temporal Slice Tokenization to extract features at progressively varying resolutions; (2) a positional-encoding-free design that prioritizes shift-invariant pattern recognition to enhance robustness against data irregularities; and (3) a Feature Flattening Strategy that retains comprehensive information for the final decision logic. Extensive experiments on a real-world dataset comprising 42,372 users demonstrate the efficacy of this approach. Quantitative results indicate that MST achieves consistent improvements over competitive baselines, registering increases of 15.8% in F1-score, 3.9% in AUC, and 16.7% in MAP. • A Hierarchical Feature Representation is proposed to synergistically capture multi-scale theft patterns. • A Positional-Encoding-Free Design enhances robustness to real-world data irregularities. • A Feature Flattening Strategy overcomes the loss of information in feature aggregation. • MST outperforms 13 baselines, achieving results with up to 15.8% and 16.7% gains in F1-score and MAP.
Zhang et al. (Tue,) studied this question.