基于时序双分支图卷积网络的交通流量预测  

Traffic Flow Prediction Based on Bi-Path Graph Convolutional Networks

作　　者：昝欣 刘煜明 刘辰昀 ZAN Xin;LIU Yuming;LIU Chenyun(Antai College of Economics&Management,Shanghai Jiao Tong University,Shanghai 200030,China;Shanghai International Port(Group)Co.,Ltd.,Shanghai 201201,China;Shanghai Big Data Center,Shanghai,200433,China)

机构地区：[1]上海交通大学安泰经济与管理学院,上海200030 [2]上海国际港务(集团)股份有限公司,上海201201 [3]上海市大数据中心,上海200433

出　　处：《信息系统学报》2024年第1期114-131,共18页China Journal of Information Systems

基　　金：2022年(第二批)上海市城市数字化转型专项资金(上港集团港口物流大数据中心项目,编号:202202013)。

摘　　要：米精准的交通流量预测是现代智能交通系统稳定运行的关键。在过去的十几年中,交通流量数据呈爆炸式增长,智能交通步人大数据时代。交通流量预测受空间道路拓扑结构和时序流量运动模式等多种因素影响,如何捕获观测节点间的复杂时空相关性,如何利用合理的先验知识辅助建模海量数据,提升在真实场景下的预测精度,成为交通流量预测的重点。现有交通流量预测方法主要使用基于隐式特征提取的设计,在综合交通时空信息,捕获节点时空依赖关系上仍存在模型能力弱,可解释性差等不足。本文提出了一种时序双分支图卷积网络(Bi-Path graph convolutional network,简称Bi-PathGCN),设计了基于时序特征提取的双分支图卷积网络。双分支图卷积网络首先利用时序特征解耦模块将时序特征拆分成低频分量和高频分量。不同分量输入由两个并行的时空特征提取模块提取有判别力的特征,并利用时空特征融合模块进行自适应加权融合。该设计显式利用了时序先验信息,提升了模型对复杂时空结构信息的捕获能力。本文还使用两个真实场景下的交通数据集评估了该方法的预测性能。实验表明,本文提出的双分支图卷积网络结构,可以有效利用不同时序分量的互补信息,简化模型的学习难度,在对比模型中取得了更好的表现,体现了本方法在精准预测交通流量方面的潜力和优越性。Accurate traffic flow prediction is the key to the stable operation of modern intelligent transportation systems.In the past decade,traffic flow data has been growing explosively,and intelligent transportation has stepped into the era of big data.Traffic flow prediction is influenced by various factors such as spatial road topology and temporal movement patterns,so how to capture the complex spatio-temporal correlation between observed nodes,enhance the prediction accuracy of the model by using proper a priori knowledge to assist in modeling massive data have become the focus of traffic flow prediction.Existing traffic flow prediction methods mainly use prediction models based on implicit feature extraction,which are still inadequate in integrating traffic spatio-temporal information and capturing spatio-temporal dependencies of nodes.This paper proposes a temporal bipartite graph convolutional network(Bi-Path GCN),and designs a bipartite graph convolutional network based on temporal feature extraction.It first splits the temporal features into low-frequency and high-frequency components using the temporal feature decoupling module.Two parallel spatio-temporal feature extraction modules extract the different component inputs with discriminative power,and adaptive weighted fusion is performed using the temporal feature fusion module.Such a design explicitly utilizes the temporal decomposition a priori information and enhances the model's ability to capture complex spatio-temporal structure information.This paper also evaluates the prediction performance of the method using two real traffic datasets.Experiments show that the two-branch graph convolutional network structure proposed in this paper can effectively utilize the complementary information of the frequency temporal components,simplify the learning difficulty of the model,and achieve better performance in the comparison models,reflecting the potential and superiority performance of proposed method in accurately predictingtraffic flow.

关 键 词：图卷积网络 时序双分支结构 时序分解 自适应加权融合 

分 类 号：U495[交通运输工程—交通运输规划与管理]