Unsupervised line loss anomaly detection based on multi-resolution temporal modeling

Accurate identification of line-loss anomalies is crucial for safeguarding the economic benefits of power utilities and ensuring the stable and secure operation of the power grid. However, existing methods struggle to effectively capture complex multi-period temporal patterns in electricity consumption data and heavily rely on anomaly labels that are often difficult to obtain. To address these challenges, this paper proposes an unsupervised end-to-end line-loss anomaly detection algorithm based on multi-resolution temporal modeling. The method first performs unsupervised clustering on unarchived users to select candidate users representing significant feeder characteristics. It then applies multi-resolution temporal modeling to process cross-feeder user data, eliminating magnitude differences to focus on pattern shapes and capturing both global and local features of electricity consumption. The dual-path attention feature extraction and fusion module deeply mines multi-scale temporal patterns embedded in user electricity data. Through multi-scale and multi-resolution integration, the model adaptively fuses information across scales and periods, constructing a powerful representation learner for electricity consumption behavior patterns. On a real-world dataset, the proposed algorithm achieves an accuracy of 98.58%, significantly outperforming existing baselines, while demonstrating excellent cross-transformer generalization and robustness to noise. This method provides power utilities with a label-free, highly generalizable, and effective solution for precise line-loss management.