Research On Traffic Flow Prediction Based On Spatial-Temporal Graph Neural Network

With the process of rapid urbanization,the urban traffic system is facing more and more serious challenges.Intelligent transportation system has become an important way to solve this problem,and traffic flow prediction is an important part of the system.Effective traffic flow forecasting is a prominent study topic since it may assist the system make decisions more effectively.In recent years,due to the excellent spatio-temporal correlation analysis ability of the spatio-temporal graph neural network,it has been widely used in the field of traffic flow forecasting and has strongly promoted the progress of this field.However,previous models rely too much on prior knowledge such as road network graph,and the spatio-temporal dependence of model modeling is incomplete.To address the above problems,two models of traffic flow prediction proposed in this thesis.The two models proposed can largely alleviate the dependence on prior knowledge,and can fully exploit the dynamics of spatio-temporal changes.Finally,the predictive ability of the models is verified by public datasets.The main contributions of this thesis are as follows:(1)We propose a graph neural network traffic flow prediction model based on attention mechanism.The spatial structure is learned through the adaptive embedding representation to replace the road network graph,thereby alleviating the model’s dependence on prior knowledge.The model is completely built with the attention mechanism,which captures the timing changes in the timing through the attention mechanism,and embeds the timing changes into the adaptive embedded representation with the spatil generation function to simulate the dynamic spatial dependence,and then uses the attention mechanism to model the complete spatiotemporal dependency to complete the final modeling prediction.The model performance is verified on four public datasets and compared with multiple models,proving that the model has advanced performance.(2)We propose a graph neural network traffic flow prediction model based on mixed mechanism and attention mechanism.Aiming at the problems of incomplete spatio-temporal modeling,insensitive local timing capture and unstable modeling in the previous model,this thesis proposes a new prediction model along the same design idea.First,the recurrent neural network is used as the basic framework to alleviate the problem that the attention mechanism is not sensitive to local timing capture.On this basis,the Mixed Graph-GRU is designed to model the spatio-temporal dependence,which integrates the graph convolution unit and the spatial attention unit,respectively modeling the static spatial dependence and the dynamic spatial dependence,and finally uses the mixed mechanism to model the complete spatio-temporal dependence.At the same time,in order to expand the time series receptive field of the overall model,the temporal attention mechanism is used to mine time series correlations in the global time series to further improve the prediction performance of the model.Finally,validation experiment is carried on the same public dataset,proving the superiority of the model.