全局和局部感知的交通速度预测模型  

Global and Local Information Aware Traffic Speed Prediction

作　　者：申岩松 李琳[1] 黄传明 SHEN Yan-song;LI Lin;HUANG Chuan-ming(School of Computer Science and Artificial Intelligence,Wuhan University of Technology,Wuhan,Hubei 430070,China;Wuhan Traffic Management Bureau,Wuhan,Hubei 430030,China)

机构地区：[1]武汉理工大学计算机与人工智能学院,湖北武汉430070 [2]武汉市交管局,湖北武汉430030

出　　处：《电子学报》2024年第9期3195-3205,共11页Acta Electronica Sinica

基　　金：国家自然科学基金(No.62276196);湖北省重点研发计划项目(No.2021BAA030)。

摘　　要：面对日益严峻的交通堵塞问题,智能交通系统获得飞速发展和广泛应用,作为基石工作的交通速度预测因此备受关注.近些年来,深度学习被广泛用于交通速度预测的研究工作,并且研究方向也从单一的建模时间相关性迁移到复杂的时空相关性,图神经网络由于契合交通路网的图结构数据这一本质属性,成为建模空间相关性的主流方法 .目前,大多数的研究工作已经注意到动态的空间相关性对交通速度预测任务的重要性.然而,基于这一发现所提出的建模思路主要预定义矩阵或自适应矩阵,属于静态矩阵,并不足以应对空间相关性的复杂和动态的特性.同时通过对真实交通速度数据集的分析,本文发现交通节点间依赖的局部波动相比交通路网的全局影响具有更强的动态性,这表明空间相关性可以从全局和局部的角度分开建模,因此本文提出了一个端到端全局和局部融合的动态图神经网络模型来进行交通速度预测.首先,交通速度流被自分解层分解为静态分量和动态分量,随后动态图生成模块为动态分量构造实时的动态图以匹配其动态性.基于构造的动态图和输入的预定义图,本文借助图卷积操作来学习这两类空间相关性的高阶表达.除此之外,本文在时间模块使用空洞因果卷积捕获交通数据中时间相关性.最后,残差连接被用来聚合时空相关性并输送给输出层完成最终的速度预测.在两个高速公路数据集和一个城市路数据集上的实验结果表明本文提出的模型相比主流模型在平均绝对误差和均方根误差两个预测指标上均优于主流模型.Facing the increasingly severe traffic congestion problem,the intelligent transportation system has been rapidly developed and widely used,and the traffic speed prediction,a cornerstone task,has attracted much attention.In re⁃cent years,deep learning has been widely used in the research of traffic speed prediction,and the research direction has also shifted from modeling time correlation to considering complex spatiotemporal correlation.The graph neural network fits the graph structure data of the traffic network and has become the mainstream method for modeling spatial correlation.To date,most research works have noted the importance of modeling dynamic spatial correlations in the task of traffic speed predic⁃tion.However,predefined or adaptive matrices for spatial feature learning are essentially static,and are not sufficient to match the complex and dynamic characteristics of spatial correlations.Moreover,through the analysis of multiple real traf⁃fic speed datasets,we find that the local fluctuations of inter-node dependencies are more dynamic than the global influence of the traffic network,which indicates that the spatial correlation can be derived from the global and local angles.Therefore,we propose an end-to-end global and local aware dynamic graph neural network model for traffic speed prediction.The traf⁃fic speed flow is first decomposed into static components and dynamic components by the self-decomposition layer,and then the dynamic graph generation module constructs a real-time dynamic graph for the dynamic components to match their dynamics.With the constructed dynamic graph and the input predefined graph,we model higher-order representations of these two classes of spatial correlations through graph convolution operations.Besides,we use causal convolution in the temporal module to capture temporal correlations in traffic data.Finally,residual connections are used to aggregate spatiotemporal correlations and feed to the output layer for final speed prediction. Experimental results on two h

关 键 词：交通速度预测 时空相关性 动态图神经网络 图生成 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]