An efficient evacuation time prediction model for different types of subway stations

Passenger evacuation time prediction is a crucial basis for subway station management to better formulate safety control measures. It becomes possible to reasonably plan the flow of passengers within the station and configure safety devices such as signage and sprinkler systems with a known and explicit time framework, thereby reducing passenger congestion and panic. However, methods based on crowd dynamics simulation require a significant amount of time and effort to build models, and the rapidity of time prediction is challenging to ensure. Real human evacuation experiments involve ethical, safety and practical operational issues. To address this challenge, an evacuation time prediction model for subway passengers is established based on the CPA-SVR machine learning method, enhancing the speed and accuracy of prediction. The reliability of simulation results is validated by comparing observed values of passenger alighting and boarding time and traffic time at stairs with simulation values from MassMotion software. Fourteen factors related to the subway station structure, passengers and train status are selected as influence factors for evacuation time. A foundation data set for the evacuation time prediction model is obtained through 179 evacuation experiments under different influence factors using the MassMotion simulation system at 32 constructed stations. The SHAP interpretation method is applied to identify the importance of influence factors in the experimental results. A CPA-SVR passenger evacuation time prediction model is established, with accuracy concentrated between 85%–100%, based on training and validation sets. Further testing with 45 additional sets of fresh experimental data demonstrates the model’s strong predictive capability for new data, indicating good generalization ability. Finally, a focused analysis of passenger evacuation behaviors at bottlenecks such as stairs, gates and exits is conducted, accompanied by relevant optimization strategies.