Research On Transportation Flow Prediction Based On Attention Mechanism And Graph Convolution Network

With the gradual increase of vehicles,various traffic conditions such as traffic congestion and accidents occur frequently.Urban traffic planning is the focus of attention in various countries.Building an intelligent transportation system is crucial for vehicle and road planning.The cornerstone of intelligent transportation systems is traffic prediction,so many researchers are increasingly paying attention to the issue of traffic data prediction.With the development of neural networks and the improvement of computer equipment performance,using neural networks to construct models and using GPU to train the constructed models has become a hot research direction in traffic prediction.The volume of traffic data is huge,and traffic data has strong temporal and spatial correlation.How to process a large amount of data and extract the spatiotemporal correlation of the data is a difficult problem in current traffic prediction.Researchers usually use Convolutional Neural Networks to extract spatial correlations,and Recurrent Neural Networks to extract temporal correlations and predict them.However,Convolutional Neural Networks are usually used to process regular grid data,while traffic data is irregular and Non-Euclidean structure data,so using Convolutional Neural Networks is difficult to get spatial correlation of traffic data.The spatiotemporal information of traffic data is usually fused,and extracting temporal and spatial correlations separately is not very effective.Therefore,it is necessary to consider simultaneously extracting spatiotemporal correlations of traffic data.This article proposes the following models to address these issues.In order to solve the problems of accurately extracting spatio-temporal correlation between traffic data,a traffic prediction model named STAT-GNN based on spatio-temporal attention mechanism was proposed.Firstly,one-hot encoding is used to generate temporal embedding and node2 vec method is used to generate spatial embedding,which is used to extract temporal and spatial information of traffic data.Secondly,Multi-Head Convolutional Self-Attention mechanism is used to extract temporal correlation of traffic data,exclude the influence of future moments,and increase the accuracy of prediction.Thirdly,Spatial Self-Attention Convolution Layer is used to extract spatial correlation of traffic data,and Add & Norm Layer is used to unify parameter dimensions and avoid gradient disappearance during training.This model can well extract the temporal and spatial correlations of traffic data.In order to solve the problems that static adjacency matrix is difficult to represent the constantly changing traffic conditions and many models cannot extract the temporal and spatial correlation simultaneously,a traffic prediction model DGARN based on dynamic graph attention recursive network is proposed.The dynamic graph generation module generates a dynamic adjacency matrix that changes over time.The dynamic adjacency matrix can represent changing spatial conditions.The model also proposes Dynamic Convolution Recurrent Layer to simultaneously capture the spatiotemporal correlation of traffic data.In order to improve the accuracy of traffic prediction,a transformation layer is added between the encoder and decoder in the model.The transformation layer can reduce the accumulated errors during the prediction process of the model and improve the accuracy of traffic prediction.In the experiment,in order to solve the problem of training difficulties,this thesis adopts the method of extracting a portion of the data to be processed and using batch processing for multi GPU training to solve the problem.Experiments were conducted on two common data sets,METR-LA and PEMS-BAY.Experimental results on two datasets show that STAT-GNN and DGARN have excellent prediction effect,which proves the effectiveness of the model proposed in this thesis.