基于多维特征融合与残差增强的交通流量预测

打开文本图片集

关键词：交通流量预测；动态图卷积网络；特征融合；残差建模；注意力机制中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2026）01-019-0161-09doi： 10.19734/j. issn.1001-3695.2026.06.0168

Multi-dimensional feature fusion and residual-enhanced learning for traffic flow prediction

Zhang Zhenlin 1 ，Guo Huijie’，Dou Tianfeng¹，Qi Kaiyuan²，Wu Dong²，Qu Zhijian1，Ren Chongguang1† （1.SchoolofueeedcogdoUeitfolbodo5i;spu Technology Co.，Ltd.，Jinan 250101，China）

Abstract：Traffcflowforecasting playsacrucialroleinintellgenttransportationsystems.Toaddress thelimitationsof existing methodsin featureutilizationandspatiotemporal dependencymodeling，thispaperdevelopedanovelmodel named MFRGCRN（multi-dimensionalfeature fusionandresidual-enhanced graphconvolutionalrecurrentnetwork）.Themodelcombinedautoencoders，depthwiseseparable convolutions，and temporal convolutions tocomprehensivelycapture spatiotemporal correlations.Itintegratedgatedrecurrntunitswithamulti-scaleconvolutionalatentionmechanismtoleacomplexdependencies，andadoptedamulti-scaleresidual enhancement moduletoprogresivelymodeldynamictraffc paterns.Experimental resultsonfourreal-worlddatasets demonstrate that theproposedmodelconsistentlyoutperformsbaseline methods inprediction accuracy.In particular，on the1-step forecastingtask of thePEMSO8 dataset，MFRGCRNachieves reductionsof approximately 7.7% in MAE， 2.9% in RMSE，and 4.5% in MAPE，highlighting its superior long-term prediction performance.The modelexhibitsstrongaccuracy，stability，androbustness，providinganefectivesolutionforcomplextraffcflowmodelingin intelligent transportation systems.

KeyWords：traffcflowprediction；dynamicgraphconvolutional network；featurefusion；residualmodeling；atentionmechanism

0 引言

在智能交通系统（ITS）快速发展的背景下，交通流量预测已成为缓解交通拥堵、提升路网效率的关键技术：随着城市化进程的加快，交通流量的不确定性和复杂性日益增加，对交通流量预测模型提出了更高的要求[1]。（剩余25252字）