基于STAtt-DGCN模型的高速公路短时交通流预测  

Short-term Traffic Flow Forecast of Expressway Based on STAtt-DGCN Model

作　　者：唐嘉立 舒宏柯 黄小峰 陈梦宇 Tang Jiali;Shu Hongke;Huang Xiaofeng;Chen Mengyu(Jiangxi Traffic Monitoring and Command Center,Nanchang 330036,China;College of Engineering and Construction,Nanchang University,Nanchang 330031,China)

机构地区：[1]江西省交通监控指挥中心,江西南昌330036 [2]南昌大学工程建设学院,江西南昌330031

出　　处：《市政技术》2024年第11期84-91,126,共9页Journal of Municipal Technology

基　　金：江西省交通运输厅科技项目(2022X0047)。

摘　　要：短时交通流精准预测是高速公路交通运行状态精细化监管的重要手段,有助于提前监测高速公路潜在车流拥挤事件并及时管控。国内外学者已经从数理统计、数据驱动的维度提出了多种短时交通流的预测方法,虽然成果颇丰,但对交通流数据在时间关联性、空间关联性方面的共同建模能力不足,导致预测精度仍然有提升的空间。基于此,笔者提出了一种时空注意力扩散图卷积模型(STAtt-DGCN),来进行高速公路交通流的短时预测。该模型依托经典的时间注意力机制、空间注意力机制和图卷积网络,设计了时空模块、时空卷积模块以及扩散图卷积网络模块,来分别建立交通流数据在时间、空间维度的关联性,从而使预测精度得到有效提升。选取了江西省某高速公路3个月的ETC数据集来验证所提模型的性能,并选用ARIMA、LSTM、STGCN等常见基线模型来进行模型的对比评估。实验结果表明:STAtt-DGCN模型几乎在每个月的数据集上都展现出较好的预测能力。以2022年4月为例,与最具挑战的STGCN基线模型相比,所提模型在平均绝对误差、均方绝对误差、平均绝对误差上分别下降了17.9%、40.0%、11.0%。这意味着STAtt-DGCN模型的预测精度相较于基准方法有较大提升,可应用于高速公路交通流精准预测。Accurate prediction of short-time traffic flow is an important means of fine supervision of highway traffic operation status,which helps to monitor potential traffic congestion events on highways in advance and control them in time.Scholars at home and abroad have proposed a variety of short-time traffic flow prediction methods from mathematical statistics and data-driven dimensions.Although the results are quite fruitful,the joint modeling capability of traffic flow data is still insufficient in terms of temporal correlation and spatial correlation,resulting in that there is still improvement room in prediction accuracy.Based on this,a spatiotemporal attention diffusion graph convolution model(STAtt-DGCN) is proposed for the short-term prediction of highway traffic flow in this study.Relying on the classical temporal attention mechanism,spatial attention mechanism and graph convolution network,a space-time block,a spatiotemporal convolution block and a diffusion graph convolution network block are designed to establish the correlation of traffic flow data in the temporal and spatial dimensions,so that the prediction accuracy can be effectively improved.In this study,a 3-month ETC dataset of a highway in Jiangxi Province is selected to verify the performance of the proposed model;The common baseline models such as ARIMA,LSTM,and STGCN are chosen for the comparative evaluation of the models.The experimental results show that the STAtt-DGCN model exhibits better prediction ability of dataset almost every month.Taking the data in April 2022 as an example,compared with the most challenging STGCN baseline model,the proposed model shows a decrease of17.9%,40.0%,and 11.0% in mean absolute error,mean square absolute error and mean absolute error respectively.This shows that the prediction accuracy of the STAtt-DGCN model is greatly improved compared with the baseline method,and can be applied to the accurate prediction of highway traffic flow.

关 键 词：短时交通流预测 高速公路 深度学习模型 时空注意力机制 扩散图卷积网络 

分 类 号：U491.14[交通运输工程—交通运输规划与管理]