The advent of smart grids has facilitated data-driven methods for detecting electricity theft, with a preponderance of research efforts focused on user electricity consumption data. The multi-dimensional power state data captured by Advanced Metering Infrastructure (AMI) encompasses rich information, the exploration of which, in relation to electricity usage behaviors, holds immense potential for enhancing the efficiency of theft detection. In light of this, we propose the Catch22-Conv-Transformer method, a multi-dimensional feature extraction-based approach tailored for the detection of anomalous electricity usage patterns. This methodology leverages both the Catch22 feature set and complementary features to extract sequential features, subsequently employing convolutional networks and the Transformer architecture to discern various types of theft behaviors. Our evaluation, utilizing a three-phase power state and daily electricity usage data provided by the State Grid Corporation of China, demonstrates the efficacy of our approach in accurately identifying theft modalities, including evasion, tampering, and data manipulation.

Electricity theft is a major source of non-technical losses in distribution networks, threatening both economic revenues and power supply reliability. This study addresses the identification of nodes exhibiting anomalous load behavior (anomalous nodes) in 10 kV distribution feeders. Based on the IEEE-33 bus benchmark system, the disturbance patterns induced by abnormal consumption are analyzed. The results show that voltage and current fluctuations intensify with increasing electrical distance from the power source, while branch loss peaks localize at the affected terminals and propagate unidirectionally along the power flow path. Building on these findings, an improved density-based spatial clustering of applications with noise (DBSCAN) method is proposed, integrating five spatial network features and sixteen temporal electrical features extracted from voltage, current, and power series. Prior to clustering, the features are standardized and reduced via principal component analysis (PCA), retaining over 90% of the cumulative variance. Validation on a hybrid dataset demonstrates that the proposed method achieves 90.7% accuracy, 87.5% recall, and an F1-score of 0.895, outperforming traditional K-means and approaching supervised CNN models without requiring labeled data. These results confirm the method’s robustness and suitability for practical deployment in distribution networks.