基于层次时空注意力图卷积的交通速度预测方法  

Traffic speed prediction based on hierarchical spatio-temporal attention graph convolutional network

作　　者：周晓燕[1] 曹威 徐超 ZHOU Xiaoyan;CAO Wei;XU Chao(General Education Department,Fuzhou Polytechnic,Fuzhou 350108,China;College of Computer and Cyber Security,Fujian Normal University,Fuzhou 350117,China)

机构地区：[1]福州职业技术学院通识教育学院,福建福州350108 [2]福建师范大学计算机与网络空间安全学院,福建福州350117

出　　处：《广西科技大学学报》2024年第4期92-99,107,共9页Journal of Guangxi University of Science and Technology

基　　金：福建省科技厅对外合作项目(2020I0014)资助。

摘　　要：针对交通速度预测方法主要考虑实际交通路网而忽略人群活动热点区域的影响问题,提出一种层次时空图卷积网络模型(hierarchical spatio-temporal graph convolutional network,H-STGCN)。首先,通过计算路段节点之间的距离构建路网结构图,并使用聚类算法得到热点区域的聚类图;其次,利用图卷积网络层和时空注意力机制层来提取交通速度数据的时空特征;最后,引入过往邻近时间段、昨天相同时间段及上周同一时间段3个时间分量的交通速度数据分别进行训练,以挖掘交通速度在相同时间段的规律。实验结果表明,模型在预测时长为30、60和120 min时,相比于经典的时空图卷积模型,平均绝对误差分别降低了6.9%、13.6%和14.6%,实现了更准确的预测,能更好地服务于智能交通系统。In response to the problem that the traffic speed prediction method mainly considers the actual road network and neglects the effect of the hot spot areas of crowd activity,a hierarchical spatiotemporal graph convolutional network(H-STGCN)model is proposed.Firstly,a road network structure graph is structured by calculating the distances between road segment nodes.Then,a clustering graph of the hot spot areas is obtained by using clustering algorithms.Secondly,the spatio-temporal features of the traffic speed data are extracted by using graph convolutional layers and spatio-temporal attention mechanism layers.Finally,three temporal components of traffic speed data,namely,historical adjacent time periods,the same time period from the previous day,and the corresponding time period from the previous week,are introduced for training respectively,so as to uncover the patterns of traffic speed during the same time periods.The experimental results show that when the prediction time of the model is 30,60 and 120 min,the mean absolute error is reduced by 6.9%,13.6%and 14.6%,respectively,compared with the classical spatio-temporal graph convolution model,which realizes more accurate prediction and can better serve the intelligent traffic system.

关 键 词：交通速度预测 图卷积 注意力机制 时空特征 智能交通系统 

分 类 号：U491.14[交通运输工程—交通运输规划与管理] U491.2[交通运输工程—道路与铁道工程]