融合自适应邻接矩阵与动态消息传递机制的时间卷积交通流预测方法

打开文本图片集

中图分类号：TP183 文献标识码：A 文章编号：1006-8228（2025）08-16-06

Abstract：Asurbantraffcnetworksbecomeincreasinglycomplex，traditionaltraffcflowpredictionmetodsstruggletoeffectively modeldnamicspatiotemporaldependenciesandroadnetworkstructures.Thispaperproposesadeepleaingmodelthatintegrates adaptiveadjacencymatrixconstructionadynamicmesagepasingmechanism，andacausaltemporalconvolutionalnetworkto effectivelyextractspatiotemporalfeaturesfromtrafficdataandimprovetheaccuracyofurbantraffcflowprediction.Themodel firstgeneratesanadaptiveadjacencymatrixthroughlearnablenodeembedings，overcomingthelimitationsofstaticgraph structuresindynamictraffcscenarios.Then，amodifiedmessgepassngmodulewithasinglegatingmechanismisdesignedto efficientlyfuseinformationfromcurrntandneighboringodes.Finally，acausaldilatedtemporalconvolutionisemployedto capturelong-termtemporaldependencies.Experimentsconductedontworeal-worldtraficdatasets（PEMS-BAYandMETR-LA） demonstratethattheproposedmodeloutperformsvariousmainstreambaselinesinbothshort-andlong-termpredictiontasks. Ablationstudies furthervalidatethesignificant contributionsofeach componenttotheoverallmodel performance.

Keywords：TraffclowPredction;SpatiotemporalFeatures;MessagePassing;AdaptiveAdjacencyMatrix；TemporalConvolution

0引言

随着城市化加快和机动车数量激增，城市交通拥堵已成为普遍且严重的问题，影响出行效率和社会发展。（剩余10096字）