Towards adoption of GNNs for power flow applications in distribution systems

An essential component of smart grid applications is the ability to solve the power flow (PF) problem in real-time. As numerical methods are too slow, the use of neural networks (NNs) is rapidly increasing. Graph Neural Networks (GNNs) and their variants have become one of the leading methods to learn graph representations. Power systems and in particular, distribution systems can be represented as graphs, and are characterized by often topology changes, which makes the consideration of the topology structure to be an important aspect when searching for a solution approach. Although GNNs have promising results for certain applications such as computer vision ones, considering its limitations, it still has a long way to go until becoming a leading candidate for PF based applications. This paper highlights the existing gaps and challenges in fully accepting ANNs and particularly GNNs as real-time solution engines for the PF problem in DSs. These gaps are analyzed under three categories: suitable architectures for the solution of the PF problem in DS, explicit vs. implicit incorporation of the DS topology information impact on the models’ generalization, and the limiting factors for GNNs implementation aimed at the solution of the PF problem in DSs. The paper also includes a discussion, suggestions and insights of overcoming these gaps in future research.