Advanced electricity theft detection in smart metering systems via Channel-Correlation enhanced hierarchical kernel networks and matrix completion

Electricity theft can result in attacks and tampering with advanced metering infrastructure. Although electricity theft has decreased with the widespread adoption of smart meters, significantly increasing the amount of measured data, the issue persists. This paper presents a novel method termed channel-correlation exploited hierarchical kernel network to address the problem of electricity theft detection, integrating matrix completion and channel optimization techniques. Initially, the proposed method addresses the issue of missing or abnormal original data by employing the alternating direction method of multipliers, enhancing the quality of data samples for training purposes. Subsequently, the Hierarchical Kernel Network utilizes different kernel sizes to extract diverse feature sets, thereby capturing comprehensive information to improve recognition accuracy. Furthermore, leveraging the channel correlation exploitation module of the compression and excitation network, the network effectively analyzes and learns the unique features of each channel, significantly enhancing classification performance. Through ablation studies and experiments conducted with varying proportions of missing data and different fraud rates, the proposed model consistently demonstrates superior performance across all metrics compared to other models. The practicality and effectiveness of the model are further validated through implementation on hardware platforms. These findings provide robust evidence of the model's efficacy and superiority in detecting electricity theft.