Deep learning-based electricity theft detection for active distribution networks with high photovoltaic penetration

Abstract

The active distribution network (ADN) with high photovoltaic penetration faces increasingly serious electricity theft threats at present. However, existing electricity theft detection is still limited by the data imbalance, the malicious alteration on photovoltaic power generation data, and the confusion between significant irregular electricity consumption fluctuations of normal and fraudulent power consumers. Therefore, a novel data-driven electricity theft detector based on deep learning is proposed in this paper. Firstly, the proposed detector contains two balanced temporal generative adversarial networks (BTGANs), which synthesize abnormal samples involving diverse electricity theft behaviors to address the data imbalance. Secondly, the photovoltaic electricity theft detector (PETD) in the proposed detector perceives the malicious alteration on photovoltaic power generation data, and thus optimizes detection on electricity theft covered up by falsified photovoltaic power generation data. Thirdly, the temporal graph convolutional neural network (TGCNN) in the proposed detector captures unnatural changes in correlations between temporal electricity consumption features of interrelated power consumers as distinctive electricity theft indicators to eliminate the confusion. Finally, the performance of the proposed detector is evaluated by experimental research on actual power data collected from industrial estates in China. In the experimental research, the proposed detector achieves high accuracy of 97.64 %, a high detection rate of 98.44 %, and a false positive rate as low as 2.40 %, which indicates that it is more successful than existing electricity theft detectors.