Electricity Theft Detection Method Based on CNN-Mamba and Multi-Dimensional Temporal Feature Cross-Enhancement

In recent years, electricity theft has become a major issue affecting the stable operation of smart grids.Nevertheless, the effectiveness of electricity theft detection is significantly constrained by several factors.These include the high dimensionality and computational intricacy of electricity data, complex behavioral patterns of electricity users, and the number of electricity thieves far exceeds the number of normal users.For this reason, this paper first proposes a multi-scale CNN-Mamba-based temporal feature extraction method to extract timeseries electricity usage features at different time scales in electricity usage data.Then, the complementary enhancement of time-series features and statistical features of electricity consumption data is realized by a cross-attention-driven multi-dimensional temporal feature cross-enhancement method.Finally, an electricity theft detection model integrating CNN-Mamba and multi-dimensional temporal feature cross-enhancement is constructed.Ablation and comparative experiments were conducted on real-world power datasets.Through experimental comparisons, statistical significance tests, model complexity analysis, and deployment performance evaluations, this approach demonstrated significant effectiveness and superiority in electricity theft detection.