Scholarlink

<p>Accurate traffic forecasting, a core function of Intelligent Transportation Systems (ITS), is critical for both road users and traffic management. However, the complex spatiotemporal dependencies inherent in traffic data pose significant challenges to existing methods, often resulting in suboptimal performance. To address this issue, we propose the Adaptive Parallel Spatio-Temporal Graph Convolution Network (APSTGCN), a novel model designed to adaptively capture these complex dependencies. APSTGCN integrates predefined road information and historical traffic data to generate dynamic graphs. It employs a parallel spatial-temporal convolution module with node attention to fully exploit potential patterns in dynamic graphs, enabling effective multi-step traffic forecasting. Extensive experiments conducted on six real-world traffic datasets demonstrate that APSTGCN significantly outperforms state-of-the-art methods. Specifically, compared to baseline methods, APSTGCN achieves improvements up to 4.16%, 5.45%, and 5.59% in MAE, RMSE, and MAPE, respectively. Code is available at <a href="https://github.com/bin1030/APSTGCN">https://github.com/bin1030/APSTGCN</a>.</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

Apstgcn: adaptive parallel spatio-temporal graph convolution network for traffic forecasting

  • 期刊论文

  • 发表日期：

    2025年7月31日

查看全文
  • Xiaohan Yu,
  • Zebin Hu,
  • Seung Jun Baek,
  • Chen Shen,
  • Chao Chen

摘要

Accurate traffic forecasting, a core function of Intelligent Transportation Systems (ITS), is critical for both road users and traffic management. However, the complex spatiotemporal dependencies inherent in traffic data pose significant challenges to existing methods, often resulting in suboptimal performance. To address this issue, we propose the Adaptive Parallel Spatio-Temporal Graph Convolution Network (APSTGCN), a novel model designed to adaptively capture these complex dependencies. APSTGCN integrates predefined road information and historical traffic data to generate dynamic graphs. It employs a parallel spatial-temporal convolution module with node attention to fully exploit potential patterns in dynamic graphs, enabling effective multi-step traffic forecasting. Extensive experiments conducted on six real-world traffic datasets demonstrate that APSTGCN significantly outperforms state-of-the-art methods. Specifically, compared to baseline methods, APSTGCN achieves improvements up to 4.16%, 5.45%, and 5.59% in MAE, RMSE, and MAPE, respectively. Code is available at https://github.com/bin1030/APSTGCN.