Graph Neural Networks for Hosting Capacity Assessment in Distribution Networks

Marco Rossi – Ricerca su Sistema Energetico

The presentation introduces a fast and accurate method for estimating node-level hosting capacity in distribution networks using Graph Neural Networks (GNNs). A large dataset of synthetic networks, reflecting typical Italian distribution grids, was generated, and hosting capacity was calculated for each node using deterministic, computationally intensive techniques. This dataset enabled the training of a GNN regressor based on the Graph Isomorphism Network architecture, which incorporates path-based electrical descriptors to capture both local and global grid dependencies. The GNN’s performance was compared to a novel linearized power flow method, showing that the GNN achieves lower prediction errors and much greater computational efficiency. Inference times are reduced by several orders of magnitude compared to traditional AC-OPF simulations, enabling real-time and scalable hosting capacity assessments. These results demonstrate the practical value of GNNs for Distribution System Operators, supporting rapid and reliable integration of distributed energy resources.