Abstract: To enhance driving safety, traffic efficiency and environmental performance of highway merging areas, the variable speed limit control strategy was investigated. Based on real-time traffic flow status obtained by the detector, with the shortest travel time and collision time as safety objectives, the optimal speed limit value was determined with the consideration of the traffic demand in different areas, and a hierarchical variable speed limit control strategy was established. By integrating prioritized experience replay and the ε-Greedy strategy into the double deep Q network (DDQN) algorithm, the optimized double deep Q network (OPDDQN) algorithm was proposed to improve the utilization rate of critical samples and enhance the stability of exploration strategy by training , thereby obtaining the corresponding control strategy. Relying on the SUMO platform to build a simulation environment, the effectiveness of the proposed strategy was verified under low, medium and high traffic flow conditions. The research results indicate that compared to the DDQN algorithm, OPDDQN reduces training time by 33%. In contrast to the no-control scenario, the proposed strategy reduces collision risk, travel time, and fuel consumption by 39.18%, 48.30%, and 34.48%, respectively, while increasing average speed by 67.05%.

Key words: traffic engineering; intelligent transportation; hierarchical variable speed limit; safety simulation; deep reinforcement learning

摘要： 为提升高速公路合流区的行车安全、通行效率与环保性能，对可变限速控制策略展开了研究。基于检测器获取的实时交通流状态，以最短行程时间与碰撞时间为安全目标，结合不同区域通行需求确定最优限速值，构建了分级可变限速控制策略；在双深度Q网络(double deep Q network, DDQN)算法中引入优先经验回放与ε-Greedy策略，提出优化双深度Q网络(optimized DDQN, OPDDQN)算法，以提高关键样本利用率并增强探索策略稳定性，进而训练得到相应的控制策略；依托SUMO平台搭建仿真环境，在低、 中、 高不同流量条件下验证了该策略的效果。研究结果表明：与DDQN算法相比，OPDDQN算法训练时间可节省33%；相较于无控制场景，所提策略使碰撞风险、行程时间、燃油消耗分别降低了39.18%、 48.30%、 34.48%，平均速度提升了67.05%。

关键词: 交通工程； 智能交通； 分级可变限速； 安全仿真； 深度强化学习