A GAN-Based Power Quality Anomaly Detection Method for Imbalanced Multivariate Time Series

Time series anomaly detection is an important and well-researched topic with extensive applications in finance, healthcare, biological data, and industrial production. However, in the field of power quality, detecting anomalies in complex and imbalanced multivariate long time series data remains a challenge. This paper proposes a GAN-based method for multivariate time series anomaly detection, which directly trains on raw data captured by power monitoring stations. It addresses the issues of the lack of original examples, difficulty in acquisition, and simulation of abnormal data in power quality. Additionally, the proposed method introduces an attention mechanism and a multi-scale feature fusion mechanism, reducing the accumulation of errors during model training and effectively addressing the problem of overfitting due to the scarcity of abnormal samples and high noise in power quality data. The comprehensive experiments conducted on four original power quality datasets demonstrate that the proposed method significantly improves the performance of anomaly detection in the field of power quality compared to baseline methods.