Augmented Pre-trained Graph Neural Networks for Grid-Supportive Flexibility Control

With the ongoing transformation in electricity grids, new components such as decentralized generation systems and high-load electric vehicle charging infrastructures are installed.To foresee congestion scenarios and lay out flexibility-providing distributed energy resources, power flow calculations are generally required.Since power flow calculations are computationally expensive, various neural network architectures have been developed in the literature to infer power flow estimates.In the present paper, we extend the application of a graph neural network (GNN) architecture initially developed for power flow approximation to infer minimum load or generation requirements for congestion mitigation.Evaluating different training procedures and models, we find that a two-step training process using a pre-trained power flow inferring GNN has a head start in training compared to end-to-end training of the GNN.The pre-trained model demonstrates robust generalization capabilities, effectively inferring optimal power values for congestion mitigation across grid topologies outside of the training distribution.Furthermore, as the second training step is not limited to the task of congestion mitigation, the approach can easily be adapted to other use cases.