Fast and accurate node-level hosting capacity estimation in distribution networks using graph neural networks

This paper presents a fast and accurate method for estimating node-level hosting capacity in distribution networks using Graph Neural Networks (GNNs). A large open dataset of synthetic networks, representative of Italian distribution grids, was created, and hosting capacity values were computed using AC optimal power flow (AC-OPF) simulations, which are computationally expensive. This dataset was used to train a GNN regressor based on the Graph Isomorphism Network architecture, integrating path-based electrical features to capture both local and global grid dependencies. The proposed GNN is compared to a fast physics-based estimation method, showing a reduction in mean absolute error of about 30%, highlighting its ability to capture the spatial and electrical relationships embedded in the network topology and to produce physically consistent estimates. In terms of efficiency, the model achieves an inference speed over 11,000 times faster than AC-OPF, enabling hosting capacity estimation in milliseconds. This significant speed-up allows real-time applications such as interactive hosting capacity maps, large-scale scenario analysis, and decision support for distribution system operators. The model’s robustness is also evaluated under out-of-distribution conditions, where some performance degradation is observed, particularly for small networks, emphasizing the importance of representative training data. Importantly, the GNN remains operational even when traditional power flow methods fail to converge, providing stable and reliable estimates without requiring manual intervention. Overall, the approach demonstrates strong potential for scalable, real-time hosting capacity assessment to support the integration of distributed energy resources.