Abstract and subjects
Fault diagnostics play a crucial role in determining the appropriate actions to take when dealing with faults in a power system. The increasing integration of inverter-based distributed energy resources has had a significant impact on fault detection, particularly with traditional overcurrent relays. This chapter introduces a novel temporal recurrent-graph neural network (R-GNN) approach that leverages emerging graph learning techniques for fault diagnostics. These models utilize voltage measurement data from critical buses to extract spatial-temporal features. Through these features, the proposed models detect fault events, classify fault type/phase, and determine fault locations. Compared with previous approaches, the temporal R-GNNs have the potential to be superior to other methods in generalization capabilities for fault diagnostics. Additionally, such a scheme potentially eliminates the need for installing current sensing transformer on all lines of the distribution system by retrieving voltage signals instead of currents. This characteristic allows for greater scalability and flexibility since the method is not constrained by the number of installed relays for fault diagnoses. This chapter will present a couple of case studies, including the practical 13-node Potsdam microgrid in Potsdam, New York, and the standard IEEE 123-node system for comparative studies of the spatial-temporal method against other neural network structures. The studies reveal the enhancement of the new approach in the power system fault diagnosis by enabling more accurate fault detection, classification, and localization, while reducing infrastructure requirements and offering improved performance.