Graph Neural Networks for Transmission Grid Topology Control: Busbar Information Asymmetry and Heterogeneous Representations

TL;DR

The paper tackles grid congestion by topology control and investigates how graph representations affect GNN performance. It identifies busbar information asymmetry in homogeneous graph representations and introduces a heterogeneous graph model (HetGNN) that captures same-busbar, other-busbar, and line-end connections. Empirical results show HetGNN outperforms both HomGNN and FCNN on in-distribution data and generalizes more effectively to out-of-distribution grid states, while simulations indicate operation performance comparable to expert policies with substantially faster inference than humans. The work demonstrates the practical potential of graph-aware topology control for real-time transmission-system operation and outlines future directions for integrating HetGNN into more advanced learning frameworks like MARL and DAgger-based methods.

Abstract

Factors such as the proliferation of renewable energy and electrification contribute to grid congestion as a pressing problem. Topology control is an appealing method for relieving congestion, but traditional approaches for topology discovery have proven too slow for practical application. Recent research has focused on machine learning (ML) as an efficient alternative. Graph neural networks (GNNs) are particularly well-suited for topology control applications due to their ability to model the graph structure of power grids. This study investigates the effect of the graph representation on GNN effectiveness for topology control. We identify the busbar information asymmetry problem inherent to the popular homogeneous graph representation. We propose a heterogeneous graph representation that resolves this problem. We apply GNNs with both representations and a fully connected neural network (FCNN) baseline on an imitation learning task. The models are evaluated by classification accuracy and grid operation ability. We find that heterogeneous GNNs perform best on in-distribution network configurations, followed by FCNNs, and lastly, homogeneous GNNs. We also find that both GNN types generalize better to out-of-distribution network configurations than FCNNs.

Figures (9)

• Figure 1: The default state of Grid2Op environment rte_case14_realistic. Numbers indicate power line or substation IDs.
• Figure 2: a: An example grid consisting of two substations and seven objects. Note that this grid is unrealistic, as real power grids should not be split. b: The homogeneous graph representation of that grid. c: The heterogeneous graph representation of that grid, where edge colors indicate edge types. d: Message passing towards node 5 in the homogeneous model. The purple reflective edge is added to reflect self-weights. e: Message passing towards node 5 in the heterogeneous model, where edge colors indicate message types.
• Figure 3: The training curves of the five models per model type.
• Figure 4: The test accuracies of the model types on the ID dataset set (left) and OOD dataset (right). The ranges indicate the maximum and minimum accuracy.
• Figure 5: The log-distributions of the classes in the ID validation set. Overlaid is the frequency by which the model predicts that class (left: HomGNN, right: HetGNN). The non-overlapping blue areas at the tails of the distributions indicate that the models predict rare classes disproportionally infrequently. This finding is consistent among the models.