Physics-Informed Learning for High Impedance Faults Detection

High impedance faults (HIFs) in distribution grids may cause wildfires and threaten human lives. Conventional protection relays at substations fail to detect more than 10\% HIFs since over-currents are low and the signatures of HIFs are local. With more $\mu$PMU being installed in the distribution system, high-resolution $\mu$PMU datasets provide the opportunity of detecting HIFs from multiple points. Still, the main obstacle in applying the $\mu$PMU datasets is the lack of labels. To address this issue, we construct a physics-informed convolutional auto-encoder (PICAE) to detect HIFs without labeled HIFs for training. The significance of our PICAE is a physical regularization, derived from the elliptical trajectory of voltages-current characteristics, to distinguish HIFs from other abnormal events even in highly noisy situations. We formulate a system-wide detection framework that merges multiple nodes' local detection results to improve the detection accuracy and reliability. The proposed approaches are validated in the IEEE 34-node test feeder simulated through PSCAD/EMTDC. Our PICAE outperforms the existing works in various scenarios and is robust to different observability and noise.