Optimising pre-disaster evacuation strategies using shared autonomous vehicles: A dynamic and stochastic approach

• Defines a dynamic SAV Evacuation Dispatching Problem for car-less residents. • Introduces TB-DRL combined with Branch-and-Bound strategy for real-time SAV routing. • Cuts total evacuation cost by up to 27.45 % while keeping route gap < 1.7 %. • Validated on Auckland data, showing strong real-world applicability and robustness. Classical mass evacuation strategies from disasters often rely on static planning and manual operations, which struggle to handle the uncertainty and time-critical nature of large-scale disasters. This study aims to develop a dynamic and intelligent evacuation strategy using Shared Autonomous Vehicles (SAVs) to improve the efficiency and equity of emergency transport for populations without private vehicles. We propose a Transformer-Based Deep Reinforcement Learning with Branch and Bound (TB-DRL) approach to address the SAV Evacuation Dispatching Problem under uncertain demand, traffic, and SAVs availability. The TB-DRL’s multi-head attention layers learn spatiotemporal correlations between demand nodes and SAV locations, while the hierarchical mask progressively prunes infeasible or low-value actions, shortening decision time without compromising optimality. Based on Auckland’s 345 km road graph and 38,993 residents dataset, the results show that TB-DRL lowers total evacuation cost by 8.53 %-27.45 % in dynamic tests and by up to 18.55 % and 16.19 % in static and dynamic city-scale scenarios, while keeping the mean route-cost gap below 1.7 %. Pareto analysis demonstrates that TB-DRL can shift smoothly between cost-driven and demand-driven objectives, while sensitivity experiments across assembly node locations, random SAV’s state and congestion levels confirm the robustness of the strategy to uncertain evacuation scenarios. The simulation based on real population data shows that using more SAVs helps reduce congestion in the later evacuation stage. However, their empty return trips can create local congestion. When SAVs handle 60 % of the evacuation tasks, the total evacuation time and efficiency improve most significantly.