{"title":"Adaptive multi-scale spatio-temporal convolutional network with reinforcement learning for dynamic lane-level traffic flow prediction","authors":"Xiaohui Yang , Shaowei Sun , Mingzhou Liu","doi":"10.1016/j.array.2025.100513","DOIUrl":null,"url":null,"abstract":"<div><div>This paper presents an Adaptive Multi-Scale Spatio-Temporal Convolutional Network and Reinforcement Learning Collaborative Optimization Lane-Level Traffic Flow Prediction Model (AST-RLM), designed to address the challenges posed by the sudden changes in microscopic driving behaviors and spatio-temporal dependencies in complex urban environments. The model achieves high-precision lane-level traffic flow prediction through dynamic graph construction mechanisms, heterogeneous perception-based multi-scale convolutional networks, and a DQN-based collaborative optimization framework. Experimental results demonstrate that AST-RLM performs exceptionally well on real-world datasets from multiple cities, containing over 10,000 lanes. The average absolute error (MAE) during the evening peak is as low as 0.033, a 38.9 % reduction compared to GraphWaveNet. The root mean square error (RMSE) for 30-min predictions is 3.98, outperforming existing models like ST-MetaNet, and the model maintains 92.4 % stability even in extreme weather conditions. Notably, during sudden events like traffic accidents, the dynamic graph module adapts in real-time to changes in topology, reducing prediction errors by 26.7 %–30.9 %, significantly improving the model's robustness and responsiveness in complex dynamic scenarios. Furthermore, AST-RLM's multi-agent reinforcement learning deployment on edge devices achieves a convergence speed 3.6 times faster than GC-RL, validating its efficiency and feasibility in real-world traffic systems.</div></div>","PeriodicalId":8417,"journal":{"name":"Array","volume":"28 ","pages":"Article 100513"},"PeriodicalIF":4.5000,"publicationDate":"2025-09-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Array","FirstCategoryId":"1085","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2590005625001407","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"COMPUTER SCIENCE, THEORY & METHODS","Score":null,"Total":0}
引用次数: 0
Abstract
This paper presents an Adaptive Multi-Scale Spatio-Temporal Convolutional Network and Reinforcement Learning Collaborative Optimization Lane-Level Traffic Flow Prediction Model (AST-RLM), designed to address the challenges posed by the sudden changes in microscopic driving behaviors and spatio-temporal dependencies in complex urban environments. The model achieves high-precision lane-level traffic flow prediction through dynamic graph construction mechanisms, heterogeneous perception-based multi-scale convolutional networks, and a DQN-based collaborative optimization framework. Experimental results demonstrate that AST-RLM performs exceptionally well on real-world datasets from multiple cities, containing over 10,000 lanes. The average absolute error (MAE) during the evening peak is as low as 0.033, a 38.9 % reduction compared to GraphWaveNet. The root mean square error (RMSE) for 30-min predictions is 3.98, outperforming existing models like ST-MetaNet, and the model maintains 92.4 % stability even in extreme weather conditions. Notably, during sudden events like traffic accidents, the dynamic graph module adapts in real-time to changes in topology, reducing prediction errors by 26.7 %–30.9 %, significantly improving the model's robustness and responsiveness in complex dynamic scenarios. Furthermore, AST-RLM's multi-agent reinforcement learning deployment on edge devices achieves a convergence speed 3.6 times faster than GC-RL, validating its efficiency and feasibility in real-world traffic systems.