Urban traffic flow prediction: a dynamic temporal graph network considering missing values

AbstractAbstractAccurate traffic flow prediction on the urban road network is an indispensable function of Intelligent Transportation Systems (ITS), which great significance for planning. However, current methods still face many challenges, such as missing values and dynamic spatial relationships in flow. In this study, a temporal graph neural considering (D-TGNM) proposed prediction. First, inspired by Bidirectional Encoder Representations from Transformers (BERT), we extend classic BERT model, called Traffic BERT, to learn associations structure. Second, propose (TGNM) mine patterns data scenarios Finally, D-TGNM model can be obtained integrating learned into TGNM model. To train design novel loss function, considers problem flow, optimize The was validated actual dataset collected Wuhan, China. Experimental results showed that achieved good under four (15% random missing, 15% block 30% missing), outperformed ten existing state-of-the-art baselines.Keywords: missingtraffic predictiongraph networksdynamic graphTraffic AcknowledgmentsThe numerical calculations paper have been done supercomputing system Supercomputing Center Wuhan University.Disclosure statementNo potential conflict interest reported author(s).Data codes availability statementThe support findings study are available ‘figshare.com’ with identifier https://doi.org/https://doi.org/10.6084/m9.figshare.19642575.Additional informationFundingThis project supported National Key R&D Program China (International Scientific & Technological Cooperation Program) [grant 2019YFE0106500], Natural Science Foundation 41871308].Notes contributorsPeixiao WangPeixiao Wang PhD candidate State Laboratory Information Engineering Surveying, Mapping Remote Sensing (LIESMARS), University. He received M.S. degree Academy Digital China, Fuzhou University 2020. His research focus spatiotemporal mining, prediction, social computing, public health.Yan ZhangYan Zhang Sensing, also joint at School Design Environment, Singapore. collaborative sensing geographic knowledge services.Tao HuTao Hu Assistant Professor Department Geography Oklahoma (OSU). Before joining OSU, he worked postdoc fellow Geographic Analysis Harvard Kent interests include geospatial big analysis (i.e., media), health geography, human mobility, crime geography.Tong ZhangTong M.Eng. cartography information (GIS) University, 2003, Ph.D. geography San Diego Diego, CA, USA, California Santa Barbara, 2007. topics computing machine learning.