How to improve traffic safety at freeway bottleneck with connected and automated vehicles? A dynamic deceleration strategy based on deep reinforcement learning

This study proposes a dynamic deceleration control strategy to improve traffic safety at freeway bottlenecks under mixed traffic with connected and automated vehicles (CAVs). A dynamic deceleration zone is implemented upstream of the bottleneck, where CAVs act as mobile actuators to induce smooth speed reduction among surrounding human-driven vehicles. A deep reinforcement learning (DRL) framework, incorporating the soft actor–critic and deep deterministic policy gradient algorithms, is designed to optimize the target speeds of CAVs within the zone. Simulation results show that the proposed strategy improves safety metrics by 30%–70% compared with uncontrolled conditions and performs better than conventional deceleration strategies in both single-lane and multi-lane scenarios. Sensitivity analysis further indicates that the optimal deceleration-zone length is 200–400 m. The findings demonstrate the potential of integrating CAVs with DRL-based control to address the limitations of traditional variable speed limit approaches at freeway bottlenecks.