Research On Traffic Flow Prediction Based On Spatiotemporal Data

As one of the basic components of Intelligent Transportation System(ITS),traffic prediction has been a hot issue in the field of transportation.In the past decades,with the expansion of urban scale,the extensive installation of high-end sensors,a large amount of traffic data continuously produced,the study of traffic prediction model has posed a great challenge.First,the expansion of the city requires the generalization ability of the traffic prediction model.However,the current advanced model based on Graph Convolutional Network(GCN)mostly relies on the access to all graph nodes,which limits its generalization ability.Second,the traffic data is a typical spatial-temporal data,it has the dual characteristics of time and space.Most of the previous studies only considered the temporal features of traffic data,but neglected the influence of the spatial structure of road network on the prediction results.Therefore,there is still much room to improve the accuracy of traffic prediction model.Thirdly,most of the spatial-temporal models have the problems of complex structure,high calculation cost and low training efficiency.How to reduce the model training cost and ensure the high prediction accuracy is also a direction worthy of study.Aiming at these problems,this paper proposes a Spatial-Temporal Gated Graph Attention Network(ST-GGAN)model.This model has a simple structure,low cost and high precision.Among them,we choose Graph Attention Mechanism(GAT)instead of GCN to extract the spatial features of road network.Compared with GCN,GAT can adapt to the development of urban planning well because it does not depend on the visit of all graph nodes.Secondly,GAT can assign different weights according to the importance of neighboring nodes,so as to accurately extract the spatial features of road network.Finally,GAT does not need feature decomposition,so it can save a lot of computation cost and improve the training efficiency of the model.In addition,the Gated Recurrent Unit(GRU)is selected to extract the temporal features of dynamic time series.GRU has a simple structure and few training parameters,so it can improve the training efficiency of the model and extract temporal features accurately.Experimental results on real datasets show that the proposed model outperforms all baseline models,and the prediction error is reduced by at least 22.4%.At the same time,compared with the best baseline method,the training efficiency of ST-GGAN is improved by nearly six times.In order to further improve the prediction accuracy of the model,this paper proposes the Bidirectional Gated Attention Network(Bi-GGAN)based on the ST-GGAN model,and we use the bidirectional GRU to extract the temporal features of traffic data.Compared with the ST-GGAN model,the prediction error of the Bi-GGAN model is reduced by 1.5%,and the training time is reduced by one third.To sum up,this paper proposes two traffic prediction models to solve the problems in the current traffic prediction field,and evaluates the performance of the two models based on a real dataset.Experimental results show that these models have achieved good results in short,medium and long term prediction.At the same time,both models significantly reduce the training time of the existing models,and the training time will not increase with the increase of the prediction time.