Scholarlink

<p>Accurately predicting traffic flow becomes significant as urban transportation networks become more complex and data-intensive. To address the challenge of capturing dynamic spatiotemporal dependencies under real-time requirements, we propose AGBi-Mamba, a traffic flow prediction method based on dynamic graph convolution and Mamba. Specifically, the model employs an FFT-based adaptive graph convolution (FFT-AGC) to learn spatial correlations. Then, we construct a two-layer selective Mamba with dilated and causal convolution (DCCS2-Mamba) to capture temporal dependencies. Finally, considering the complex dependencies in time series, we employ a bidirectional Mamba structure for comprehensive data modeling. The model is implemented with GPU-based parallel training and inference to leverage the scalability and efficiency of HPC environments. Experiments with three real datasets show that the proposed AGBi-Mamba model is more effective than the baseline models. The overall predictive performance of our model improves by an average of 5.04%, 4.84%, and 3.34% in MAE, RMSE, and MAPE, respectively.</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

AGBi-Mamba: a traffic flow prediction model based on adaptive graph convolution and bidirectional Mamba networks

  • 期刊论文

  • 发表日期：

    2025年11月26日

查看全文
  • Guoyan Li,
  • Xue Zhang,
  • Tianying Gao,
  • Yang Cao

摘要

Accurately predicting traffic flow becomes significant as urban transportation networks become more complex and data-intensive. To address the challenge of capturing dynamic spatiotemporal dependencies under real-time requirements, we propose AGBi-Mamba, a traffic flow prediction method based on dynamic graph convolution and Mamba. Specifically, the model employs an FFT-based adaptive graph convolution (FFT-AGC) to learn spatial correlations. Then, we construct a two-layer selective Mamba with dilated and causal convolution (DCCS2-Mamba) to capture temporal dependencies. Finally, considering the complex dependencies in time series, we employ a bidirectional Mamba structure for comprehensive data modeling. The model is implemented with GPU-based parallel training and inference to leverage the scalability and efficiency of HPC environments. Experiments with three real datasets show that the proposed AGBi-Mamba model is more effective than the baseline models. The overall predictive performance of our model improves by an average of 5.04%, 4.84%, and 3.34% in MAE, RMSE, and MAPE, respectively.