Research On Spatiotemporal Traffic Flow Prediction Based On Multi-Graph Convolution Gated Recurrent Unit

Traffic flow information is essential for effective traffic planning,operation,management,control and guidance,which is an important input for intelligent transportation systems(ITS).Traffic flow prediction methods analyze the variation patterns of historical traffic flow and predict the possible future conditions based on the current traffic state information.It is well recognized that accurate prediction of traffic flow is essential for effective traffic management and control.A lot of research has been conducted on the traffic flow prediction problem.Traditional prediction methods based on theories related to mechanics and statistics and new generation prediction methods based on machine learning,especially deep learning theory,have been developed.Traditional prediction methods have limited model complexity and cannot fully explore the spatial-temporal correlation of traffic flow,but the implementation process is simple and generalized.Machine learning and deep learning-based prediction methods have complex and diverse model structures.They are capable of mining deeper traffic flow information and optimizing model parameters automatically for different scenarios,with significantly better prediction accuracy than traditional prediction methods.However,deep learning traffic flow prediction models are currently in the development stage,and the analysis of traffic scenarios is lacking in the process of model design and hyperparameter calibration.In order to make the prediction model more closely integrated with the traffic scenario and have higher prediction accuracy,we analyze the spatial-temporal characteristics of the freeway network in the and accordingly establish a traffic flow prediction model based on deep learning method.The main contents of the thesis are as follows:Firstly,the correlation analysis was carried out on the traffic flow data in the time and space dimensions.The Pearson’s correlation coefficient was used to analyze the spatial correlation and three temporal correlations(i.e.,for short-term,daily and weekly)of traffic flow.We further propose a Spatial-Temporal Multi-Graph Convolutional Network(STMGN)based on Graph Convolutional Network(GCN),Gated Recurrent Unit(GRU)and Transformer architecture for traffic flow characteristics.GCN is used to deal with the spatial characteristics of traffic flow.In this thesis,we design the framework of GCN for the specific traffic flow prediction task.The adjacency matrix and correlation coefficient matrix of the road network are used as the input of the two GCN models respectively.The output of the two GCN models are fused using the gated fusion mechanism to improve the GCN model’s ability to extract spatial information.We combine GRU and Transformer architecture to extract traffic flow temporal correlation,where GRU is used to obtain short-term temporal features and transformer architecture is used to extract long-term features.Considering the multi-period properties of traffic flow,we propose three different data organization methods for short-term,daily and weekly characteristics for the data pre-processing to generate the input data for the prediction model.The proposed data organization is combined with the STMGN model to obtain the merged traffic flow prediction model MSTMGN by data fusion.Finally,a field freeway dataset from California,USA,and a urban road network simulation dataset from Changsha are used to analyze the prediction performance of the proposed models.The results show that data organization and data fusion can effectively improve the prediction accuracy of the model.We further compare the prediction effects of the proposed STMGN model and MSTMGN model with the classical baseline model,and the results show that the proposed model outperforms other prediction models at different prediction steps and different time periods,which illustrates the effectiveness of the proposed traffic flow prediction model.