Answers are grounded in this paper and cite their sources.
Unsupervised analysis of multivariate electricity consumption time series based on machine learning has become a core technical approach for electricity theft detection. Without relying on manually labeled theft samples, such methods can mine the inherent correlation and anomaly patterns hidden in electricity consumption data, thereby supporting the identification of theft behaviors under relaxed supervised conditions. However, mainstream unsupervised detection methods rarely consider the multi-temporal fluctuation characteristics of electricity data and the coupling relationships among multiple consumption variables, which to some extent weakens the generalization ability of algorithms across different data sources. As a result, it is difficult to effectively identify stealthy theft behaviors such as meter-tampering electricity theft. To tackle the above challenges, this paper proposes a Multi-Scale Graph Attention Multi-Head Aggregation Autoencoder (MSGME). The model extracts electricity consumption features at multiple temporal scales through multi-scale modules and constructs them into multi-scale graphs. On this basis, a contrastive graph attention network is designed to strengthen the correlation modeling between multiple consumption variables, and multi-head attention aggregation is integrated to improve feature representation capability. Under the unsupervised learning paradigm, the proposed MSGME realizes an effective unsupervised training and detection workflow by combining reconstruction loss and contrastive loss. Experimental results on real-world electricity consumption datasets verify the effectiveness and superiority of the MSGME model in electricity theft detection.
No takes yet. Share an insight, caveat, or question.
Ai et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: