Scholarlink

<p>In traffic flow prediction, Graph Convolutional Networks (GCNs) are widely used due to their powerful ability to model non-Euclidean road networks. However, current GCN-based methods have following limitations: (1) The lack of distinction between the diffusion and non-diffusion components in traffic flow causes a decrease in prediction accuracy. (2) Stacking GCN layers reduces the distinguishability and diversity of graph nodes, leading to an over-smoothing issue. (3) Predefined or adaptive static graphs are inadequate for simulating the dynamic spatial correlations between road nodes. To address these problems, this paper proposes a new traffic prediction method called Multi-Branch Heterogeneous Spatial-Temporal Graph Convolutional Network (MHSTGCN). MHSTGCN consists of three parallel heterogeneous branches. Specifically, we use branches 1 and 2 with GCN, and branch 3 without GCN to model diffusion and non-diffusion components, respectively. To alleviate the over-smoothing issue, we equally distribute the GCN layers that would have been stacked in a single branch between branches 1 and 2. Meanwhile, we add dynamic graphs to capture dynamic spatial correlations to model the dynamics of traffic. The effectiveness and robustness of the proposed method are demonstrated by its state-of-the-art performance on four public traffic flow datasets.</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

Multi-branch heterogeneous spatial-temporal graph convolutional network for traffic flow forecasting

  • 期刊论文

  • 发表日期：

    2026年2月17日

查看全文
  • Tingyu Lin,
  • Qunyong Wu,
  • Keyue Wang,
  • Shiyu Yang

摘要

In traffic flow prediction, Graph Convolutional Networks (GCNs) are widely used due to their powerful ability to model non-Euclidean road networks. However, current GCN-based methods have following limitations: (1) The lack of distinction between the diffusion and non-diffusion components in traffic flow causes a decrease in prediction accuracy. (2) Stacking GCN layers reduces the distinguishability and diversity of graph nodes, leading to an over-smoothing issue. (3) Predefined or adaptive static graphs are inadequate for simulating the dynamic spatial correlations between road nodes. To address these problems, this paper proposes a new traffic prediction method called Multi-Branch Heterogeneous Spatial-Temporal Graph Convolutional Network (MHSTGCN). MHSTGCN consists of three parallel heterogeneous branches. Specifically, we use branches 1 and 2 with GCN, and branch 3 without GCN to model diffusion and non-diffusion components, respectively. To alleviate the over-smoothing issue, we equally distribute the GCN layers that would have been stacked in a single branch between branches 1 and 2. Meanwhile, we add dynamic graphs to capture dynamic spatial correlations to model the dynamics of traffic. The effectiveness and robustness of the proposed method are demonstrated by its state-of-the-art performance on four public traffic flow datasets.