Scholarlink

<p>Traffic flow prediction is a critical yet challenging task due to the complex spatiotemporal dependencies in road networks. We propose MapsTSF, a novel framework that enhances traffic forecasting by integrating three innovative components. Spatial path embedding uses multiple traversal algorithms to capture dynamic spatial flow patterns across road networks. The hybrid Mamba2-transformer model combines Mamba2’s selective state-space modeling with transformer’s global attention to effectively capture intricate spatiotemporal dependencies. Additionally, the cross adaptive periodic sparse forecasting mechanism decouples periodic and trend features, reducing computational overhead while preserving high accuracy. Experiments on real-world datasets demonstrate that MapsTSF outperforms baseline methods by an average of 18.80% in MAE, 15.69% in RMSE, and 19.80% in MAPE. With its scalable and efficient design, MapsTSF is well-suited for real-time traffic management in smart cities and navigation applications, providing precise and reliable forecasts for large-scale, dynamic road networks.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MapsTSF: efficient traffic prediction via hybrid Mamba 2-transformer spatiotemporal modeling and cross adaptive periodic sparse forecasting

  • 期刊论文

  • 发表日期：

    2025年5月7日

查看全文
  • Bing Wang,
  • Chaoqi Cai,
  • Xingpeng Zhang,
  • Chunlan Zhao,
  • Chi Zhang,
  • Youming Zhang

摘要

Traffic flow prediction is a critical yet challenging task due to the complex spatiotemporal dependencies in road networks. We propose MapsTSF, a novel framework that enhances traffic forecasting by integrating three innovative components. Spatial path embedding uses multiple traversal algorithms to capture dynamic spatial flow patterns across road networks. The hybrid Mamba2-transformer model combines Mamba2’s selective state-space modeling with transformer’s global attention to effectively capture intricate spatiotemporal dependencies. Additionally, the cross adaptive periodic sparse forecasting mechanism decouples periodic and trend features, reducing computational overhead while preserving high accuracy. Experiments on real-world datasets demonstrate that MapsTSF outperforms baseline methods by an average of 18.80% in MAE, 15.69% in RMSE, and 19.80% in MAPE. With its scalable and efficient design, MapsTSF is well-suited for real-time traffic management in smart cities and navigation applications, providing precise and reliable forecasts for large-scale, dynamic road networks.