基于时空图卷积网络与强化学习的高速公路应急车道智能开启决策研究

Intelligent Decision-Making for Highway Emergency Lane Opening Based on Spatio-Temporal Graph Convolutional Network and Reinforcement Learning

摘要: 高速公路拥堵是城市交通管理的核心挑战，科学开启应急车道可有效缓解拥堵。传统决策依赖人工经验，缺乏实时性与预测性。为此，本研究提出融合时空图卷积网络(ST-GCN)、自注意力机制与近端策略优化(PPO)的智能决策框架。首先，利用YOLOv8从监控视频中提取车流量、车速和密度；其次，构建ST-GCN捕获交通流时空依赖，引入自注意力机制动态计算节点权重；然后，设计PPO智能体，以交通状态预测为输入，输出应急车道开启的最优决策；最后，建立多维度评价体系。实验结果表明，交通流预测RMSE分别为3.83 (车流量)、4.21 (车速)、0.07 (密度)，拥堵检测准确率达92.1%，应急车道开启决策的精确率和准确率分别为70.0%和87.5%。本研究实现了从状态预测到决策的端到端智能框架，为高速公路实时拥堵管理提供了科学决策支持。

Abstract: Highway congestion is a core challenge in urban traffic management. Scientifically opening emergency lanes can effectively alleviate congestion. Traditional decision-making relies on manual experience and lacks real-time and predictive capabilities. Therefore, this study proposes an intelligent decision-making framework that integrates the Spatial-Temporal Graph Convolutional Network (ST-GCN), the self-attention mechanism, and Proximal Policy Optimization (PPO). Firstly, YOLOv8 is used to extract traffic volume, speed, and density from surveillance videos; secondly, ST-GCN is constructed to capture the spatio-temporal dependencies of traffic flow, and the self-attention mechanism is introduced to dynamically calculate node weights; then, a PPO agent is designed, with traffic state prediction as the input and outputting the optimal decision for opening emergency lanes; finally, a multi-dimensional evaluation system is established. Experimental results show that the traffic flow prediction RMSE is 3.83 (traffic volume), 4.21 (speed), and 0.07 (density), the congestion detection accuracy rate is 92.1%, and the precision and accuracy of the emergency lane opening decision are 70.0% and 87.5% respectively. This study has achieved an end-to-end intelligent framework from state prediction to decision-making, providing scientific decision-making support for real-time highway congestion management.