基于改进的时空图卷积网络的交通流预测模型研究  

Research on Traffic Flow Prediction Model Based on Improved Spatial-Temporal Graph Convolutional Networks

作　　者：庄伟卿 曹勇博 ZHUANG Weiqing;CAO Yongbo(School of Internet Economics and Business,Fujian University of Technology,Fuzhou 350014;Shaanxi Provincial Transportation Planning and Design Research Institute Co.,Ltd,Xi'an 710065)

机构地区：[1]福建理工大学互联网经贸学院,福州350014 [2]陕西省交通规划设计研究院有限公司,西安710065

出　　处：《系统科学与数学》2024年第1期217-235,共19页Journal of Systems Science and Mathematical Sciences

基　　金：国家社科基金一般项目(22BGL007)资助课题。

摘　　要：智能交通系统(ITS)作为协调城市交通的核心正快速发展,而交通流量预测是智能交通系统的重要组成部分,被视为成功部署ITS的关键因素.由于交通网络的复杂空间拓扑结构,使交通流量表现出高阶非线性及动态时空复杂性.为更好地对交通网格数据进行预测,文章提出了一种新的时空网络模型DCSTGCN,具备以下特点:1)将切比雪夫多项式(Ch)应用于图卷积神经网络,结合自扩散卷积将传统的固定式交通拓扑结构进行转换,使其更具有随机性、动态性;2)加入空间Transformer模型,在考虑数据异质性问题的同时,利用多头自注意力机制考虑节点、本地邻居节点以及非本地节点的多属性问题,从高维的子空间进行考虑节点之间的隐藏特征信息;3)将时间Transformer与1×1的二维卷积神经网络(Conv2d)相结合,对交通流时间序列信息进行多重权重分配获取重要的时间特征,利用Conv2d网络进行预测输出.通过试验验证,表明该方法模型优于多种对比基线模型.As the core of coordinating urban traffic,intelligent transportation system(ITS)is developing rapidly,as well as,the traffic flow prediction is an important part of ITS,which is regarded as the key factor for successful deployment of ITS.Because of the complex spatial topological structure of the traffic network,the traffic flow shows higher-order nonlinearity and dynamic spatial-temporal complexity.In order to better predict the traffic grid data,this paper proposes a new spatial-temporal network model DCSTGCN,which has the following characteristics:1)The Chebyshev polynomial(Ch)is applied to the graph convolutional neural network,and the traditional fixed traffic topology is converted with self-diffused convolution to make it more random and dynamic;2)The spatial Transformer model is added.While considering the data heterogeneity,the multi head self attention mechanism is used to consider the multi attribute problems of nodes,local neighbors,and non local nodes,and the hidden feature information between nodes is considered from the high-dimensional subspace;3)Combining the temporal transformer with a 1×12D convolutional neural network(Conv2d).Multiple weights are assigned to the traffic flow time series information to obtain important time features,and the Conv2d network is used to predict the output.The experimental verification shows that the method model is better than a variety of comparative baseline models.

关 键 词：交通流预测 切比雪夫多项式 图卷积 TRANSFORMER 随机游走 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] U491[自动化与计算机技术—控制科学与工程]