Graph Convolutional Network-Based Federated Learning Method for Distribution System Topology Identification

Distribution system topology identification is the basic function for system operation but is hindered by data privacy concerns and uncertainties arising from high penetration of renewable energy sources (RES). To tackle these issues, a graph convolutional network-based federated learning method (GraphFed) is proposed for topology identification while ensuring data privacy. Specifically, to remove the data leakage risk, the federated learning (FL) algorihm is proposed for topology identification in a decentralized learning framework. Besides, the proposed graph convolutional network (GCN) leverages a novel node embedding-based graph shift operator to automatically represent the graphical features of the power system data to achieve enhanced topology identification accuracy. Comparative experiments are conducted on the IEEE 33-node distribution system, demonstrating the effectiveness of the proposed GraphFed method.