《计算机应用》唯一官方网站 ›› 2026, Vol. 46 ›› Issue (4): 1323-1333.DOI: 10.11772/j.issn.1001-9081.2025040522

• 前沿与综合应用 • 上一篇    

基于双层多尺度动态GCN模型的城市交通流量预测

李文浩, 郭银章()   

  1. 群智计算与云计算实验室(太原科技大学),太原 030024
  • 收稿日期:2025-05-14 修回日期:2025-08-04 接受日期:2025-08-07 发布日期:2025-08-22 出版日期:2026-04-10
  • 通讯作者: 郭银章
  • 作者简介:李文浩(2001—),女,河南南阳人,硕士研究生,主要研究方向:深度学习、交通流预测
  • 基金资助:
    山西省重点实验室开放课题(CICIP2023001)

Urban traffic flow prediction based on dual-layer multi-scale dynamic graph convolutional network model

Wenhao LI, Yinzhang GUO()   

  1. Swarm Intelligence Computing and Cloud Computing Laboratory (Taiyuan University of Science and Technology),Taiyuan Shanxi 030024,China
  • Received:2025-05-14 Revised:2025-08-04 Accepted:2025-08-07 Online:2025-08-22 Published:2026-04-10
  • Contact: Yinzhang GUO
  • About author:LI Wenhao, born in 2001, M. S. candidate. Her research interests include deep learning, traffic flow prediction.
  • Supported by:
    Open Project of Shanxi Province Key Laboratory(CICIP2023001)

摘要:

针对现有交通流量预测模型不能有效利用交通区域层特征与节点层特征的融合信息,以及缺乏对时空特征的动态表达的问题,提出一种用于城市交通预测的双层多尺度动态图卷积网络(DM-DGCN)模型。首先,采用双层网络架构融合区域与节点的时空特征,同时处理节点和区域的交通流量数据;其次,在空间维度上,构建空间动态图卷积网络(S-GCN)模块提取动态空间相关性;再次,在时间维度上,为了捕捉不同语义环境下的潜在时间依赖关系,设计多尺度时间卷积网络(MSTCN)模块,同时,为了将空间关系学习思想引入时间域,设计时间动态图卷积网络(T-GCN)模块,构建随时间动态变化的时间关系矩阵;最后,设计基于注意力机制的动态融合模块,整合区域层特征与节点层特征,通过融合层生成最终预测结果。在济南交通数据集和西安交通数据集上的实验结果表明,在60 min预测任务中:DM-DGCN模型在济南数据集上的平均绝对误差(MAE)和均方根误差(RMSE)相较于时空关键图神经网络(STPGNN)模型分别降低了5.79%和5.56%;在西安数据集上,相较于分层时空图常微分方程(HSTGODE)模型,所提模型在MAE和RMSE上分别降低了3.73%和4.19%。以上验证了DM-DGCN模型优于现有的基线模型,能有效挖掘交通数据中的动态多尺度时空依赖关系,并准确预测未来的交通流量。

关键词: 交通流量预测, 时空特征融合, 图卷积网络, 动态时空图, 多尺度时间依赖

Abstract:

To address the limitations of the existing traffic flow prediction models in utilizing fusion information of traffic regions and node layer features effectively, as well as these models’lack of dynamic representation of spatio-temporal features, a Double-layer Multi-scale Dynamic Graph Convolutional Network (DM-DGCN) was proposed for urban traffic prediction. Firstly, a double-layer network architecture was adopted to fuse the spatio-temporal features of regions and nodes, so as to handle node and region traffic flow data simultaneously. Secondly, in the spatial dimension, a Spatial-dynamic Graph Convolutional Network (S-GCN) module was constructed to capture dynamic spatial correlations. Thirdly, in the temporal dimension, a Multi-Scale Temporal Convolutional Network (MSTCN) module was designed to capture potential temporal dependencies under different semantic environments. At the same time, the idea of spatial relationship learning was introduced into the temporal domain by designing a Temporal-dynamic Graph Convolutional Network (T-GCN) module, so as to construct a dynamic time-varying temporal relationship matrix. Finally, a dynamic fusion module based on the attention mechanism was designed to integrate region and node layer features, and the final prediction results were generated through the fusion layer. Experimental results on the Jinan and Xi’an traffic datasets show that for the 60-minute prediction task, DM-DGCN model reduces the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) by 5.79% and 5.56%, respectively, compared to Spatio-Temporal Pivotal Graph Neural Network (STPGNN) model on the Jinan dataset; and by 3.73% and 4.19%, respectively, compared to Hierarchical Spatio-Temporal Graph Ordinary Differential Equation network (HSTGODE) model on the Xi’an dataset. The above verifies that DM-DGCN model outperforms the existing baseline models, captures dynamic multi-scale spatio-temporal dependencies in traffic data effectively, and predicts future traffic flow accurately.

Key words: traffic flow prediction, spatio-temporal feature fusion, Graph Convolutional Network (GCN), dynamic spatio-temporal graph, multi-scale time dependency

中图分类号: