A hybrid machine learning and microsimulation framework for simulating metro station evacuation

Mass Rapid Transit is a globally efficient transport mode, yet its stations host large-scale pedestrian activity, leaving them vulnerable to hazards that can trigger evacuations. This study develops a pedestrian evacuation microsimulation framework capturing behavior across varying panic levels during metro station evacuations. Latin Hypercube Sampling and Random Forest modeling generate pedestrian behavior parameter combinations for calibrating the model and simulating panic conditions. Calibration uses a Genetic Algorithm, with validation via Geoffrey E. Havers statistics from CCTV data. Results show average speed fluctuates under low panic but remains stable under medium and high panic. Under low panic, most pedestrians evacuate in the later half; medium panic spreads evacuations evenly; high panic peaks in the first and last thirds. At peak hour (540 people), clearance times were 5.25–6.47 min, 5.12–5.25 min, and 4.90–5.07 min for low, medium, and high panic, respectively. These findings support countermeasures for safe, efficient MRT evacuations.