面向交通流预测的双分支时空图卷积神经网络  

Double Branch Spatial-temporal Graph Convolutional Neural Network for Traffic Flow Prediction

作　　者：刘赏[1] 陈浩 陈小玉 贺娇娇 LIU Shang;CHEN Hao;CHEN Xiaoyu;HE Jiaojiao(School of Science and Technology,Tianjin University of Finance and Economics,Tianjin 300222,China)

机构地区：[1]天津财经大学理工学院,天津300222

出　　处：《信息与控制》2023年第3期391-404,416,共15页Information and Control

基　　金：教育部人文社会科学基金(19YJA630046);天津市科技计划项目(22ZLZKZF00480);天津市自然科学基金(22JCYBJC01550)。

摘　　要：针对目前交通流预测中未充分考虑到交通流量与密度、交通流量与速度之间的关联性特征信息,以及忽略多尺度时间特征的问题,提出了一种面向交通流预测的双分支时空图卷积神经网络。首先,依据交通数据的周期性,将交通数据划分为近期与周期两种时间粒度的数据输入;其次,在每个输入分支中,先通过关联性门控线性单元(AGLU)提取流量与密度、流量与速度之间的关联性特征信息;然后,通过图卷积层和多尺度时间卷积层提取关联性特征中的空间与时间上下文信息,并采用预测卷积层输出近期、周期双分支预测结果;最后,通过门控机制融合预测结果,从而实现交通流量的准确预测。实验结果表明,所提模型在交通流预测准确性与稳定性方面整体优于其他模型。Currently,traffic flow prediction does not completely consider the correlation feature information between traffic flow and density and traffic flow and speed.In addition,it also ignores multi-scale temporal features.Thus,in this study,we propose a double-branch spatial-temporal graph convolutional neural network for traffic flow prediction.First,we categorize traffic data into two temporal granularity data inputs,i.e.,near-term input and periodic input,according to the periodicity of traffic data.Second,we use an associative gated linear unit to extract the correlation feature information between traffic flow and density and traffic flow and speed in each input branch.Third,we use the graph convolutional layer and multi-scale temporal convolutional layer to extract the spatial-temporal context information of the correlation feature.We also use the prediction convolutional layer to obtain the near-term and periodic double-branch prediction results.Finally,we achieve an accurate prediction of the traffic flow by fusing the prediction results through the gating mechanism.Our experimental results show that this model is superior to other models in terms of the accuracy and stability of traffic flow prediction.

关 键 词：交通流预测 双分支时空图卷积神经网络 关联性门控线性单元 多尺度时间卷积 智能交通 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]