Traffic Prediction Based On Multi-graph Convolutional Neural Network

As a very significant research field in the Intelligent Transportation System,traffic prediction is of great importance to travel planning,traffic management and congestion control.In recent years,various graph convolutional network models have emerged one after another,which has greatly promoted the development of the field of traffic prediction.However,the existing traffic prediction models still have problems such as incomplete spatial feature mining,insufficient temporal feature mining,lack of adaptability of the model and so on.Focusing on the problems in traffic forecasting and combining its domestic research status,this paper designs better traffic forecasting models and algorithms,which can further improve the performance of traffic forecasting.The main research contents are as follows:(1)In order to solve the problem of poor prediction performance due to incomplete spatial feature mining,this paper proposes a traffic prediction method based on multi-view graph convolutional neural network.This method extracts a variety of spatial relationships in traffic data,which regarded as different graphs,and then designs a fusion algorithm to generate a structural features of multiple graphs,fusing multiple graphs into a new graph containing multiple spatial relations.Then we apply a graph convolutional neural network to the new graph for extracting spatial features,using the gated recurrent unit neural network to extract temporal features,so that the model can achieve better traffic state prediction performance.Experiments show that on the PEMS228 dataset,compared with the best traditional machine learning methods,the RMSE of FMGCN is reduced by 14.4%,and the MAE is reduced by 5.2%.Compared with the single-graph convolution model,the RMSE of FMGCN is reduced by 6.0%,and the MAE is reduced by 7.8%.(2)Aiming at the problem that lacking the overall context characteristics of the temporal and spatial information in the traffic prediction method and then lead to the unstable prediction performance,this paper proposes a prediction method based on the multi-graph convolutional attention mechanism.This method adds an attention mechanism to the temporal dimensions o for focusing on the information of key nodes in the temporal dimensions.At the same time,the graph convolution al gated recurrent unit neural network is proposed to minimize feature loss.By improving the adaptive learning ability of the prediction algorithm to the overall context feature,the effect of improving the prediction performance is achieved.Experiments show that the 15 min,30min and 45 min prediction results of AMGCGRU on the on the LOS-LOOP dataset are better than the FMGCN method,the RMSE is reduced by 4.4%,2.0% and 1.1%,and the MAE is reduced by 7.2%,1.7% and 4.6%,which prove the advancement and effectiveness of the method proposed in this paper.