Non-Technical Losses Detection in Electric Distribution Systems Using BERT and GAN

Abstract

Non-technical losses have caused lots of revenue loss in many electric utility companies around the world. In current practices, manual analysis on collected power consumption data first. Then, on-site inspections are conducted. With recent advances of machine learning techniques, several works have been developed to solve this task in a more effective manner. However, most existing machine learning approaches still require the feature extraction step. Moreover, most current studies overlook the imbalanced dataset from consumer's power meters. This paper proposes a new approach to deal with these problems by integrating two deep machine learning techniques. First, we use Bidirectional Encoder Representations from Transformers (BERT) to remove the feature extraction step. Meanwhile, the generative adversarial network (GAN) is considered to generate fake data to increase the number of the minority class in the imbalanced dataset. The effectiveness of the proposed method has been evaluated on various metrics. Experimental results demonstrated that the proposed method can indeed improve the recall and F1-score significantly.