基于动态时空特征周期性编码的公路交通流量预测模型  

Highway traffic flow prediction model based on periodic encoding with dynamic spatio-temporal features

作　　者：张韫 张波 王剑楠 申田 刘明 ZHANG Yun;ZHANG Bo;WANG Jiannan;SHEN Tian;LIU Ming(Shaanxi Transportation Holding Group Co.,Ltd.,Xi’an,Shaanxi 710000,China;Operation Management Branch of Shaanxi Transportation Holding Group Co.,Ltd.,Xi’an,Shaanxi 710000,China;Shaanxi Transportation Holding Supply Chain Management Group Co.,Ltd.,Xi’an,Shaanxi 710000,China;School of Materials Science and Engineering,Xi’an University of Technology,Xi’an,Shaanxi 710048,China)

机构地区：[1]陕西交通控股集团有限公司,陕西西安710000 [2]陕西交通控股集团有限公司运营管理分公司,陕西西安710000 [3]陕西交控供应链管理集团有限公司,陕西西安710000 [4]西安理工大学材料科学与工程学院,陕西西安710048

出　　处：《公路交通科技》2025年第3期21-33,共13页Journal of Highway and Transportation Research and Development

基　　金：2023年陕西省交通运输厅科研项目(23-90X)。

摘　　要：【目标】针对现有的基于图卷积神经网络(GCN)的交通流预测模型在提取路网节点动态时空相关性和长距离时间依赖性方面存在的不足,提出了一种基于动态时空特征周期性编码的公路交通流量预测模型。【方法】首先,根据交通流数据在不同周期尺度下的动态相关性,设计周期性交通流数据嵌入模块,捕捉隐藏在交通流数据中的多尺度时间依赖特性。然后,为了克服传统GCN模型使用预定义的邻接矩阵聚合节点空域特征,忽略路网异质性的问题,构建基于动态拓扑编码的空域特征提取模块,利用近邻空域拓扑编码和相似空域拓扑编码,融合注意力机制,捕获路网数据动态空间的相关性。其次,为了克服传统GCN仅使用一维卷积提取时域特征造成长距离节点时域关系提取受限的问题,构建基于注意力机制约束的时域特征编码模块,通过时域卷积网络模块,时域相关性模块和关键帧相关性模块,提取不同时间尺度下的时间依赖关系。最后,通过自适应周期时空特征融合模块进一步融合节点的时空特征。在PEMS04,PEMS07,PEMS08这3个真实交通流数据上将本研究所提出方法与其他9种主流的交通流预测模型进行对比分析。【结论】所提出模型在3个数据集上的预测精度指标均优于对比算法,具有较好的预测精度。[Objective]To address the limitations of current graph convolutional neural network(GCN)based traffic flow prediction models in capturing dynamic spatio-temporal correlations and long-term temporal dependencies of road network nodes,a highway traffic flow prediction model based on periodic encoding with dynamic spatio-temporal features was proposed.[Method]First,a periodic traffic flow data embedding module was designed to capture the multi-scale temporal dependencies inherent in traffic flow data,according to the dynamic correlation of traffic flow data with various cycle scales.Then,to mitigate the issues of pre-defined adjacency matrices for aggregating node spatial features while neglecting road network heterogeneity by using traditional GCN models,a spatial feature extraction module based on dynamic topological encoding was constructed.This module employed the neighboring spatial topological encoding and similar spatial topological encoding,integrating an attention mechanism,to capture the dynamic spatial correlations within road network data.Furthermore,to overcome the problem that the traditional GCN uses only one-dimensional convolution to extract time-domain features,resulting in limited extraction of long-distance node time series,a temporal feature encoding module based on attention mechanism was constructed.Through the temporal convolution network module,correlation feature extraction module,and key frame temporal feature extraction module,the temporal dependency relations with various temporal scales were extracted.Finally,the adaptive periodic spatio-temporal feature fusion module was implemented to further integrate the spatio-temporal features of nodes.The proposed method was evaluated against other 9 state-of-the-art traffic flow prediction models on 3 real traffic flow benchmark datasets,i.e.,PEMS04,PEMS07,and PEMS08.[Conclusion]Compared with competing methods,the proposed model achieves the superior prediction accuracy.It is with higher prediction accuracy.

关 键 词：智能交通 交通流预测 动态时空特征编码 动态拓扑编码 时域特征编码 

分 类 号：U491[交通运输工程—交通运输规划与管理]