交通运输系统工程与信息

交通运输系统工程与信息 ›› 2026, Vol. 26 ›› Issue (2): 125-136.DOI: 10.16097/j.cnki.1009-6744.2026.02.012

• 智能交通系统与信息技术 • 上一篇    下一篇

网联自动驾驶车辆专用车道动态宏微观协同部署方法

谷远利*1,2 ，宇泓儒1 ，陈龙1 ，邓社军3 ，陆文琦4   

  1. 1. 北京交通大学，综合交通运输大数据应用技术交通运输行业重点实验室，北京100044； 2. 河北省高校交通基础设施数智化应用技术研发中心，河北沧州061000； 3. 扬州大学，土木与交通学院，江苏扬州225127；4.香港科技大学，土木及环境工程学系，香港999077
  • 收稿日期:2025-12-29 修回日期:2026-03-06 接受日期:2026-03-12 出版日期:2026-04-25 发布日期:2026-04-20
  • 作者简介:谷远利（1973—），男，辽宁海城人，教授，博士。
  • 基金资助:
    河北省交通运输厅科技计划项目 (CZ202510)；河北省高校交通基础设施数智化应用技术研发中心开放课题 (HBWE 2025KF-01)。

Dynamic Macro-Micro Collaborative Deployment Method for Connected and Automated Vehicle Exclusive Lanes

GU Yuanli*1,2, YU Hongru1, CHEN Long1, DENG Shejun3, LU Wenqi4   

  1. 1. Key Laboratory of Transport Industry of Big Data Application Technologies for Comprehensive Transport, Ministry of Transport, Beijing Jiaotong University, Beijing 100044, China; 2. Hebei Higher Institute of Transportation Infrastructure Research and Development Center for Digital and Intelligent Technology Application, Cangzhou 061000, Hebei, China; 3. School of Civil Engineering and Transportation, Yangzhou University, Yangzhou 225127, Jiangsu, China; 4. Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology, Hong Kong 999077, China
  • Received:2025-12-29 Revised:2026-03-06 Accepted:2026-03-12 Online:2026-04-25 Published:2026-04-20
  • Supported by:
    Hebei Provincial Department of Transportation Science and Technology Program (CZ202510)；Hebei Higher Institute of Transportation Infrastructure Research and Development Center for Digital and Intelligent Technology Application Open Call for Research Proposals (HBWE 2025KF-01)。

PDF

摘要： 针对当前混合交通环境下网联自动驾驶汽车专用车道（Connected and Automated Vehicle Lane,CAVL）部署适应性差、宏微观决策割裂与实时响应能力不足等问题，本文提出一种融合渗流理论与多智能体深度强化学习的动态宏微观协同控制（Dynamic Macro-Micro Collaborative Deployment, DMMCD）方法。宏观层面，基于路网实时交通流状态动态更新渗流阈值，聚合高渗透率路段形成CAVL骨干网络，确保混行路网拓扑结构的连通性；微观层面，构建基于多智能体深度Q网络（Multi-AgentDeepQ-Network,MADQN）的分布式训练框架，通过多目标奖励机制优化车道级CAVL部署策略。渗流层通过动作空间宏观骨架强约束，使宏观策略直接引导微观决策；微观层依托端到端模型，实现对实时交通流的动态响应；车道级交通状态再反馈至宏观层，驱动骨干网络迭代更新，实现宏微观决策的耦合联动。最后，基于7×7的双向4车道拓扑路网进行仿真验证。结果表明：在CAV渗透率为0.1~0.9范围内，本文方法平均出行时间较分层人工蜂群算法与无专用车道方案分别降低6.77%和28.15%，平均行驶速度提升9.18%和23.94%，系统CO2排放量降低11.35%以上；当CAV渗透率超过0.5时，本文方法优势进一步扩大，出行时间降幅可达31.12%以上。此外，本文智能体单次策略生成时间仅0.12ms，能够满足实时CAVL管控需求。相关成果能够为城市路网环境下的CAVL动态部署提供理论指导。

关键词: 智能交通, 专用车道部署, 多智能体强化学习, 网联自动驾驶, 车路协同

Abstract: Aiming at the problems of poor adaptability, separation of macro-micro decision-making and insufficient real-time response in the deployment of Connected and Automated Vehicle Lanes (CAVL) under the current mixed traffic environment, this study proposes a Dynamic Macro-Micro Collaborative Deployment (DMMCD) method integrating percolation theory with multi agent deep reinforcement learning. At the macro level, the percolation threshold is dynamically updated according to the real-time traffic flow state of the road network, and road segments with high CAV penetration rates are aggregated to form a CAVL backbone network while ensuring the topological connectivity of the mixed traffic road network. At the micro level, a distributed training framework based on the Multi-Agent Deep Q-Network (MA-DQN) is constructed, and the lane-level CAVL deployment strategy is optimized through a multi-objective reward mechanism. The percolation layer enables macro strategies to directly guide micro decision-making via the macro-skeleton strong constraint on the action space; the micro layer realizes dynamic response to real-time traffic flow by virtue of an end-to-end model, and the lane-level traffic state send feedback to the macro layer to drive the iterative update of the backbone network, thus achieving the coupling and linkage of macro-micro decision-making. The simulation verification is carried out based on a 7×7 topological road network with bidirectional four lanes. The results show that within the CAV penetration rate range of 0.1 to 0.9, the proposed method reduces the average travel time by 6.77% and 28.15%, increases the average travel speed by 9.18% and 23.94%, and reduces the system CO2 emissions by more than 11.35%, compared with the hierarchical artificial bee colony algorithm and the scheme without dedicated lanes, respectively. When the CAV penetration rate exceeds 0.5, the advantage of the proposed method is further amplified, which shows a travel time reduction of more than 31.12%. In addition, the average strategy generation time per agent of the proposed model is 0.12 milliseconds, which meets the requirements of real-time CAVL management and control. The study and results provide theoretical basis for the dynamic deployment of CAVLin urban road network environment.

Key words: intelligent transportation, exclusive lane deployment, multi-agent reinforcement learning, connected automated driving, vehicle-infrastructure cooperation

中图分类号: