关键词: 智能交通, 协同换道, 深度强化学习, 智能网联汽车, 施工区

Abstract: To address traffic bottlenecks and safety risks induced by work zones on urban expressways, this study proposes a coordinated lane-changing method for connected and automated vehicles (CAV) based on the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. Considering the spatiotemporally deterministic characteristics of work zones, a staged control strategy is developed, consisting of "advance preparation and platoon splitting in the upstream warning zone-guided lane changing in the transition zone-steady mixed traffic maintenance in the work zone". Incorporating the concept of "particle-flow integration", a two-layer Markov Decision Process (MDP) model for CAV coordinated lane-changing in work zones is established. An extended reward function is designed to accommodate platoon reconfiguration and merging tasks. The problem is solved using an improved Predictive Recurrent Multi-Agent Deep Deterministic Policy Gradient (PR-MADDPG) algorithm, which encodes work zone partitioning information into the state space, enabling agents to make anticipatory decisions based on future traffic conflict types. Simulation experiments are conducted on the SUMO-Python co-simulation platform. The results demonstrate that the proposed method can reliably trigger control behaviors such as platoon splitting, gap creation, cooperative yielding, and downstream reformation under various work zone scenarios. In three types of lane-closure scenarios, the standard deviation of lane changing speed remains below 3.5, indicating small speed fluctuations and smooth lane- changing maneuvers. Furthermore, compared with the improved Deep Q-Network (DQN) algorithm, within a CAV penetration rate range of 31%~60%, the average traffic flow increases by 7.6% and 7.1% in the inner dedicated lane and outer mixed lane scenarios, respectively, while the peak flow increases by 7.9% and 8.5%. The most significant improvement is observed in the middle mixed-lane closure scenario, where, at densities of 20~40 veh·km-1, the traffic flow is increased by 12.6% compared to the improved DQN algorithm. Moreover, the PR MADDPG strategy reduces the traffic conflict probability (Time-To-Collision, TTC<4 s) by an average of 79.8% and the occurrence of extreme deceleration maneuvers by approximately 80%. Furthermore, horizontal comparisons with mainstream multi-agent algorithms such as Q-Value Mixing (QMIX) and Multi-Agent Proximal Policy Optimization (MAPPO) demonstrate that, taking the middle lane closure scenario as an example, the peak flow of the proposed method is improved by 6.8% and 3.9%, respectively, and the extreme deceleration maneuvers are reduced by 75.4% and 63.7%, exhibiting superior cooperative control capabilities in complex and restricted spaces. Ablation studies validate the independent performance gains of the zone encoding, dual- layer MDP architecture, and policy clipping mechanism in enhancing decision predictability, ensuring macroscopic coordination, and maintaining training stability. These results indicate that the proposed strategy can effectively mitigate upstream traffic disturbances in work zones and enhance traffic efficiency and operational stability under mixed traffic conditions.

Key words: intelligent transportation, coordinated lane-changing, deep reinforcement learning, connected and automated vehicle, work zone