Data-driven evacuation time prediction for train passengers under floods: A differentiated deep learning framework

Predicting the evacuation time of subway train passengers is a key issue in flood disaster prevention and emergency management. However, traditional experiments are time-consuming and costly, failing to meet real-time requirements. To address this issue, this paper proposes a differentiated deep learning framework named DCBAK, which integrates an advanced differentiated creative search (DCS), convolutional neural networks, bidirectional long short-term memory networks, attention mechanisms, and kernel density estimation. The designed DCS is innovatively applied to optimize hyperparameters in the integrated network, and an enhanced fitness calculation method is proposed to improve the quality of time prediction. Kernel density estimation synthesizes the contributions and distribution characteristics of point prediction results, enhancing the smoothness and credibility of evacuation time prediction results. The introduction of SHAP structure increases the transparency and interpretability of prediction results. Data on train passenger evacuation collected by PathFinder are used to demonstrate the effectiveness of the proposed method. The results show that: (1) The proposed DCBAK approach achieves a high coefficient of determination above 0.90 and even up to 0.99, indicating a high degree of data fitting; (2) Under a 95% confidence interval, the prediction interval coverage probability of the evacuation time reaches 0.95, accurately explaining the uncertainty range of results; (3) The most significant factor affecting the evacuation time of train passengers in flood disasters is the train stopping scenario. This study emphasizes the potential of differentiated deep learning networks in addressing the challenge of rapid evacuation time prediction in emergency evacuation management.