Time prediction of human evacuation from passenger ships based on machine learning methods
File(s)
Published version
Author(s)
Type
Journal Article
Abstract
During an emergency evacuation scenario, accurately and timely predicting human evacuation time beforehand is crucial for developing an efficient evacuation plan. This study aims to develop an innovative simulation-based framework in which series state-of-the-art Machine Learning (ML) models are applied to predict human evacuation time from passenger ships. It also develops a multi-dimensional decision-making approach to evaluate their performance from the perspectives of high prediction accuracy and timeliness to support rapid response during emergencies. Firstly, an agent-based modelling technique incorporating two objectives and seven influential factors specific to human evacuation scenarios onboard ships is used to simulate the evacuation process. Then, the evacuation model is validated using three indicators to ensure its accuracy and relevance. Secondly, nine state-of-the-art ML models are applied to predict and analyse human evacuation time. To further investigate the role of feature interactions and enhance predictive accuracy, an additional model called the Attention-enhanced Light Gradient Boosting Machine (Attention-LightGBM) is proposed. Additionally, four statistical indicators are utilised to monitor the performance of each model. Finally, a new weighted selection method based on analytic hierarchy process and entropy weight method is created to conduct a comprehensive assessment from the perspectives of accuracy and timeliness. The findings reveal that the Attention-LightGBM demonstrates significant advantages in prediction accuracy, while the LightGBM excels in prediction timeliness. This study not only provides theoretical and technical support for emergency management onboard ships but also suggests methodological advancements for future research on complex human evacuation scenarios from passenger ships. The source code is publicly available at: https://github.com/AdvMarTech/Eva_Predict_ML.
Date Issued
2025-10-01
Date Acceptance
2025-08-28
Citation
Transportation Research Part A: Policy and Practice, 2025, 200
ISSN
0191-2607
Publisher
Elsevier
Journal / Book Title
Transportation Research Part A: Policy and Practice
Volume
200
Copyright Statement
© 2025 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
License URL
Publication Status
Published
Article Number
104664
Date Publish Online
2025-09-12