Prediction of hazardous transmission lines after power grid cascading failures using the node and edge attributed graph edge-attention residual network

With the increasing integration of renewable energy and complex hybrid AC/DC grid topologies, power systems face heightened cascading failure risks under N-k contingencies. Following such failures, topological reconfiguration and load adjustments can cause some transmission lines to enter a “subcritical overload” state,where power flow exceeds stability limits without triggering protection leading to latent faults and secondary collapse risks. To tackle this, this paper proposes a Node and Edge Attributed Graph Edge-Attention Residual Network (NEA-GEAT-Res) for predicting potentially overloaded lines. Using IEEE test cases with random and clustered faults, we simulate cascading failures via an AC-Cascading Failure Model (AC-CFM). Post-failure, lines are classified as failed, normal, or hazardous. Based on the NEA-GNN framework, our model introduces cross-layer residual connections to preserve initial features and mitigate over-smoothing, alongside an edge attention mechanism that dynamically weights critical line information. Experiments show that NEA-GEAT-Res achieves F1-scores of 96.42 % (IEEE 39-bus) and 84.59 % (IEEE 118-bus), improving over baseline NEA-GNN by 16.79 % and 14.87 %, and outperforming other mainstream models. Notably, adding topological features benefits baseline models (3 %–11 % F1 improvement) but not NEA-GEAT-Res, indicating our model effectively captures dynamic grid characteristics through residual and attention mechanisms. This work reveals GNN feature sensitivity in hazardous line prediction and suggests hybrid feature modeling avenues, providing a high-accuracy solution for proactive defense after cascading failures.