Toward Explainable GNN-Based Dynamic Stability Assessment via Joint Feature Attribution

Graph Neural Networks (GNNs) offer a fast approach to dynamic stability assessment of power grids, and edge-conditioned architectures are especially well suited because they incorporate transmission-line parameters directly into message passing. For deployment, however, power grid operators need explanations that account for all inputs influencing a prediction: bus states, line parameters, and global operating context. Existing post-hoc GNN explanation methods are not designed for this setting: they focus mainly on node- and graph-level classification tasks, provide limited treatment of edge features, and often ignore graph-level features. We propose a joint attribution framework for GNN predictions that decomposes a prediction into node, edge, and graph-level contributions while preserving completeness, establishing a first formal guarantee for trustworthy joint explanations. In this work, we focus on edge-level regression for dynamic stability assessment.