铁道科学与工程学报2025，Vol.22Issue(11)：4819-4833,15.DOI:10.19713/j.cnki.43-1423/u.T20250164
基于时序聚类和图卷积的城市路网交通流预测
Traffic flow prediction in urban road networks based on time series clustering and graph convolution
摘要
Abstract
Traffic flow prediction in urban road networks,which can provide reliable future traffic flow information,is the key link to achieving urban area traffic control.To improve the accuracy of traffic flow prediction in urban road networks,a traffic flow prediction model based on time series clustering and deep learning(KS-GCN model)was proposed with the concern of spatiotemporal correlation of traffic flow.First,the time series clustering algorithm(i.e.,the K-shape algorithm)was used to build the prediction sub-region module,analyze the similarity of the sequence shapes of traffic flow at different locations in the urban road network,and divide traffic flow prediction sub-regions.Then,the graph convolutional network was utilized to build the spatial correlation module.The spatial correlation of traffic flow data and the road network topological structure of prediction sub-regions were analyzed.Next,the long short-term memory neural network was selected to build the temporal correlation module,and the temporal rules in the output of the spatial correlation module were investigated.Finally,the memory information in the temporal correlation module was input into fully connected layers,and the traffic flow prediction values of the urban road network were obtained.In addition,the traffic flow data in urban road networks in Sacramento,California,USA,and Chongqing,China,were selected to test the KS-GCN model.In addition,the ablation experiment was designed to discuss the mechanism of the KS-GCN model.The experimental results are included as follows.Compared with the KS-GCN model without the prediction sub-region module,the spatial correlation module or the temporal correlation module,the KS-GCN model can reduce the mean absolute percentage error,the mean absolute error,and the root mean square error by 2.72%～4.03%,0.55%～4.01%,and 0.99%～3.59%,respectively.Compared with the traditional clustering algorithm,the K-shape algorithm can minimize the mean absolute percentage error,the mean absolute error,and the root mean square error by 3.30%,0.34%,and 1.52%,respectively.Compared with the existing prediction models,the KS-GCN model can reduce the mean absolute percentage error,the mean absolute error,and the root mean square error by 6.31%～38.03%,1.68%～40.78%,and 1.48%～37.50%,respectively.The KS-GCN model performs best for traffic flow prediction in the urban road network,and can provide more accurate traffic flow prediction data for urban area traffic control.关键词
智能交通/城市路网交通流预测/图卷积神经网络/时序聚类算法/预测子区Key words
intelligent transportation/urban road network traffic flow prediction/graph convolutional network/time series clustering algorithm/prediction sub-region分类
交通工程引用本文复制引用
张文松,郭怡情,袁颖,姚荣涵,乔延峰,邹延权..基于时序聚类和图卷积的城市路网交通流预测[J].铁道科学与工程学报,2025,22(11):4819-4833,15.基金项目
国家自然科学基金资助项目(52172314) （52172314）
河北省自然科学基金资助项目(F2024403007) （F2024403007）
河北省教育厅科学研究项目(BJK2024090) （BJK2024090）
