Research on the Measurement and Prevention Method of Railway Network Delay Risk Propagation Based on Time-Indexed Graph Attention Network and Cascade Failure Dynamics

To characterize delay propagation and dispatch feedback under sudden disturbances in railway networks, this study formulates a risk measurement and control framework integrating time-indexed graph attention (TI-GAT), cascade-failure dynamics, and deep reinforcement learning. The TI-GAT block is evaluated independently on each 5 min network state and contains no temporal convolution, recurrent unit, or temporal self-attention. Its state-dependent edge weights parameterize the CFD transition, while temporal evolution is carried by the recursively updated load and delay-risk states. At each decision interval, the continuous DDPG action is projected onto an admissible dispatch set defined by block occupancy, minimum headway, station-track availability and capacity, section speed authorization, and route-conflict interlocking. The projected detention, speed, and routing variables then modify node release, section impedance, route allocation, and the subsequent CFD transition, whose updated load and risk states form the next policy observation. Prediction performance is evaluated against held-out chronological operational records, with terminal station–time targets derived from observed actual-versus-scheduled arrival/departure deviations. Capacity-sensitivity and scheduling-control outcomes are generated within the action-conditioned TI-GAT–CFD simulation environment. Within a 120 min prediction window, the model’s accuracy and F1 score are 0.864 ± 0.005 and 0.831 ± 0.007, respectively; in a 60 min ablation experiment, the MAE and false negative rate are 11.3 ± 0.9 min and 6.2 ± 0.3%, respectively. In the simulation-based scheduling experiment, the global delay convergence time is 82.3 ± 2.3 min, the total delay attenuation rate is 65.8 ± 1.3%, and the post-training inference time for a 100-train disturbance scenario is 1.83 ± 0.06 s. The predictive metrics therefore quantify agreement with held-out observations, whereas the control metrics characterize closed-loop in silico behavior under the modeled dynamics and feasibility constraints.