A Framework for Assessing the Generalizability of GNN-Based AC Power Flow Models

AC power flow analysis is essential for grid planning, operation, and optimization, but traditional methods like the Newton-Raphson algorithm can be computationally expensive for complicated systems.Simplified approaches, such as DC power flow, perform well in transmission grids but struggle in distribution grids due to high resistance-to-reactance (R/X) ratios that affect voltage stability and power losses.Recent advances in machine learning, particularly Graph Neural Networks (GNNs), offer a flexible, topology-aware alternative for power flow solvers, but existing models lack generalizability and fail to incorporate critical topological features.This work presents the first framework for assessing the generalization performance of these graph-based solvers, including a generalization score to evaluate this model robustness.By analyzing graph similarity statistics and their correlation with model performance, we explore the generalization impact of embedding simple graph characteristics into GNN models for distribution grids, and ultimately demonstrate that neighbourhood-aware GNNs produce more generalizable solutions.These findings address key challenges in power flow analysis, advancing the applicability of GNNs to complex distribution networks.