交通流量预测的时间异质性图注意力网络  

Time heterogeneous graph attention network for traffic flow prediction

作　　者：陈雷 赵耀帅[3,4] 林彦 郭晟楠 万怀宇 林友芳[1,2] CHEN Lei;ZHAO Yaoshuai;LIN Yan;GUO Shengnan;WAN Huaiyu;LIN Youfang(School of Computer and Information Technology,Beijing Jiaotong University,Beijing 100044,China;Beijing Key Lab of Traffic Data Analysis and Mining,Beijing Jiaotong University,Beijing 100044,China;Key Laboratory of Intelligent Passenger Service of Civil Aviation,Beijing 101318,China;Travelsky Technology Limited,Beijing 101318,China)

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044 [2]北京交通大学交通数据分析与挖掘北京市重点实验室,北京100044 [3]民航旅客服务智能化应用技术重点实验室,北京101318 [4]中国民航信息网络股份有限公司,北京101318

出　　处：《山东大学学报（工学版）》2023年第5期29-36,共8页Journal of Shandong University（Engineering Science）

基　　金：中央高校基本科研业务费项目(2019JBM024)。

摘　　要：采用注意力模型研究交通流量预测问题,提出并设计一种基于时间异质性结合噪声滤除的交通流量预测方法,有效预测美国加州高速公路未来1 h的交通流量。在构建预测方案过程中,分析交通流量数据特性,分别针对相对时间间隔和绝对时间进行建模,挖掘时间异质性;使用基于节点固有属性的动态噪声滤除方法,解决空间中噪声干扰问题;对预测模型的工作性能和结果进行详细分析,并结合基线模型进行对比评价。试验结果表明,挖掘时间异质性并动态滤除噪声的改进注意力机制预测模型具有一定的预测精度。Towarding the traffic flow prediction research problem,attention model was employed to propose and design a traffic flow prediction method based on time heterogeneity combined with noise filter,which could effectively predict the traffic flow in the upcoming 1 h within California freeway.During constructing the prediction solution,the characteristics of traffic flow data were analyzed,and the relative time interval and absolute time were respectively modeled to explore time heterogeneity.A dynamic noise filtering method based on the inherent properties of nodes was proposed to solve the problem of noise interference in space.The performance and results of the prediction model were analyzed in detail and compared with the baseline model.The experimental results showed that the improved attentional mechanism prediction model with time heterogeneity and dynamic noise filter had higher prediction accuracy.

关 键 词：交通流预测 时空依赖性 注意力网络 图神经网络 门控机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]