基于深度强化学习的道路交叉口生态驾驶策略研究

Signalized Intersection Eco-driving Strategy Based on Deep Reinforcement Learning

在互联和自动驾驶环境下,生态驾驶具有显著的潜力,可提高交通效率并降低能源消耗和排放。本文探讨一种基于深度强化学习算法的生态驾驶策略,该算法可优化互联自动驾驶汽车(CAV)的纵向操纵和横向决策;将状态空间分为与车辆动态特性相关的局部变量,以及与信号交叉口相关的全局变量,确保CAV与环境之间的充分互动;奖励函数综合考虑了车辆的驾驶要求,与信号灯的协同作用以及全局节能激励因素;此外,设计一个典型的城市道路场景训练模型。结果表明,在信号灯和智能体输出协同控制下,本文提出的策略可以实现CAV的生态驾驶,并确保CAV准确驶入目标车道;在动态交通环境下进行仿真显示,通过控制多辆CAV引导人工驾驶车辆,本文方法可将交叉路口的通行能力提高约17.90%,并将交通系统的燃料消耗和污染物排放降低约8.76%。

Eco-driving in a connected and autonomous driving environment has great potential to improve traffic efficiency,energy saving,and emission reduction.This paper proposes a prosocial eco-driving strategy based on the deep reinforcement learning algorithm that optimizes the longitudinal manipulation and lateral decision-making of the connected and automated vehicle(CAV).The state space is divided into the local variables related to dynamic vehicle characteristics and the global variables associated with signalized intersection to ensure adequate interaction between the CAV and the roadway environment.The designed reward function integrates the vehicle driving requirements,synergy with signals and global energy saving incentives.In addition,this study developed a typical urban road intersection scenario to train the model.The results show that the proposed strategy can achieve eco-driving of the CAV in collaboration with the signal and output lateral control to ensure the vehicle travels to the target lane.In addition,simulations in a dynamic traffic environment reveal that the proposed method can improve the capacity at the intersection by about 17.90%and reduce the traffic system's fuel consumption and pollutant emissions by approximately 8.76%through the control of multiple CAVs to guide the human-driven vehicles.