Abstract:Accurate detection of electricity theft incidents is critical to the stable operation of power systems and the economic interests of power utilities. Traditional electricity theft detection methods mainly model one-dimensional load data, which neither fully explores the periodic characteristics of load sequences nor takes full advantage of multi-source data for auxiliary analysis, resulting in restricted pattern mining capability and failure to meet detection requirements under complex operational scenarios. To address these deficiencies, this paper proposes a multi-source collaborative intelligent electricity theft detection method integrated with multi-scale periodic features. This approach first analyzes the inherent correlation between multi-source meteorological data and load profiles to select highly correlated influencing factors, conducts spectral clustering (SC) by combining the screened factors with load data, and identifies suspected electricity theft days according to clustering results. Meanwhile, realistic theft samples are constructed based on the mathematical formulation of multiple predefined electricity theft modes. Subsequently, a channel attention weighting mechanism is introduced to refine valid information from multi-source data, and grouped convolutions are adopted to extract local spatio-temporal features. A dynamic gating residual fusion strategy is further designed to facilitate adaptive trade-offs between raw load features and enhanced features. In addition, the autocorrelation function is utilized to conduct periodicity detection, the time series is explicitly divided based on periodic attributes, and independent modeling is implemented for intra-period and inter-period features. Finally, an improved Inception-based multi-scale convolutional neural network (CNN) module is adopted to achieve accurate identification of electricity theft behaviors. Experimental results on the UMass Smart* and SGCC datasets demonstrate that the proposed method can fully exploit the discriminative potential of periodic features and multi-source information. It achieves a detection accuracy of 95.18%, a recall rate of 93.83%, an F1-score of 95.07%, and a low false positive rate of only 0.85%. Comparative analyses with other state-of-the-art detection methods confirm that the proposed strategy delivers superior comprehensive performance and possesses stronger discriminative capability for complex electricity theft.