Generalization of Graph Neural Network Models for Distribution Grid Fault Detection

TL;DR

This work tackles fault detection in distribution grids facing evolving topologies due to switching, reconfiguration, and DER integration. It proposes an RGNN framework that combines temporal GRU encoding with spatial GNNs, specifically evaluating GraphSAGE and GATv2 against RGCN and GRU baselines on the IEEE 123-bus network. The key contributions include the first application of GraphSAGE and GATv2 within an RGNN for this task, a comprehensive generalization benchmark across varying PMU configurations, and the finding that RGATv2 offers superior robustness to topology changes. The results demonstrate that RGATv2 maintains high detection performance as monitors are added or removed, whereas pure RNNs and some GNN baselines exhibit substantial degradation, highlighting the practical potential for topology-adaptive fault detection in modern grids with DERs.

Abstract

Fault detection in power distribution grids is critical for ensuring system reliability and preventing costly outages. Moreover, fault detection methodologies should remain robust to evolving grid topologies caused by factors such as reconfigurations, equipment failures, and Distributed Energy Resource (DER) integration. Current data-driven state-of-the-art methods use Recurrent Neural Networks (RNNs) for temporal modeling and Graph Neural Networks (GNNs) for spatial learning, in an RNN+GNN pipeline setting (RGNN in short). Specifically, for power system fault diagnosis, Graph Convolutional Networks (GCNs) have been adopted. Yet, various more advanced GNN architectures have been proposed and adopted in domains outside of power systems. In this paper, we set out to systematically and consistently benchmark various GNN architectures in an RNN+GNN pipeline model. Specifically, to the best of our knowledge, we are the first to (i) propose to use GraphSAGE and Graph Attention (GAT, GATv2) in an RGNN for fault diagnosis, and (ii) provide a comprehensive benchmark against earlier proposed RGNN solutions (RGCN) as well as pure RNN models (especially Gated Recurrent Unit (GRU)), particularly (iii) exploring their generalization potential for deployment in different settings than those used for training them. Our experimental results on the IEEE 123-node distribution network show that RGATv2 has superior generalization capabilities, maintaining high performance with an F1-score reduction of $\sim$12% across different topology settings. In contrast, pure RNN models largely fail, experiencing an F1-score reduction of up to $\sim$60%, while other RGNN variants also exhibit significant performance degradation, i.e., up to $\sim$25% lower F1-scores.

Figures (3)

• Figure 1: Illustration of the (a) per-monitorshared RNN for local inference per PMU, (b) aggregated RNN model, and (c) RNN+GNN pipeline architecture. The GNN layers consists of message-passing, dropout, and batch normalization, with each variant using its respective architecture (see \ref{['subsec:sfe_gnn']}). Further, $N$ represents the number of monitorPMU nodes, $F$ is the number of features (e.g., phase voltages, currents), $S$ is the time series sequence length, $H$ denotes the dimension of the GRU hidden state representation (and hence output), and $H'$ is the dimension of the GNN layer's output node representations.
• Figure 2: IEEE 123-node feeder with fault locations and voltage measurements.
• Figure 3: F1 Scores for Fault Event Detection using RNN-only models (GRU-based; left) and GNN-based models (right). All models are trained on the 11-monitorPMU setup and evaluated across all configurations (7--25 monitorsPMUs; see \ref{['table:dataset_subsets']} for details). Error bars show the 90% confidence intervals across 5 models trained with different random seeds.