Power Distribution Network Topology Detection Using Dual-Graph Structure Graph Neural Network Model

Abstract

Topology detection (TD) in the context of power distribution networks (PDNs) is a fundamental requirement for a wide range of applications, such as fault localization and load management. PDNs suffer from a lack of real-time topological information due to insufficient data on switch statuses and an increasing number of switching actions caused by reconfigurations and the control of distributed energy resources (DERs). On this basis, in this paper, a novel near real-time TD method for PDNs is proposed. This method is built on a specialized graph neural network (GNN) design using data from micro-phasor measurement units ( $\mu $ PMUs), leveraging the strengths of both graph-based learning and conventional deep learning (DL) approaches. More specifically, the developed TD method implements a novel dual-graph structure GNN (DGS-GNN) model to transform the TD problem into an inductive link prediction task for a multi-graph dataset. During the training phase, a node attribute similarity graph is created, and the resulting node embeddings are aligned with the actual topology graph (ATG) using a structure-aware loss function. In the inference phase, however, unlike standard GNN models that require structural information as input, the ATG is recovered based solely on node attributes. The developed method enables TD using a limited number of phasor measurements with low inference time and superior generalization capability for unseen scenarios. Its strong performance in large-scale PDNs with varying configurations, as well as its robustness to uncertainties from DERs and noisy environments, is demonstrated on the IEEE 33- and 123-Bus benchmarks and a standard 240-Bus test system. The proposed method outperforms its DL-based counterparts in scenarios where full or partial system topology should be detected.