Modeling spatio-temporal coupling and emergency-induced perturbations in urban traffic networks: A multi-scale dynamical approach

Urban traffic networks exhibit complex spatio-temporal co-evolutionary characteristics and non-linear dynamic dependencies. While existing graph neural network (GNN) models achieve significant fitting precision, the asymmetric physical linkage between upstream and downstream flows is often overlooked, thereby limiting the robustness of traffic flow prediction under stochastic perturbations. In this study, the synergistic evolution of traffic states is investigated through a multi-scale dynamical framework (MuPD).A hierarchical structure is proposed to model the long-range dependence of traffic propagation, incorporating a dynamic weight allocation strategy to quantify the directional coupling strength between network nodes, and an adaptive sensing mechanism to dynamically regulate the observational scop. Specifically, a self-supervised mechanism is introduced to monitor system fluctuations. Rigorous theoretical proofs are provided to demonstrate that the proposed mechanism effectively encodes the Granger causality of traffic changes and maintains convergence stability under the constraints of the network’s topological eigenvalues.The effectiveness of the proposed framework is validated across large-scale urban road networks under both steady-state and volatile regimes. Extensive experiments indicate that MuPD consistently yields superior performance. Crucially, under emergency-induced perturbations, MuPD exhibits superior robustness compared to competitive spatio-temporal models like STWave and DASTNet, significantly reducing MAE by 11.61% to 40.26%, RMSE by 11.30% to 26.71%, and MAPE by 11.62% to 58.67%. Furthermore, sensitivity analysis confirms the systemic stability even under severe data sparsity, underscoring the ability to capture the underlying causal structure of urban infrastructures. This research provides new physical insights into the collective dynamics of intelligent transportation and the enhancement of complex network resilience.