基于门控时空图网络和TCN的交通流预测方法  

Traffic Flow Forecasting Method Based on Gated Spatial-temporalSpatiotemporal Graph Network and TCN

作　　者：黄河 谢军义 李志晖 孙霞[2] 彭挺[3] HUANG He;XIE Junyi;LI Zhihui;SUN Xia;PENG Ting(China Railway Seventh Bureau Group Third Engineering Co.,Ltd.,Xi′an 710032,China;School of Information Science and Technology,Northwest University,Xi′an 710127,China;Key Laboratory for Special Area Highway Engineering of Ministry of Education,Chang′an University,Xi′an 710064,China)

机构地区：[1]中铁七局集团第三工程有限公司,西安710032 [2]西北大学信息科学与技术学院,西安710127 [3]长安大学特殊地区公路工程教育部重点实验室,西安710064

出　　处：《交通科技》2024年第6期126-131,共6页Transportation Science & Technology

基　　金：陕西省重点研发计划项目(2021KW-63)资助。

摘　　要：当前交通流预测研究普遍采用GCN学习空间图结构,但缺乏保留图中重要节点特征的能力,且忽略时间序列之间的长距离依赖关系。针对上述问题,提出一种结合门控时空图神经网络和TCN的交通流预测方法。首先该模型采用门控图神经网络GGNN学习空间图结构并保留关键节点特征信息,然后利用TCN捕获时间序列之间的长距离依赖关系。在PeMSD04和PeMSD082种公开的交通流数据集上进行对比实验、消融实验和超参数实验。实验表明,GGNN-TCN模型在MAE、RMSE和MAPE 3种指标上的整体性能明显优于基线模型,消融实验结果验证GGNN和TCN组件有利于提升模型整体性能,参数实验表明当GGNN层数为2模型整体性能最优。Current research on traffic flow prediction commonly employs graph convolutional networks(GCNs)to learn spatial graph structure.However,these approaches often lack the ability to retain important node features within the graph and overlook long-distance dependencies between time series.To address these issues,a traffic flow prediction method that combines gated spatiotemporal graph neural network and temporal convolutional networks(TCNs)was proposed.First,the gated graph neural network(GGNN)was used in the model to learn the spatial graph structure while preserving key node feature information.Then,TCN was employed to capture long-distance dependencies between time series.Finally,comparison study,ablation study,and hyperparameter study were conducted on the publicly available PeMSD04 and PeMSD08 traffic flow datasets.Experimental results show that the GGNN-TCN model significantly outperforms baseline models in terms of MAE,RMSE,and MAPE.The ablation study confirms that both the GGNN and TCN components contribute positively to the overall model performance,while parameter study indicate that the model achieves optimal performance when the number of GGNN layers is set to 2.

关 键 词：空间图结构 长距离依赖 交通流预测 时间序列 

分 类 号：U491.112[交通运输工程—交通运输规划与管理]