Scholarlink

<p>Accurate traffic forecasting optimizes traffic management and urban planning, increasing efficiency and cutting congestion. However, the complex spatio-temporal dependencies inherent in traffic data present substantial challenges to achieving high predictive precision. Existing research often focuses on short-term high-frequency fluctuations in traffic flow while neglecting long-term low-frequency trends. Additionally, it emphasizes interactions between neighboring nodes but fails to account for the global correlations between distant nodes in traffic networks. To address these issues, this study proposes a Temporal Feature-Enhanced Graph Convolutional Recurrent Network (TFEGCRN) for traffic forecasting. TFEGCRN includes a long-term temporal feature extractor to capture low-frequency trend information, along with a short-term spatio-temporal module to model high-frequency variations and complex spatio-temporal dependencies. By incorporating an adaptive adjacency matrix, TFEGCRN dynamically constructs connection weights between nodes. The time graph module enables TFEGCRN to capture both global and local temporal dependencies by modeling each time step as a graph node, leveraging a multi-head attention mechanism to enhance long-term sequence representation. Furthermore, the local–global convolution module is introduced to process bidirectional inflow and outflow information, significantly improving TFEGCRN’s ability to capture multi-scale spatial relationships in dynamic traffic networks. We conducted extensive experiments on four real-world traffic datasets, and the results demonstrate that TFEGCRN significantly outperforms other baselines, including reducing MAE, RMSE, and MAPE by about 2.63%, 2.25%, and 4.91% on the PeMS03 dataset, and by 1.16%, 0.93%, and 1.81% on the PeMS08 dataset, showcasing robust performance, practical value, and a stronger capability in capturing multi-scale spatio-temporal dependencies.</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

Temporal feature-enhanced graph Convolutional Recurrent Network for traffic forecasting

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Zheng Wu,
  • Lai Wei

摘要

Accurate traffic forecasting optimizes traffic management and urban planning, increasing efficiency and cutting congestion. However, the complex spatio-temporal dependencies inherent in traffic data present substantial challenges to achieving high predictive precision. Existing research often focuses on short-term high-frequency fluctuations in traffic flow while neglecting long-term low-frequency trends. Additionally, it emphasizes interactions between neighboring nodes but fails to account for the global correlations between distant nodes in traffic networks. To address these issues, this study proposes a Temporal Feature-Enhanced Graph Convolutional Recurrent Network (TFEGCRN) for traffic forecasting. TFEGCRN includes a long-term temporal feature extractor to capture low-frequency trend information, along with a short-term spatio-temporal module to model high-frequency variations and complex spatio-temporal dependencies. By incorporating an adaptive adjacency matrix, TFEGCRN dynamically constructs connection weights between nodes. The time graph module enables TFEGCRN to capture both global and local temporal dependencies by modeling each time step as a graph node, leveraging a multi-head attention mechanism to enhance long-term sequence representation. Furthermore, the local–global convolution module is introduced to process bidirectional inflow and outflow information, significantly improving TFEGCRN’s ability to capture multi-scale spatial relationships in dynamic traffic networks. We conducted extensive experiments on four real-world traffic datasets, and the results demonstrate that TFEGCRN significantly outperforms other baselines, including reducing MAE, RMSE, and MAPE by about 2.63%, 2.25%, and 4.91% on the PeMS03 dataset, and by 1.16%, 0.93%, and 1.81% on the PeMS08 dataset, showcasing robust performance, practical value, and a stronger capability in capturing multi-scale spatio-temporal dependencies.