Channel Independent Attention Network for Electricity Theft Detection

Electricity theft severely impairs the economic benefits to businesses and endangers public safety. In recent years, deep learning models for automated electricity theft detection have been advancing rapidly. However, the pattern of electricity usage data exhibits complex dependencies in cases of electricity theft, posing significant challenges for accurate detection. To cope with this difficulty, we propose a Channel Independent Attention Network (CIAN) for electricity theft detection. In particular, electricity consumption data is firstly processed by a sequence patchfication strategy, which enables dependency capturing over suitable time intervals. A parallel structure is then constructed, where a self-attention based channel independent encoder and a mix convolution module are cooperated for learning global and local feature patterns, respectively. Lastly, the features are fused to predict whether the electricity consumption sequence data involves electricity theft. Experimental results on a real-world electricity dataset demonstrate that the proposed method surpasses other competitive electricity theft detection approaches across various evaluation metrics.