摘要
针对交通流预测模型中路网空间结构刻画和交通流时空特性挖掘不充分的问题,构建一种新型的有向时空图,通过定义节点相对临近度来表征路网结构关系,通过学习邻域节点对预测节点的影响权重来表征节点间时空维度的作用关系,从而能更好表达交通流的时空特性.将时空图作为预测模型的输入,采用图卷积获取交通流数据空间依赖关系,采用门控循环神经网络获取交通流数据的时空依赖关系,建立一种基于时空图卷积循环神经网络的交通流预测模型(STG-CRNN).在美国公路交通数据集上对模型预测效果进行验证,其结果表明:STG-CRNN模型的预测结果在平均绝对误差、均方根误差和平均绝对百分误差方面,均优于自回归移动平均模型、门控循环单元模型,以及扩散卷积循环神经网络模型.
Aiming at the problem of insufficient characterization of road network spatial structure and inadequate mining of traffic flow spatiotemporal characteristics in the traffic flow prediction model,a new type of directed spatiotemporal graph is constructed.In which the spatial relationship between nodes are characterized by defining the relative proximity,and the spatiotemporal relationship between nodes are characterized by learning the influence weights of neighborhood nodes on the prediction node,so as to better express the temporal and spatial characteristics of traffic flow.Taking spatiotemporal graphs as the input of the prediction model,graph convolution is used to obtain the spatial dependence of traffic flow data,and the gated recurrent neural network is used to obtain the spatiotemporal dependence of traffic flow data,and a traffic flow based on the spatiotemporal graph convolution recurrent neural network(STG-CRNN)is established.The model prediction effect is verified on the U.S.highway traffic data set,and the results show that the STG-CRNN model is better than the autoregressive moving average model,the gated recurrent unit model,and the diffusion convolutional recurrent neural network model in terms of the mean absolute error,root mean square error,and mean absolute percentage error.
作者
谷振宇
陈聪
郑家佳
孙棣华
GU Zhen-yu;CHEN Cong;ZHENG Jia-jia;SUN Di-hua(School of Automation,Chongqing University,Chongqing 400044,China;School of Business Administration,Chongqing City Management College,Chongqing 401331,China)
出处
《控制与决策》
EI
CSCD
北大核心
2022年第3期645-653,共9页
Control and Decision
基金
国家自然科学基金项目(62073049)
重庆市教委科学技术研究项目(KJQN202003303)。
关键词
交通流预测
有向时空图
相对临近度
时空依赖性
图卷积网络
循环神经网络
traffic flow prediction
directed spatiotemporal graph
relative proximity
temporal and spatial dependencies
graph convolutional network
recurrent neural network
