Graph Neural Network Power Flow Solver for Dynamical Electrical Networks

Abstract

This work presents a novel graph neural network (GNN) based power flow solver that focuses on electrical grids examined as dynamical networks. The proposed method employs a specific type of GNN that considers different types of nodes to allow the direct translation to the generator and load buses, and the branches present in power systems. The GNN model is trained with modified versions of IEEE test cases by permuting the configuration of the electrical grid for the distinct samples used during training. The training is carried out in an unsupervised manner, and the results are compared with a conventional and trustworthy Newton-Raphson based method. The presented results are shown to be accurate, and perform acceptably well even when tested on grids of different size from the ones they observed during training.