Scholarlink

<p>Traffic flow forecasting is a key component of intelligent transportation systems (ITSs), playing a critical role in congestion mitigation, efficiency improvement, and traffic management optimization. With the rapid growth of traffic data, supercomputing capabilities have become increasingly important for processing large-scale data and ensuring real-time updates, thereby supporting timely and effective decision-making. This work investigates an aggregated graph attention network with a depthwise separable convolution fusion network (AGAT-DSCFN), designed to effectively capture multidimensional traffic dependencies for real-time traffic flow forecasting. Specifically, we propose an aggregated graph attention network (AGAT) to capture the interconnections between road nodes based on the topology of dynamic traffic networks. Moreover, to efficiently handle large-scale data and ensure real-time updates, we construct a temporal-channel feature coupling (TCFC) module with a depthwise separable convolution fusion network (DSCFN), enabling independent processing of each feature channel while capturing the interdependencies among traffic conditions. Finally, we design a periodic embedding layer (PEL) to accurately model periodic features in traffic data. Experimental results demonstrate that AGAT-DSCFN accurately explores complex spatiotemporal dependencies across four real-world traffic datasets, meeting the real-time demands of traffic flow prediction.</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

An aggregated graph attention network with depthwise separable convolution fusion for traffic flow forecasting

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Dawen Xia,
  • Jian Zhang,
  • Wenchao Weng,
  • Wenlin He,
  • Yanmin Liu,
  • Yang Hu,
  • Fuchu Zhang,
  • Yantao Li,
  • Huaqing Li

摘要

Traffic flow forecasting is a key component of intelligent transportation systems (ITSs), playing a critical role in congestion mitigation, efficiency improvement, and traffic management optimization. With the rapid growth of traffic data, supercomputing capabilities have become increasingly important for processing large-scale data and ensuring real-time updates, thereby supporting timely and effective decision-making. This work investigates an aggregated graph attention network with a depthwise separable convolution fusion network (AGAT-DSCFN), designed to effectively capture multidimensional traffic dependencies for real-time traffic flow forecasting. Specifically, we propose an aggregated graph attention network (AGAT) to capture the interconnections between road nodes based on the topology of dynamic traffic networks. Moreover, to efficiently handle large-scale data and ensure real-time updates, we construct a temporal-channel feature coupling (TCFC) module with a depthwise separable convolution fusion network (DSCFN), enabling independent processing of each feature channel while capturing the interdependencies among traffic conditions. Finally, we design a periodic embedding layer (PEL) to accurately model periodic features in traffic data. Experimental results demonstrate that AGAT-DSCFN accurately explores complex spatiotemporal dependencies across four real-world traffic datasets, meeting the real-time demands of traffic flow prediction.