Scholarlink

<p>Urban traffic Internet of Things sensors continuously collect multidimensional time-series data, whose volume grows exponentially as the road network expands. Modeling the dynamic spatio-temporal correlation of traffic flow requires substantial computational load, while intelligent transportation scenarios demand low prediction latency, and both of these requirements necessitate the parallel and distributed processing capabilities of high-performance computing. Despite the significance of traffic flow prediction for urban optimization, capturing its complex spatio-temporal dependencies remains challenging due to limitations in existing methods. To solve this, we propose an efficient dynamic graph convolutional interactive network (DGCINet), which integrates graph convolution into interactive learning, adopts adaptive dynamic graph convolution, and the spatio-temporal Transformer for feature extraction. Across four real datasets, this model’s predictive performance was significantly better than that of the nine baseline methods, reducing the average MAE by 6.3%. Additionally, real-time performance analysis confirms that DGCINet possesses excellent real-time processing capabilities.</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

Dynamic graph convolution and interaction network for traffic flow forecasting

  • 期刊论文

  • 发表日期：

    2026年2月28日

查看全文
  • Hongxia He,
  • Xi Li,
  • Peng Chen,
  • Wenyu Shan,
  • Shufeng Cao,
  • Lei Wu

摘要

Urban traffic Internet of Things sensors continuously collect multidimensional time-series data, whose volume grows exponentially as the road network expands. Modeling the dynamic spatio-temporal correlation of traffic flow requires substantial computational load, while intelligent transportation scenarios demand low prediction latency, and both of these requirements necessitate the parallel and distributed processing capabilities of high-performance computing. Despite the significance of traffic flow prediction for urban optimization, capturing its complex spatio-temporal dependencies remains challenging due to limitations in existing methods. To solve this, we propose an efficient dynamic graph convolutional interactive network (DGCINet), which integrates graph convolution into interactive learning, adopts adaptive dynamic graph convolution, and the spatio-temporal Transformer for feature extraction. Across four real datasets, this model’s predictive performance was significantly better than that of the nine baseline methods, reducing the average MAE by 6.3%. Additionally, real-time performance analysis confirms that DGCINet possesses excellent real-time processing capabilities.