Urban Flood Emergency Evacuation Path Optimization Research

Abstract

Existing methods for optimizing urban flood emergency evacuation routes often fail to account for the dynamic evolution of inundation and the heterogeneity of evacuees' ages. To address this gap, this paper proposes a personalized evacuation route optimization approach that integrates Agent-Based Modeling (ABM) with Deep Q-Network (DQN) reinforcement learning. The proposed method incorporates water-depth-dependent speed reduction models and instability thresholds specific to children, adults, and the elderly into both the state space and reward function, enabling DQN to iteratively learn optimal routes that balance safety and efficiency. Using the "7·20" extreme rainstorm in Zhengzhou as a case study, we constructed a 100-year return-period pluvial flooding scenario and conducted evacuation simulations under three activation timings-0, 1, and 2 hours after rainfall onset-comparing the proposed method against the traditional A* algorithm. Results demonstrate that the experimental group reduced average evacuation time by 17.8%-42.1% relative to the control group, with particularly pronounced improvements for children and the elderly (up to 45.1%). The safety metric F-value in the experimental group exceeded that of the control group by more than 0.3 under deep-water conditions and exhibited a more gradual decline. These findings confirm that the DQN-based approach, which integrates age heterogeneity and dynamic water depth, substantially enhances both the timeliness and safety of emergency evacuation during urban floods, offering more reliable route planning for high-risk populations and providing a scientific basis for optimizing urban flood emergency response protocols.