Distribution System Topology Identification Using Graph Neural Networks

Topology identification is a critical precursor to several distribution system analyses, such as state estimation, power flow analysis, and volt/VAR optimization. However, the inability of utilities to capture the status of all switches due to monetary and resource constraints necessitates developing topology identification strategies. While several methods have been proposed to identify network topology, they have certain draw-backs. Some limitations include evaluating all possible topologies and large computational times, rendering them unsuitable for real-time application. Therefore, this paper proposes a graph neural network (GNN) approach for topology identification in distribution systems to overcome these limitations. The topology identification problem is formulated as a link-prediction problem on the graph representation of distribution systems. The status of each switch is predicted by edge scores computed for the link connecting the two nodes in the graph representation of the system. Bus measurements, such as voltage phasor and loads, are used as node-level features in the proposed GNN. The proposed approach has been evaluated on a three-phase unbalanced 559-node distribution system (modeled on the IEEE 37-node distribution system). The approach showed comparable performance with existing techniques while utilizing significantly lower computational time.

I. Introduction

Distribution system applications such as state estimation, fault detection, and control require accurate knowledge of networks' topology. However, due to insufficient measurements and monitoring, utilities fail to capture the status of all switches in the network. Several methods proposed for state estimation and control cannot be implemented without the knowledge of network topology and parameters [1]–[5]. Therefore, topology identification techniques have received significant research interest [6]–[17]. The following subsection provides a brief review of techniques proposed for topology identification.