From simulation to safety strategy for unplanned mass gatherings: Agent-based evacuation optimization with lessons from Wuhan’s mega-event

Unlike organized events, spontaneous crowd gatherings are more likely to cause local congestion and even stampede-related disasters due to the lack of effective guidance and control. Currently, crowd control at such events primarily relies on empirical flow restrictions and on-site guidance, but these strategies lack quantitative analysis and optimization for large-scale crowds. As a result, their effectiveness is difficult to evaluate under complex gathering scenarios, and their underlying mechanisms remain insufficiently explained. To this end, a large-scale evacuation simulation model was constructed based on the AnyLogic platform to quantitatively investigate the synergistic effects of delay and guidance strategies, as well as to evaluate the effectiveness and interdependencies of various safety strategies. Using XGBoost and SHAP methods, this study quantified the impact and contributions of pedestrian characteristics, guidance deployment, and spatial configuration across three safety strategies: no guidance, guidance only, and delay-guidance coordination. The NSGA-II algorithm was then used to optimize collaborative delay-guidance strategies for different conditions. The simulation results show that: (1) The coordination of delay and guidance strategies significantly alleviates congestion and improves evacuation efficiency. Compared with guidance-only strategies, overcrowding is reduced by 19.3 % and evacuation time by 9.0 %. However, with a higher walking speed and sufficient guide resources, excessive delays (e.g., 120 s) may reduce efficiency. (2) With increased crowd size, increasing dynamic guidance combined with static deployment at key nodes can significantly enhance safety and conserve resources. (3) Higher guidance acceptance notably improves evacuation efficiency. This study provides simulation-based decision support for large-scale event evacuation strategies.