ISSN1006-7167CN31-1707/TRESEARCHANDEXPLORATIONINLABORATORY第41卷第12期Vol.41No.122022年12月Dec.2022DOI:10.19927/j.cnki.syyt.2022.12.026基于ConvLSTM的高速公路交通流预测仿真研究吴剑云1,于安双1,2(1.青岛大学商学院,山东青岛266071;2.上海大学悉尼工商学院,上海201800)摘要:交通流通常具有复杂时空关联性,且易受天气、速度等外部因素的影响。为提高高速公路关键节点交通流预测的准确性,设计一种基于ConvLSTM网络且融合时空关联性和外部因素的交通流预测模型———STE-ConvLSTM。构建交通流、速度、天气时空矩阵,将其延深度方向堆叠,通过滑动窗口模型将其处理为类图像时间序列数据,利用ConvLSTM网络提取交通流的时空关联性和外部因素特征;利用卷积层实现交通流预测多变量多步输出。实验结果表明,相较于传统的交通流预测模型,该模型在交通流多步预测方面的预测准确度有所提升。关键词:ConvLSTM网络;深度学习;交通流预测;高速公路中图分类号:TP183;U491文献标志码:A文章编号:1006-7167(2022)12-0132-06SimulationStudyonHighwayTrafficFlowPredictionBasedonConvLSTMWUJianyun1,YUAnshuang1,2(1.BusinessSchool,QingdaoUniversity,Qingdao266071,Shandong,China;2.SILCBusinessSchool,ShanghaiUniversity,Shanghai201800,China)Abstract:Trafficflowusuallyhascomplextemporalandspatialcorrelations,andissusceptibletoexternalfactorssuchasweatherandtrafficspeed.Inordertoimprovetheaccuracyoftrafficflowpredictionatkeynodesoftheexpressway,atrafficflowpredictionmodel,STE-ConvLSTMmodel,basedonconvolutionalLSTMnetworkandintegratingspatio-temporalcorrelationandexternalfactorsisdesigned.Firstly,thespatio-temporalmatrixoftrafficflow,speedandweatherisconstructed,andstackedinthedirectionofdepth,whichisprocessedintotimeseriesdataofimage-likeimagesbyslidingwindowmodel,andthecharacteristicsofspatio-temporalcorrelationandexternalfactorsoftrafficflowareextractedbyConvLSTMnetwork;finally,themulti-variablemulti-stepoutputoftrafficflowpredictionisrealizedbyusingconvolutionallayers.Experimentalresultsshowthatcomparedwiththetraditionaltrafficflowpredictionmodel,thepredictionaccuracyofthemodelintermsofmulti-steppredictionof...