欢迎访问《汽车安全与节能学报》，
汽车安全与节能学报

汽车安全与节能学报 ›› 2025, Vol. 16 ›› Issue (3): 463-469.DOI: 10.3969/j.issn.1674-8484.2025.03.012

• 智能驾驶与智慧交通 • 上一篇    下一篇

基于多维注意力机制的高速公路交通流量预测方法

虞安军1(), 励英迪2, 杨哲懿2, 付崇宇3, 童蔚苹2, 余佳2,*(), 刘云海4, 刘志远2   

  1. 1.江西赣粤高速公路股份有限公司，南昌 330029，中国
    2.东南大学 交通学院，南京 211189，中国
    3.东南大学 网络空间安全学院，南京 211189，中国
    4.苏州航礼交通科技有限公司，苏州 215024，中国
  • 收稿日期:2024-09-10 修回日期:2025-01-12 出版日期:2025-06-30 发布日期:2025-07-01
  • 通讯作者: 余佳，讲师。E-mail：yujiazijin@163.com。
  • 作者简介:虞安军（1977—），男（汉），湖北，高级工程师。E-mail：7425212@qq.com。
  • 基金资助:
    交通运输部行业重点科技项目(2022-ZD6-075);江苏省自然科学基金攀登项目(BK20232019);江苏省应用数学科学研究中心(BK20233002)

Highway traffic flow prediction approach based on multi-dimensional attention mechanism

YU Anjun1(), LI Yingdi2, YANG Zheyi2, FU Chongyu3, TONG Weiping2, YU Jia2,*(), LIU Yunhai4, LIU Zhiyuan2   

  1. 1. Jiangxi Ganyue Expressway Co., Ltd., Nanchang 330029, China
    2. School of Transportation, Southeast University, Nanjing 211189, China
    3. School of Cyber Science and Engineering, Southeast University, Nanjing 211189, China
    4. Suzhou Hangli Transportation Technology Co., Ltd., Suzhou 215024, China
  • Received:2024-09-10 Revised:2025-01-12 Online:2025-06-30 Published:2025-07-01

RichHTML

4

PDF

83

摘要：

为了实现精准的交通流量预测，提高高速公路智慧管理水平，该文构建了一种基于多维注意力机制的交通流量预测模型，并在樟吉高速公路真实交通数据集上开展对比实验，以验证模型的准确性及预测精度。模型基于图神经网络（GNN）和时间卷积网络（TCN）提取交通流空间和时间维度的特征，结合多维注意力机制挖掘时空数据中的关键信息，同时引入多任务学习架构，通过基于同方差不确定性的损失函数来平衡不同任务共同学习，以提高模型的泛化能力和鲁棒性。 结果表明：该模型在测试集上的均方根误差（RMSE）和平均绝对误差（MAE）分别为7.467和5.133，相较基准模型有更好的预测精度；提出的该交通流量预测方法可有效地挖掘交通流的时空特性，描述真实交通运行状态，对高速公路交通流量做出精准预测。

关键词: 交通流预测, 图神经网络（GNN）, 时间卷积网络（TCN）, 多维注意力机制

Abstract:

A traffic flow prediction model based on a multi-dimensional attention mechanism was proposed to achieve precise traffic flow prediction and enhance the intelligent management level of expressways. Comparative experiments were conducted on real traffic datasets from the Zhangji Expressway to verify the accuracy and predictive accuracy of the model. The model extracted spatial and temporal features of traffic flows using graph neural networks (GNN) and temporal convolutional networks (TCN), respectively. It integrated a multi-dimensional attention mechanism to mine key information within spatiotemporal data. Additionally, a multi-task learning architecture was introduced, employing a loss function based on homoscedastic uncertainty to balance the joint learning of different tasks, thereby enhancing the generalization ability and robustness of the model. The results show that the root mean square error (RMSE) and mean absolute error (MAE) of the model on the test set are 7.467 and 5.133, respectively, demonstrating superior predictive accuracy compared to baseline models. The proposed prediction method can effectively uncover the spatiotemporal characteristics of traffic flows, describe the actual state of traffic operations, and make accurate predictions of the traffic flow on expressways.

Key words: traffic flow prediction, graph neural network (GNN), temporal convolutional network (TCN), multi-dimensional attention mechanism

中图分类号: