Journal of Frontiers of Computer Science and Technology ›› 2022, Vol. 16 ›› Issue (2): 384-394.DOI: 10.3778/j.issn.1673-9418.2009097

• Artificial Intelligence • Previous Articles     Next Articles

Two-Stream Graph Convolutional Neural Network for Dynamic Traffic Flow Fore-casting

LI Zhaoyang1, LI Lin1,+(), TAO Xiaohui2   

  1. 1. School of Computer Science and Technology, Wuhan University of Technology, Wuhan 430070, China
    2. School of Sciences, University of Southern Queensland, Toowoomba 4350, Australia
  • Received:2020-08-14 Revised:2020-10-20 Online:2022-02-01 Published:2020-11-06
  • About author:LI Zhaoyang, born in 1997, M.S. candidate. Her research interests include machine learning and data mining.
    LI Lin, born in 1977, Ph.D., professor, Ph.D. supervisor. Her research interests include data analytics, machine learning, information retrieval, Web personalization & recommendation, social media mining, natural language processing, etc.
    TAO Xiaohui, born in 1969, Ph.D., associate professor. His research interests include data analytics, machine learning, knowledge engineering, information retrieval and health informatics.
  • Supported by:
    Fundamental Research Funds for Wuhan University of Technology(205210003)

面向动态交通流预测的双流图卷积网络

李朝阳1, 李琳1,+(), 陶晓辉2   

  1. 1.武汉理工大学 计算机科学与技术学院,武汉 430070
    2.南昆士兰大学 理学院,澳大利亚 图文巴 4350
  • 通讯作者: + E-mail: cathylilin@whut.edu.cn
  • 作者简介:李朝阳(1997—),女,山东泰安人,硕士研究生,主要研究方向为机器学习、数据挖掘。
    李琳(1977—),女,湖南衡阳人,博士,教授,博士生导师,主要研究方向为数据分析、机器学习、信息检索、Web个性化推荐、社交媒体挖掘、自然语言处理等。
    陶晓辉(1969—),男,广西桂林人,博士,副教授,主要研究方向为数据分析、机器学习、知识工程、信息检索、健康信息学。
  • 基金资助:
    武汉理工大学自主创新研究基金(205210003)

Abstract:

Accurate traffic flow prediction can provide decision-making basis for traffic management departments and early warning of road conditions for drivers, which is a crucial issue in the field of transportation. In recent years, related studies have used the characteristics of graph convolution neural network (GCN) in processing non-Euclidean structure data to model the spatial correlation of traffic flow data from complex road networks. However, existing traffic flow forecasting methods based on graph convolution fail to fully consider the directionality and dynamics of spatial correlation. Considering that dynamic traffic flow presents stable spatial correlation constrained by fixed road structure and dynamic spatial correlation influenced by traffic environment changes, this paper proposes an end-to-end two-stream graph convolution network (TSGCN) for dynamic traffic flow forecasting. Firstly, real-time traffic flow data are decomposed into stable components and dynamic components with different spatial correlations. Specifically, the stable components are constrained by the physical road network and traffic habits, while the dynamic components represent the fluctuations caused by changes in traffic conditions (such as traffic congestion and bad weather). Then, the stable and dynamic spatial correlations are extracted through the two-stream graph convolution layer. Finally, this paper uses the parameterized skip connection to fuse the spatial-temporal correlations to obtain the final prediction results. Experimental results on two published real-world traffic flow da-tasets show that the proposed model is better than several popular baselines.

Key words: traffic ?ow forecasting;, graph convolution neural network (GCN), spatial-temporal correlation

摘要:

准确的交通流预测能够为管理部门提供合理的决策依据,为驾驶员提供实时的道路状况预警,是交通领域至关重要的问题。近年来,相关研究利用图卷积神经网络(GCN)处理非欧式空间结构的特点,对来自复杂路网的交通流数据进行空间相关性建模。然而,现有基于图卷积的交通流预测方法未能充分考虑空间相关性的有向性和动态性这两个重要特点。考虑到动态交通流呈现出由固定道路结构约束的稳定空间相关性和受交通环境变化影响的动态空间相关性,提出了一种用于动态交通流预测的端到端双流图卷积网络(TSGCN)。首先,将实时交通流数据分解为具有不同空间相关性的稳定分量和动态分量。其中,稳定分量表示受路网约束和交通习惯影响的部分,动态分量则代表因交通状况变化(如交通拥堵和恶劣天气)引起的波动。然后,通过双流图卷积层提取稳定和动态的空间相关性。最后,使用参数化跳过连接方法来融合时空相关性以获得最终的预测结果。在两个公开的真实交通数据集上的实验结果表明,提出的模型优于对比的交通流预测方法。

关键词: 交通流预测, 图卷积神经网络(GCN), 时空相关性

CLC Number: