Abstract
The sudden changes in light and slope at the entrance of mountain tunnels can easily lead to unstable traffic flow, especially in mixed traffic flow containing ICV. The existing control methods are difficult to simultaneously address three challenges: dynamically changing lighting and slope environments, vehicle data privacy protection requirements, and insufficient adaptability to different tunnel scenarios. This study developed the MF3DQN-TF control framework to address these issues by integrating environmental perception and vehicle control. The framework first establishes a correlation model between light gradient and slope resistance, converting environmental risks into quantifiable control signals; Secondly, design an intelligent weight allocation mechanism to automatically increase the decision-making priority of environmental factors in strong light areas; Finally, a distributed training architecture is adopted to achieve knowledge sharing in multi tunnel scenarios while protecting vehicle data privacy. The verification results show that in typical tunnel testing scenarios, this framework significantly improves performance compared to traditional methods: the amplitude of speed fluctuations is reduced by about 40% compared to conventional control methods, the risk of rear end collisions is reduced to one-third of that of traditional reinforcement learning schemes, and the communication transmission volume is only half of that of typical federated learning methods. The verification results show that in typical tunnel testing scenarios, this framework reduces speed standard deviation by 41.4% (vs. PID control), cuts high-risk TTC < 2 s events by 62.5% (vs. centralized DQN), and nearly halves communication volume (46.2% reduction vs. FedProx). These improvements stem from the collaborative processing mechanism of the framework for environmental risks and vehicle status, providing a new technological path for safety control in complex tunnel environments.