Distribution System Anomaly Detection Based on AnoGAN Embedded with Cross-Stitch Units

We consider the anomaly detection problem in distribution systems. Many distribution systems are not observable due to their limited numbers of real-time meters, thus state estimation based approaches are not applicable. Moreover, the number of abnormal samples are usually small comparing to the diversity of anomalies, imposing challenges to machine learning based methods as well. To this end, AnoGAN, an unsupervised learning method, is applied to the distribution system anomaly detection in this paper. Only normal samples are needed during its training process. In the testing process, it generates corresponding samples based on its knowledge trained from normal ones, which are compared against real measurements to detect if there are possible anomalies. To achieve better granularity, we propose to partition a large distribution system into sub-networks, establish parallel AnoGANs, and employ Cross-stitch units to incorporate their correlations. Simulations have been done to show the satisfactory accuracy and efficiency of the proposed approach in detecting anomalies in distribution system.