Optimization of bus-subway network resilience based on dynamic passenger flow prediction and potential station failure identification

To address the risk of cascading failures caused by large passenger flows in subways, this study proposes a resilience optimization method that integrates dynamic passenger flow prediction with the identification of potential failure sites. First, this paper constructs a multi-layer bus-subway network model based on the Space L method and designs a Bayesian-optimized graph convolutional gated recurrent unit (BO-GCGRU). This model utilizes self-attention dynamic adjacency matrices to capture the spatiotemporal evolution patterns of passenger flows, achieving high-precision short-term predictions of passenger flows at subway stations (e.g., Xi'an Zhonglou Station) with an mean absolute percentage error (MAPE) of only 4.91%. Second, this paper combines passenger flow prediction with station capacity to identify potential fault nodes (e.g., Longshouyuan Station with a load rate of 93%) and establishes a cascading failure simulation model that integrates dynamic congestion penalties and random utility theory. Simulation results indicate that adjusting the bus mode share to 30%–40% can effectively mitigate the potential risk of failure at Longshouyuan Station, keeping the overall network performance loss rate at approximately 10%. This method provides theoretical basis and decision support for enhancing the resilience of urban public transportation networks.