A Resilience-aware Spatio-temporal Graph Neural Network for Adaptive Routing in Disrupted Road Networks

Unplanned disturbances such as collisions, weather events and infrastructure failures routinely compromise the continuity of urban mobility.Traditional route guidance systems are ill-equipped to cope with cascading disruptions that propagate through the network's intertwined topology.Recent advances in Graph Neural Networks (GNNs) make it possible to reason over relational data and infer hidden dependencies, yet most studies focus solely on prediction and overlook the resilience of the road network itself.In this work we introduce a resilience-aware spatio-temporal graph neural framework that models both the structural vulnerability and the temporal dynamics of traffic flows.Each intersection and road segment is represented as a node and edge in a dynamic graph augmented with real-time sensor observations and historical incident information, with incidents inferred through sustained, abnormal deviations in observed traffic conditions rather than relying on explicit annotations.A multi-layer spatio-temporal GNN learns rich embeddings of each component and a novel node-edge influence function quantifies the extent to which localized disturbances propagate through neighbouring components, producing normalized resilience scores with a consistent physical interpretation across time and network scale.These resilience scores drive a modified Dijkstra-like routing algorithm that dynamically reassigns flow away from vulnerable segments, enabling rapid recovery during disruptions under realistic update latency and partial observability constraints.Experiments on real-world (METR-LA) and largescale simulated (CityFlow+OpenStreetMap) datasets demonstrate that our framework reduces the Mean Resilience Score Deviation by 36 %, shortens Traffic Flow Recovery Time by 40 %, and increases Emergency Rerouting Efficiency by nearly 20 % relative to baseline GNN and rule-based methods, with performance reported as averaged outcomes over multiple randomized runs.A comparative analysis against recent large-scale STGNN architectures reveals that the proposed method retains competitive predictive accuracy while introducing a resilience dimension absent in prior work.These results underscore the potential of resilience-guided spatio-temporal GNNs for intelligent transportation systems and motivate further integration with reinforcement-learning strategies for proactive routing.