Graph neural networks for transmission grid topology control: Busbar information asymmetry and heterogeneous representations

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