Electricity Theft Detection Based on Contrastive Learning and Non-Intrusive Load Monitoring

Electricity theft has caused enormous damage to grid's safety and economy globally, bringing plentiful attention to electricity theft detection. However, the inherent problems of data imbalance, data sparsity, and data shift due to the residents' will or the seasons remain challenges. Besides, there is a lack of appliance-level information provided by non-intrusive load monitoring to improve the data graininess in electricity theft detection. To address these problems, an electricity theft detection algorithm based on contrastive learning and non-intrusive load monitoring is proposed. Firstly, a semi-supervised learning architecture composed of Gramian angular field encoding for sequence visualization and contrastive learning architecture featuring few-shot learning is established for load monitoring and initial detection considering the inherent operating characteristics of appliances. Furthermore, after the filtration of typical regular-switched appliances by Kendall's coefficient of concordance, in-depth detection of abnormal operation routines of appliances is conducted for suspicious residents, aiming to improve the fault-tolerant ability and resist the emergence of unknown electricity theft methods. Finally, the electricity theft probability is computed to confirm fraudulent users. Both the public and practical datasets are utilized to verify the effectiveness of the proposed study, and the results show an overall better performance compared to other state-of-the-art algorithms.