Scholarlink

<p>Accurately modeling complex spatiotemporal dependencies remains a central challenge in traffic flow prediction, particularly for capturing long-range temporal recurrences and hierarchical spatial structures. To address these issues, we propose the Hierarchical Memory-Guided Graph Fusion Network (HMGF-Net). For spatiotemporal modeling, HMGF-Net incorporates a Momentum-Driven Memory Refinement Block (MD-MRB), which dynamically stores, retrieves, and refines prototypical traffic patterns through an Adaptive Momentum Gated Unit (AMGU), thereby enhancing pattern-level representation learning. For spatial modeling, we further introduce a Hierarchical Graph Fusion Block (HGFB) to integrate dynamic memory features with multi-scale static graph representations. Extensive experiments on four real-world traffic forecasting benchmarks, namely PEMS03, PEMS04, PEMS07, and PEMS08, demonstrate that HMGF-Net achieves superior overall performance over representative baselines while maintaining favorable computational efficiency. Specifically, compared with the strongest competing baseline under the same evaluation protocol, HMGF-Net reduces the Mean Absolute Error (MAE) on PEMS08 by approximately 2.2%, the Mean Absolute Percentage Error (MAPE) on PEMS07 by approximately 1.7%, and the MAE on PEMS04 by approximately 1.3%. In addition, under identical hardware and training settings, HMGF-Net lowers the per-epoch training and inference time by approximately 11.7% and 5.4%, respectively. The implementation is publicly available at <a href="https://github.com/sherru81/HMGF-Net">https://github.com/sherru81/HMGF-Net</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

HMGF-Net: Hierarchical memory-guided graph fusion network for traffic flow prediction

  • 期刊论文

  • 发表日期：

    2026年7月16日

查看全文
  • Lian Xiong,
  • Xing Guo,
  • Yuxi Feng,
  • Wanchang Li

摘要

Accurately modeling complex spatiotemporal dependencies remains a central challenge in traffic flow prediction, particularly for capturing long-range temporal recurrences and hierarchical spatial structures. To address these issues, we propose the Hierarchical Memory-Guided Graph Fusion Network (HMGF-Net). For spatiotemporal modeling, HMGF-Net incorporates a Momentum-Driven Memory Refinement Block (MD-MRB), which dynamically stores, retrieves, and refines prototypical traffic patterns through an Adaptive Momentum Gated Unit (AMGU), thereby enhancing pattern-level representation learning. For spatial modeling, we further introduce a Hierarchical Graph Fusion Block (HGFB) to integrate dynamic memory features with multi-scale static graph representations. Extensive experiments on four real-world traffic forecasting benchmarks, namely PEMS03, PEMS04, PEMS07, and PEMS08, demonstrate that HMGF-Net achieves superior overall performance over representative baselines while maintaining favorable computational efficiency. Specifically, compared with the strongest competing baseline under the same evaluation protocol, HMGF-Net reduces the Mean Absolute Error (MAE) on PEMS08 by approximately 2.2%, the Mean Absolute Percentage Error (MAPE) on PEMS07 by approximately 1.7%, and the MAE on PEMS04 by approximately 1.3%. In addition, under identical hardware and training settings, HMGF-Net lowers the per-epoch training and inference time by approximately 11.7% and 5.4%, respectively. The implementation is publicly available at https://github.com/sherru81/HMGF-Net.