Research On Traffic Flow Forecast Based On Graph Neural Network

With the development of information technology and artificial intelligence,the Intelligent Transportation System(ITS)has gradually improved and become one of the important solutions to traffic problems.Traffic flow forecasting is an important part of ITS,which is of great significance to traffic planning,control,and management.Accurate and rapid traffic flow prediction is very important for traffic management departments to customize accurate traffic management plans and improve traffic efficiency.Because traffic flow data is typical spatio-temporal data,the mining of spatio-temporal relationship is the focus of the traffic flow forecasting.However,the spatio-temporal relationship of traffic flow is highly dynamic and nonlinear,which increases the difficulty of improving the prediction accuracy and efficiency.In order to solve the above problems,the following two traffic flow forecasting methods based on the graph neural network are proposed in this subject.(1)In view of the difficulty in improving the accuracy of the prediction,this subject mining the spatio-temporal dependence and spatio-temporal similarity in traffic flow,expounds on the limitations of existing methods,and proposes a traffic flow prediction method based on spatio-temporal dependence and similarity perception.This method combines two modules to model the complex spatio-temporal relationship in traffic flow.One of the modules is the spatio-temporal similarity feature module(STSF),which is newly designed in this subject,and is used to model the spatio-temporal similarity.The other module is a spatio-temporal convolution module with attention mechanism,which can dynamically model spatio-temporal dependence.A large number of experimental results show that the prediction accuracy of this method is improved by 3.76% on average compared with the current popular methods.(2)In view of the difficulty in improving operational efficiency,this subject analyzes the shortcomings of computational redundancy and incomplete modeling in the current models used to model the spatio-temporal relationship,and proposes a fast and dynamic temporal graph traffic flow prediction method to optimize it.This method consists of a temporal module and a spatial module.The temporal module is composed of an improved fast temporal convolution network,which can improve the modeling efficiency while modeling the temporal relationship.The spatial module is mainly composed of graph convolution based on self-attention and the dynamic convolution matrix proposed in this subject,which is used to model spatial relationships.A large number of experimental results show that compared with previous methods,the fast and dynamic temporal graph method can improve the training speed by 30% while ensuring the prediction accuracy.