Meta-MSCC: A foundation model for adaptive CAV control in highway weaving segments☆,☆

Artikel i vetenskaplig tidskrift, 2025

The connected and automated vehicle (CAV) technology offers new opportunities for active traffic management in highway weaving segments. Current CAV-based control strategies have not fully accounted for their coupling impacts on safety and efficiency in highway weaving segments. Moreover, most existing active traffic control models still have limitations in adaptability and transferability across diverse weaving scenarios. To address these issues, this study proposes a novel multi-strategy cooperative control model based on meta-deep reinforcement learning algorithm, named Meta-MSCC, which serves as a lightweight domain-specific foundation model for optimizing both safety and efficiency in highway weaving segments under mixed connected automated traffic. A total of 22,306 real-world vehicle trajectories, including 1,488 near-crash events, are used to determine an acceptable collision risk level as the safety benchmark. Additionally, 20 high-risk and congested weaving scenarios are constructed through microscopic simulations to train and validate the Meta-MSCC. This study further develops a multi-level CAVbased control strategy by integrating microscopic (trajectory), mesoscopic (speed), and macroscopic (flow) control approaches. Then, a multi-strategy cooperative control model is put forward based on the actor-critic framework, which can effectively capture coupling impacts of this strategy on safety and efficiency. After incorporating a meta-learning algorithm (i.e., meta-deep deterministic policy gradient) into the actor-critic framework, the Meta-MSCC is finally formed as a lightweight domain-specific foundation model, enhancing adaptability and transferability across diverse weaving scenarios. The results show that the multi-level CAV-based control strategy achieves greater improvements in safety and efficiency than single-level strategies. Furthermore, the proposed Meta-MSCC outperforms other classical DRL-based models, with collision risk reduced by 49.1% and time delay by 12.6%, while maintaining stable optimization performance at CAV penetration rates exceeding 60%. The findings of this study might contribute to the enhancement of active traffic management and the optimization of intelligent road infrastructures.