Distribution Fault Location Using Graph Neural Network with Both Node and Link Attributes

This paper proposes a graph neural network (GNN) based approach to locate fault spots within the power distribution systems. The GNN model is extended from graph convolutional networks (GCNs), and includes a set of graph processing layers and followed by a set of dense connected layers. The graph processing layers incorporated with node and link attributes are used to map system topology, bus measurements and branch parameters into hidden node embeddings, and dense connected layers are used to relevant fault locations to node embeddings. The node attributes of the graph include measured phase voltage and current measurements, and branch impedance, admittance and regulation parameters are integrated into link attributes of the graph. The fault locations are represented as output features of nodes for the graph, in which only terminal buses of faulted branch have non-zero values corresponding to faulted phases. The developed approach is applicable to both short-circuit faults and ground faults. Numerical examples are given to demonstrate the effectiveness of the proposed method. The test results showed that the faulted section can be estimated with high percentage of accuracy, and more importantly there is no need for re-training in case of topology changes.