Traffic Flow Prediction Based On Spectral Temporal Graph Neural Network

In recent years,spatiotemporal series data prediction has been widely used in various fields,such as weather,traffic,nerve,etc.The prediction of traffic flow is a typical example.Traffic flow includes not only vehicles but also people flow.This paper mainly studies the passenger flow forecast of subway stations,which helps to improve the efficiency of people’s travel.With the continuous development of deep learning and graph neural network,more and more scholars began to try to apply it to traffic flow prediction,and constantly put forward more methods that can be used for traffic flow prediction.In the real world,traffic flow data has obvious temporal and spatial dependence.Therefore,when modeling the urban traffic flow prediction,it is necessary to fully consider the dynamic space-time correlation.This paper systematically studies and summarizes the existing traffic flow prediction methods at home and abroad.By comparing the advantages and disadvantages of various models,it provides an improved idea and sufficient theoretical basis for the follow-up research of this paper.The data were cleaned and pretreated to sort out the traffic data set with 5-minute and 10-minute intervals used in the experiment.The traditional method of learning the relationship between sequences often needs to be defined and assumed in advance.In this paper,the potential correlation layer of the spectral time graph neural network(Stem GNN)can be used to automatically generate the graph structure.In addition,the discrete Fourier transform and graph Fourier transform in the model can capture the correlation and time dependence between sequences in the frequency domain,and achieve effective prediction.In addition,based on the Stem GNN model,some further improvements are proposed.First,the spectral convolution layer in the spectral time graph neural network is replaced by Cheb Net,and the module is improved to obtain stronger expression ability and more adjacent node information.The second is to add a multitask learning model to the spectrum time graph neural network,including the prediction task of subway station passenger flow,the classification task of rise and fall,and the change point detection of passenger flow.Finally,seven baseline models are set for comparison,and the best experimental results are obtained on both data sets,which verifies the effectiveness of the improved model proposed in this paper...