跳到主要导航 跳到搜索 跳到主要内容

CDGNet: A Cross-Time Dynamic Graph-Based Deep Learning Model for Vehicle-Based Traffic Speed Forecasting

  • Yuchen Fang
  • , Haiyong Luo*
  • , Fang Zhao*
  • , Poly Z.H. Sun
  • , Yanjun Qin
  • , Liang Zeng
  • , Bo Hui
  • , Chenxing Wang
  • *此作品的通讯作者
  • Beijing University of Posts and Telecommunications
  • CAS - Institute of Computing Technology
  • Shanghai Jiao Tong University
  • Tsinghua University
  • University of Tulsa

科研成果: 期刊稿件 › 文章 › 同行评审

摘要

Vehicle-based traffic speed forecasting aims to predict the average speed of vehicles on the road in the future, which is an essential side information in intelligent vehicles and beneficial to safe autonomous driving, yet is very challenging due to the complex and dynamic spatio-temporal dependencies in real-world vehicle-based traffic. To extract intricate correlations among multiple vehicle-based speed time series, previous methods have used graph convolution networks. However, the conventional static and dynamic graphs fail to reflect the traffic evolution and the hysteresis spatial influence caused by vehicle movement. To address this issue, we propose a novel cross-time dynamic graph-based deep learning model, named CDGNet, for vehicle-based traffic speed forecasting. The model is able to effectively capture hysteresis spatial dependencies between each time slice and its historical time slices through the cross-time dynamic graph-based GCN. Meanwhile, a gating mechanism is integrated into our cross-time dynamic graph, which conforms to the sparse correlation in the real world. Besides, GCNs are incorporated into a novel encoder-decoder architecture to forecast multi-step speed. Experimental results on three real-world vehicle-based traffic speed datasets demonstrate the superiority of our CDGNet over various state-of-the-art spatio-temporal forecasting methods and the effectiveness of each component. We additionally provide a visualization of our cross-time dynamic graph to show the capability of assisting intelligent vehicles to avoid congestion.

源语言英语
页(从-至)683-696
页数14
期刊IEEE Transactions on Intelligent Vehicles
卷10
期1
DOI
出版状态已出版 - 2025
已对外发布是

学术指纹

探究 'CDGNet: A Cross-Time Dynamic Graph-Based Deep Learning Model for Vehicle-Based Traffic Speed Forecasting' 的科研主题。它们共同构成独一无二的学术指纹。

引用此