Support Vector Regression for Bus Travel Time Prediction Using Wavelet Transform
JournalofHarbinInstituteofTechnology(NewSeries),Vol.26,No.3,2019DOI:10.11916/j.issn.1005?9113.18025
SupportVectorRegressionforBusTravelTimePredictionUsing
WaveletTransform
YangLiu1,YanjieJi1?,KeyuChen2andXinyiQi1
(1.SchoolofTransportation,SoutheastUniversity,Nanjing210096,China;
2.GuangzhouUrbanPlanning&DesignSurveyResearchInstitute,Guangzhou510060,China)Abstract:Inordertoaccuratelypredictbustraveltime,ahybridmodelbasedoncombiningwavelettransformtechniquewithsupportvectorregression(WT?SVR)modelisemployed.Inthismodel,waveletdecompositionisusedtoextractimportantinformationofdataatdifferentlevelsandenhancestheforecastingabilityofthemodel.AfterwavelettransformdifferentcomponentsareforecastedbytheircorrespondingSVRpredictors.Thefinalpredictionresultisobtainedbythesummationofthepredictedresultsforeachcomponent.TheproposedhybridmodelisexaminedbythedataofbusrouteNo.550inNanjing,China.TheperformanceofWT?SVRmodelisevaluatedbymeanabsoluteerror(MAE),meanabsolutepercenterror(MAPE)andrelativemeansquareerror(RMSE),andalsocomparedtoregularSVRandANNmodels.TheresultsshowthatthepredictionmethodbasedonwavelettransformandSVRhasbettertrackingabilityanddynamicbehaviorthanregularSVRandANNmodels.Theforecastingperformanceisremarkablyimprovedtoobtainwithin6%MAPEfortestingsectionIand8%MAPEfortestingsectionII,whichprovesthatthesuggestedapproachisfeasibleandapplicableinbustraveltimeprediction.
Keywords:intelligenttransportation;bustraveltimeprediction;wavelettransform;supportvectorregression;hybridmodel
CLCnumber:U12 Documentcode:A ArticleID:1005?9113(2019)03?0026?09
1 Introduction
Bustraveltimepredictionisvitalcomponentofadvancedpublictransportationsystem(APTS)andadvancedtravelerinformationsystem(ATIS).Withtherapiddevelopmentofcommunicationandnetworktechnology,anaccurateandreal?timetraveltimeforecastisincreasinglyimportant.Forbusoperationmanagement,itcanhelpoptimizebusrouteplanning,stopsiteanddistancebetweenstationsselection,andchooseappropriateroadsectiontoimplementbusprioritytragedy,whichwillrealizebetterbuspriorityonthepremiseoflimitedtrafficsupply.Ontheotherhand,real?timeanddynamicbusarrivaltimeforecastreleasedbymobilecommunicationapplicationscanhelppassengersmakemoresuitabletravelplans,whichnotonlyreducesthelongwaitingprocess,butalsoimprovestheservicelevelofpublictransportation
andattractsmorepassengers.
Previously,variousmethodshavebeenadoptedbyresearcherstoforecastbustraveltimeusinghistoricalaveragemodel[1],timeseriesmodel[2],statisticalregressionmodel[3]andkalmanfilteralgorithms[4].However,thepredictionofbustraveltimeisverycomplexandhighlynonlinearinnatureasitdependsuponmanyinfluencefactorssuchasridership,trafficflow,weatherandtrafficsignalsinbussystem.Itisdifficultforthosepredictingmethodstoconsideralloffactors,sothepredictionquality,inpractice,isunsatisfactory.
Intherecentdecade,machinelearningmodelshavebettercapabilitytohandlenonlinearmappingproblemsthatarecomplexinnature,particularlyinthefieldoftraveltimepredictionwhereanartificialneutralnetwork(ANN)hasbeenwidelyapplied.ParkandRilettanalyzedtheperformanceofANNapplicationsinbustraveltimemodeling[5];Chien
Received2018-03-17.
SponsoredbytheProjectsofInternationalCooperationandExchangeoftheNationalNaturalScienceFoundationofChina(GrantNo.51561135003)andtheScientificResearchFoundationofGraduatedSchoolofSoutheastUniversity(GrantNo.YBJJ1842).?Correspondingauthor.E-mail:jiyanjie@seu.edu.cn.
·26·