Accurate pedestrian flow prediction via a graph- and sequence-based spatiotemporal mixture-of-experts for assessing collective dynamics features of crowd crush risk in subway stations

Accurate prediction of pedestrian flow at critical subway nodes (e.g. gates, escalators, staircases) prone to bottlenecks is fundamental to assessing collective dynamics (e.g. crowd crush risk). However, large peak and off-peak contrasts, spatial interdependencies, and heterogeneous node capacities undermine the predictive performance of single-expert models. To address such complexity, we propose a hybrid sequence- and graph-based Mixture of Experts approach, ResGAT-BX, comprising i) a Graph Attention Network to model spatial dependencies and ii) Bidirectional GRU and extended long short-term memory experts to capture temporal variability, thereby improving adaptability across conditions. Using ResGAT-BX predictions, we quantify node-level risk via crowd-gathering density, duration, and throughput fluctuation amplitude. Experiments on a real-world Qingdao subway dataset demonstrate i) ResGAT-BX’s robustness via parameter-sensitivity analyses and ii) ResGAT-BX’s consistently superior performance over state-of-the-art baselines for both direct flow volume prediction and derived crowd crush risk anticipation across multiple metrics and regimes, including peak surges and off-peak fluctuations. The dataset used in this study has been publicly released on GitHub: https://github.com/niupeng111/Node-Pedestrian-Flow-Dataset .