中文核心期刊
CSCD来源期刊
中国科技核心期刊
RCCSE中国核心学术期刊
重庆交通大学学报（自然科学版）

重庆交通大学学报（自然科学版） ›› 2025, Vol. 44 ›› Issue (10): 51-59.DOI: 10.3969/j.issn.1674-0696.2025.10.07

• 交通+大数据人工智能 • 上一篇    

基于多时空特征和图注意力网络的交通流预测模型研究

朱政泽，熊宇恒   

  1. (湖北汽车工业学院 智能网联汽车学院，湖北 十堰 442000)
  • 收稿日期:2024-10-09 修回日期:2025-02-10 发布日期:2025-11-06
  • 作者简介:朱政泽(1988—)，男，湖北十堰人，讲师，博士，主要从事智慧交通车辆决策控制方面的研究。E-mail：zhengze.zhu@huat.edu.cn 通信作者：熊宇恒(2000—)，男，湖北天门人，硕士，主要从事智慧交通方面的研究。E-mail：202211224@huat.edu.cn
  • 基金资助:
    湖北省自然科学基金计划项目(2023AFB481)；湖北汽车工业学院博士科研启动基金项目(BK202307)

Traffic Flow Prediction Model Based on Multi-spatiotemporal Features and Graph Attention Network

ZHU Zhengze，XIONG Yuheng   

  1. (Institute of Intelligent Networked Vehicle, Hubei University of Automotive Technology, Shiyan 442000, Hubei, China)
  • Received:2024-10-09 Revised:2025-02-10 Published:2025-11-06

摘要： 准确预测未来的交通流量是解决城市交通拥堵问题的关键。针对交通流预测中时间周期性特征提取不充分，空间特征挖掘不全面等问题，提出一种基于多时空特征与图注意力网络的交通流预测模型（multi-spatiotemporal features and graph attention network, MSTFGAN）。首先按照小时周期、日周期和周周期构建3种不同时段的周期性数据，作为模型的输入，接着使用带门控机制的时间卷积网络（temporal convolutional network, TCN）分别提取3种周期性数据的时间特征；其次，构建一种自适应邻接矩阵，通过运用可训练参数来学习交通传感器节点间的关联性，进而揭示交通网络中的潜在空间特性；然后，利用图注意力网络（graph attention network, GAT）为每个路网节点动态分配不同的权重，捕获交通传感器节点之间动态的空间相关性；继而，设计一个门控融合模块，自适应地融合隐式空间特征和动态空间特征；最后，基于注意力机制融合时间和空间特征，以准确地进行交通流预测。在两个真实世界高速公路数据集PeMS04和PeMS08上进行了实验验证。研究结果表明：与效果最好的基线模型DSTAGNN相比，MSTFGAN的平均绝对误差（EMA）降低了3.802%，均方根误差（ERMS）降低了3.780%，平均绝对百分比误差（EMAP）降低了3.356%，说明MSTFGAN能有效提高交通流预测的精度，准确地预测交通流趋势。

关键词: 交通工程；交通流预测；多时空特征；图卷积网络；图注意力

Abstract: The accurate prediction of future traffic flow is the key of solving the urban traffic congestion problem. Aiming at the problems of insufficient time-periodic feature extraction and incomplete spatial feature mining in traffic flow prediction, a traffic flow prediction model based on multi-spatiotemporal features and graph attention network (MSTFGAN) was proposed. Firstly, three kinds of periodic data with different time periods were constructed according to the hourly, daily and weekly cycles as the input of the model, and then the temporal features of the three kinds of periodic data were extracted by using temporal convolutional network (TCN) with a gating mechanism. Secondly, an adaptive neighbor matrix was constructed to learn the correlation between the traffic sensor nodes by applying the trainable parameter, thereby revealing the potential spatial characteristics in the traffic network. Then, a graph attention network (GAT) was used to assign different weights to each road network node dynamically to capture the dynamic spatial correlation between the traffic sensor nodes; furthermore, a gated fusion module was designed to adaptively fuse implicit spatial features and dynamic spatial features. Finally, temporal and spatial features were fused based on an attention mechanism for accurate traffic flow prediction. Experimental validation was performed on two real-world highway datasets, that is, PeMS04 and PeMS08. The research results show that compared with the best-performing baseline model DSTAGNN, MSTFGAN reduces the mean absolute error (EMA) by 3.802% , the root mean square error (ERMS) by 3.780%, and the mean absolute percentage error (EMAP) by 3.356%, which indicates that MSTFGAN can effectively improve the accuracy of the traffic flow prediction and accurately predict the traffic flow trend.

Key words: traffic engineering; traffic flow forecasting; multi-spatiotemporal features; graph convolutional network; graph attention

中图分类号: