【作者】 刘博；

【导师】 宋现敏；

【作者基本信息】 吉林大学 ， 交通运输硕士（专业学位）， 2025， 硕士

【摘要】 平面交叉口是城市路网的关键节点,也是交通冲突和拥堵的高发区域。科学的管控措施能够有效组织各向交通流,提升通行效率与安全性。传统管控方式主要通过信号灯分配通行权,车辆状态(如位置、速度等)仅作为环境变量,无法直接调控。随着自动驾驶和车联网技术的发展,智能网联车辆(Connected and Automated Vehicles,CAV)与人工驾驶车辆(Human-Driven Vehicles,HDV)混行将成为常态,推动交叉口管控向智能化、网联化转型。如何解析混行交通流运行规律,建立协同控制方法,成为当前研究的重点。智能网联环境下可以通过动态分配通行权和规划车辆轨迹之间的协同控制,实现对混行交通流的有效管控。目前基于数学模型的协同控制手段在环境感知和实时决策方面呈现出一定的局限性。采用数据驱动和基于深度强化学习(Deep Reinforcement Learning,DRL)的方法可以利用实时获取的多维度交通信息优化协同控制策略,提高城市交叉口控制系统的响应速度和处理能力。鉴于此,本文立足城市交通智能网联化的实际需求,基于DRL方法,在智能网联交叉口协同控制优化方面做了如下工作:考虑到与传统交叉口控制中HDV只能被动响应控制策略不同,CAV可以提供轨迹数据用于交叉口信号控制,也可以与控制系统交互从而进行自主决策,主动调整自身的路径、车道、速度,进而间接影响周围车辆的驾驶行为。首先对CAV车辆的微观驾驶行为进行建模与分析,然后针对混合智能网联环境下信号交叉口交通流协同控制这一场景展开研究,以实现交通运行效率、安全、生态和公平的协同优化为目标,提出一种基于异构多智能体与反馈寻优机制的流粒协同控制模型,通过交通信息感知与预处理模块、交叉口信号控制系统和最小生态控制单元的协同工作,实现混合智能网联环境下交叉口车速引导和信号的协同控制。为了验证所提流粒协控模型的效果,设计了基于异构图多智能体强化学习的IHG-MAAC算法对智能网联混行交通流协同控制问题进行优化求解。对于具有图结构的交通状态信息,通过异构图网络捕捉CAV、HDV与信号控制器之间的异构特征,获得更有效的特征表示,结合基于演员-评论家(Actor-Critic,AC)框架的多智能体强化学习实现协同优化。通过分布式训练和边缘计算架构,显著降低了计算复杂度,在保证优化效果的同时满足了实时性要求。最后以数值模拟实验为基础,利用模拟仿真软件SUMO搭建仿真环境,首先验证了CAV跟驰模型MIDM的有效性,然后从算法收敛性、有效性、鲁棒性等多个角度验证了所提算法的优越性,并进行了消融实验证明在智能网联协同控制问题场景下对算法做出的设计可以显著提升算法的求解能力。更多还原

【Abstract】 Planar intersections are critical nodes in urban road networks and hotspots for traffic conflicts and congestion.Scientific management measures can effectively organize traffic flows in all directions,improving traffic efficiency and safety.Traditional control methods primarily allocate right-of-way through traffic signals,treating vehicle states(such as position and speed)merely as environmental variables without direct regulation.With the advancement of autonomous driving and vehicle-to-everything(V2X)technologies,the mixed operation of connected and automated vehicles(CAVs)and human-driven vehicles(HDVs)has become the norm,driving the transformation of intersection control toward intelligence and connectivity.Understanding the operational patterns of mixed traffic flows and establishing collaborative control methods have become key research priorities.In a connected and automated environment,effective management of mixed traffic flows can be achieved through the coordinated control of dynamic right-of-way allocation and vehicle trajectory planning.Current model-based collaborative control methods exhibit limitations in environmental perception and real-time decision-making.Data-driven approaches and deep reinforcement learning(DRL)methods can optimize collaborative control strategies by leveraging real-time multi-dimensional traffic information,thereby enhancing the responsiveness and processing capabilities of urban intersection control systems.In light of this,this study addresses the practical needs of urban traffic intelligence and connectivity,leveraging DRL methods to make the following contributions in the optimization of collaborative control for smart and connected intersections:Unlike traditional intersection control,where HDVs can only passively respond to control strategies,CAVs can provide trajectory data for intersection signal control and interact with the control system to make autonomous decisions,actively adjusting their paths,lanes,and speeds,thereby indirectly influencing the driving behavior of surrounding vehicles.This study first models and analyzes the microscopic driving behavior of CAVs,then investigates the scenario of collaborative traffic flow control at signalized intersections in a mixed connected environment.Aiming to achieve coordinated optimization of traffic efficiency,safety,ecological impact,and fairness,a flow-granularity collaborative control model based on heterogeneous multi-agent and feedback optimization mechanisms is proposed.This model integrates traffic information perception and preprocessing modules,intersection signal control systems,and minimal ecological control units to achieve coordinated speed guidance and signal control at intersections in a mixed connected environment.To validate the effectiveness of the proposed flow-granularity control model,a heterogeneous graph multi-agent reinforcement learning algorithm,IHG-MAAC,is designed to optimize the collaborative control of mixed connected traffic flows.For traffic state information with graph structures,a heterogeneous graph network is employed to capture the heterogeneous features among CAVs,HDVs,and signal controllers,enabling more effective feature representation.Combined with the actor-critic(AC)framework-based multi-agent reinforcement learning,collaborative optimization is achieved.Through distributed training and edge computing architectures,computational complexity is significantly reduced,meeting real-time requirements while ensuring optimization performance.Finally,numerical simulation experiments are conducted using the SUMO simulation software.The effectiveness of the proposed CAV car-following model,MIDM,is first validated.Then,the superiority of the proposed algorithm is verified from multiple perspectives,including convergence,effectiveness,and robustness.Ablation experiments are conducted to demonstrate that the algorithmic design enhances its problem-solving capabilities in the context of collaborative control for smart and connected intersections.更多还原