An Electricity Theft Detection Based on Image Coding Technique and ConvNeXt

The deep learning (DL) techniques have been an effective means for electricity theft detection. However, most existing works are based on 1-D time series data, which makes it challenging to fully capture and utilize the multidimensional features of the electricity data. Therefore, a novel electricity theft detection approach based on Gramian angular summation field (GASF) and ConvNeXt is proposed. Specifically, this article designs a GASF-based image coding strategy to convert 1-D electricity consumption data into 2-D images. In this case, the multidimensional features of electricity consumption behavior can be preserved in the GASF images. And then, the image classifier is designed that utilize ConvNeXt as a feature extractor to learn multidimensional features of GASF images, thereby achieving electricity theft detection. Experimental results based on a real-world dataset show that the proposed detector reaches an accuracy of 0.973, which is superior to state-of-the-art methods. Furthermore, the detection performance of different anomaly detectors against smart electricity theft behaviors is evaluated, and the economic losses (ELs) they can recover for power companies are quantified. The results indicate that the proposed detector, by virtue of its strong capability in identifying interclass differences among electricity theft behaviors, can thereby minimize the ELs sustained by power companies.