A knowledge-data dual-driven framework for intelligent flood evacuation in subway stations

Traditional building evacuation planning often lacks real-time adaptability in subway floods due to over-reliance on simulations, which bring complex physics and calibration challenges, leading to delays in risk assessment and jeopardizing occupant safety. To address this challenge, this study proposes a novel knowledge-data dual-driven framework for intelligent flood evacuation management in subway stations. The framework integrates rapid data-driven prediction enhanced by simulation-derived knowledge, and fast optimization guided by knowledge-based risk assessment within a decision-support system, aiming to improve real-time responsiveness and occupant safety. The novel components in this framework include a red-billed blue magpie-optimized deep learning model for evacuation time and density prediction with SHAP interpretability, a cloud-based fuzzy evaluation system for flood risk quantification, a multi-objective path optimizer balancing evacuation time and slip-fall risks, and a convolutional genetic algorithm for efficient solution generation. A real subway station case study is conducted by using Fluent and PathFinder to validate the proposed method, demonstrating that: (1) The prediction model achieves a 6.44% improvement over traditional TCN-GRU methods. (2) The cloud model-based weighting method effectively quantifies safety risks, providing data support for emergency decisions. (3) The path optimization method reduces evacuation time by 46.44 s and peak crowd density by 0.8477 p/m², outperforming conventional methods by over 15.4%. These advancements position the framework as a transformative decision-support tool for intelligent building operations in underground structures, directly contributing to sustainable and safe built environments.