Benchmarking Graph Neural Networks for Power Flow Prediction in Distribution Systems

The search for methods able to control large numbers of distributed energy resources (DERs) without compromising grid security has led to the consideration of machine learning (ML) methods for predicting power flow (PF). Previous research has shown positive potential of this approach, but ML methods that are reliable and fast at the same time remain to be found. This study contributes to this stream of research by evaluating the performance of graph neural networks (GNNs) for power flow prediction in distribution systems. In particular, we benchmark different GNN architectures like graph convolutional networks (GCNs) and graph attention networks (GATs) as well as variants thereof with multi-layer perceptrons (MLPs) as baseline. We restrict ourselves to purely data-driven variations of GNNs in an inductive setting to compare the predictive performance and computational efficiency of these different neural networks. Our results show that none of the evaluated GNNs surpasses MLPs due to the restrictions of message passing. The results contribute to the understanding of how different GNN architectures perform in PF prediction and why. This could inform further research on GNNs applied to complex power system control tasks, eventually helping to efficiently integrate DERs into distribution grids.