A Safe Reinforcement Learning Based Predictive Position Security Control in a Mixed Ramp Confluence Scene

Safety and efficiency are the unremitting pursuit of road traffic, and the research goal of this article is to improve operational efficiency as much as possible while ensuring safety in a Mixed Ramp Confluence Scene. In the scenario of ramp merging, a large number of methods have been proposed to better ensure safety and efficiency, but the control results in mixed traffic flow are not satisfactory. We propose a Predictive Position Security Control (PPSC): by integrating the concept of safe reinforcement learning, we limit the dangerous situations in the scene of ramp convergence and improve driving safety. In addition, by improving the first-in-first-out (FIFO) sorting method, a Future Position Projection (FPP) method considering vehicle speed was designed, which improved the merging efficiency of ramps and vehicle operating speed. The proposed model was simulated using the SUMO platform and compared with other advanced methods. The experimental results showed that PPSC had good performance: in the scenario where the main ramp flow is 1500/500veh/h with a 10% penetration rate, compared with the IDM model, it achieved significant improvements in multiple indicators, reduced emergency braking rate by 76.31%, increased average speed by 67.19%, and reduced average waiting time by 35.47%. Finally, we also conducted robustness analysis to verify the stability of PPSC.