Assessing Explanations of Graph Neural Networks for Dynamic Stability Assessment of Power Grids

As Graph Neural Networks (GNNs) are increasingly deployed for power grid stability predictions, understanding whether their learned behavior aligns with physical grid dynamics becomes essential for safe deployment. We present a systematic analysis addressing two questions: do post-hoc explanations accurately reflect GNN model behavior, and does that behavior align with the physics of grid dynamics? We apply gradient-based methods (five variants including Saliency, InputXGradient, and Integrated Gradients) and a game-theory-based method (Shapley Value Sampling) to a Dirac–Bianconi Graph Neural Network (DBGNN) achieving a skill score of 0.903 on a fault-ride-through (FRT) probability prediction task. Our analysis reveals that the model primarily relies on node type-level information, learning to distinguish inverters (and subtypes via the Normal Form (NF) parameter) from loads, rather than continuous power flow and network features within node types. Nevertheless, it incorporates meaningful topological context: neighborhood composition modulates predictions in physically intuitive ways, with influence decaying with graph distance. We validate the explanation methods on the IEEE39-AC dataset with known ground truth. Beyond explanation, we present a proof-of-concept counterfactual analysis: edge removals improve predicted stability by 45% on average at the most critical nodes while preserving global grid stability, reducing the search space for physics-based validation.