Real-time prediction of tunnel evacuation time using machine-learning surrogate models

Real-time prediction of tunnel evacuation time using machine-learning surrogate models

Článek WoS

While advanced microscopic models provide detailed behavioural representation, their computational cost remains a significant barrier to real-time operational use. This study explores the use of machine-learning surrogate models for real-time estimation of tunnel required safe egress time. A synthetic dataset of 50,910 evacuation simulations was created using procedurally generated simulations with variable tunnel geometry and pseudo-random sampling using the Pathfinder evacuation simulator. Two complementary surrogate model classes were investigated: a tree-based extreme gradient boosting ensemble and a multilayer perceptron. Models were trained using nested cross-validation and validated against 10 full-scale tunnel evacuation experiments conducted in three countries. The extreme gradient boosting model demonstrated stable generalisation performance, low prediction error and consistent conservative bias across all validation datasets (R2: 0.99, MAE: 9.2 s). The multilayer perceptron showed reduced robustness under experimental conditions (R2: 0.91, MAE: 32.1 s). Results indicate that tree-based surrogate models are able to mimic required safe egress time predictions performed with more computationally expensive agent-based simulators. The proposed approach enables rapid evacuation-time assessment suitable for operational tunnel emergency decision support.

While advanced microscopic models provide detailed behavioural representation, their computational cost remains a significant barrier to real-time operational use. This study explores the use of machine-learning surrogate models for real-time estimation of tunnel required safe egress time. A synthetic dataset of 50,910 evacuation simulations was created using procedurally generated simulations with variable tunnel geometry and pseudo-random sampling using the Pathfinder evacuation simulator. Two complementary surrogate model classes were investigated: a tree-based extreme gradient boosting ensemble and a multilayer perceptron. Models were trained using nested cross-validation and validated against 10 full-scale tunnel evacuation experiments conducted in three countries. The extreme gradient boosting model demonstrated stable generalisation performance, low prediction error and consistent conservative bias across all validation datasets (R2: 0.99, MAE: 9.2 s). The multilayer perceptron showed reduced robustness under experimental conditions (R2: 0.91, MAE: 32.1 s). Results indicate that tree-based surrogate models are able to mimic required safe egress time predictions performed with more computationally expensive agent-based simulators. The proposed approach enables rapid evacuation-time assessment suitable for operational tunnel emergency decision support.

Evacuation, Road tunnels, Real-time assessment, Agent-based models, Machine learning

Evacuation, Road tunnels, Real-time assessment, Agent-based models, Machine learning

UHLÍK, O.; GONZALES-VILLA, J.; CUESTA, A.; RONCHI, E.; JUŘÍK, V.; JURÁNKOVÁ, R.; APELTAUER, T.; APELTAUER, J.

30.06.2026

Elsevier

Safety science

203

červen

Nizozemsko

12

https://www.sciencedirect.com/science/article/pii/S0925753526002390?dgcid=author