Research On Traffic Flow Prediction Of Spatiotemporal Sequence Based On Graph Convolution

The development of urban modernization and the rapid growth of population has brought huge challenges to the transportation capacity.Intelligent Transportation System is an important part of intelligent cities,and traffic flow prediction plays an important role in the intelligent transportation system.In recent years,the improvement of the ability of road network sensors to collect data has provided a large amount of data support for the field of traffic forecasting.On the other hand,the rapid development of deep learning computing power has also brought rich research avenues for traffic prediction.The essence of traffic forecasting is to obtain information on future road network conditions.Efficient and accurate traffic forecasting can not only help people plan to travel more rationally,save travel time and cost,but also help traffic management departments plan traffic and dynamic control more effectively.Traffic flow prediction is a typical spatial-temporal data prediction problem,which contains complex and dynamic nonlinear spatial-temporal relationships.Therefore,how to model spatial-temporal sequence data and effectively obtain the temporal,spatial,and spatial-temporal correlations contained therein is of great significance.Firstly,in the multi-view modeling research,this thesis proposes a Multi-View SpatialTemporal Adaptive Graph Convolutional Network(MVST-AGCN).Temporal,spatial,and spatial-temporal correlations are modeled using three views,respectively.Specifically,the short-term dependencies of the data are obtained based on a temporal view of residual connections and gating mechanisms.Using a spatial view of adaptive adjacency matrix graph convolution,spatial information is learned through multiple layers of trainable node embedding matrices.Based on the spatial-temporal view of the traditional Transformer encoder,it avoids the problem of error accumulation caused by recurrent recursive learning,and can better model the spatial-temporal relationship.Compared with the existing state-of-the-art methods,the proposed method achieves the best performance on four real datasets.Meanwhile,through a series of experiments,the effects of graph convolution modules,model hyperparameters and multi-view fusion methods on the prediction results are explored.Secondly,in multi-task modeling research,this thesis proposes a Multi-Task Dynamic Graph Convolutional Network(MTDGCN)to simultaneously predict traffic flow and speed.A dynamic graph is constructed for each time slice to obtain spatial topology information,and a dynamic graph convolution module is designed to further effectively mine the spatial correlation of nodes.Aiming at the high computational cost of traditional Transformers,a probabilistic attention network is used to reconstruct spatial-temporal views.Design a multi-task learning module to simultaneously predict the traffic flow and speed of the road network.The multi-task module can mine potential relationships between sub-tasks and avoid repeated training of sub-tasks.Through comparative experiments,compared with MVST-AGCN,MTDGCN improves the accuracy of the traffic prediction subtask by 6.04% on average and the speed prediction subtask by 4.43% on average.The experimental results show the feasibility of the multi-task module and the dynamic graph convolution module.Finally,based on the model proposed in this thesis,a demonstration system for traffic flow prediction based on graph convolution is designed and implemented.The main function of the system is the demonstration of geographic information and traffic node prediction.