Research On The Prediction Method Of Single Road And Road Network Traffic Flow Based On Deep Learning

The traffic flow prediction task is to learn a mapping relationship by using historical traffic data of a specific area,and then estimate the number of vehicles passing through the area in a certain period of time in the future through the mapping relationship.Traffic flow prediction is an indispensable part of intelligent transportation system,which can not only help traffic decision-makers maintain traffic order,but also help travelers choose better travel plans.Due to the characteristics of high dimensional,nonlinear and spatio-temporal correlation of traffic flow data,it is difficult to construct an accurate and fast traffic flow prediction method.The traffic flow prediction method based on deep learning can be trained by big data,and has strong ability of nonlinear relationship modeling and deep spatio-temporal feature extraction,which has an incomparable accuracy in traffic prediction tasks.In order to take into account the prediction accuracy and calculation time,this article conducts a more in-depth study of traffic flow prediction methods based on deep learning under both single road section and road network traffic backgrounds,the main research contents are as follows:Considering the high complexity of deep learning methods that improve the accuracy of traffic flow prediction by extracting spatio-temporal correlation features,a single road traffic flow prediction method based on the residual gated convolutional neural network with wide attention(WA-RGCNN)is proposed to ensure high prediction accuracy and short calculation time.Firstly,a parallel one-dimensional residual gated convolutional network(1D-RGCNN)is used to extract the spatio-temporal correlation features of historical traffic flow,historical traffic occupancy and historical traffic speed,which shortens the time required for the extraction of spatio-temporal correlation features.Secondly,a wide attention(WA)module is designed to weight the extracted spatio-temporal correlation features,which can enhance useful information and suppress the influence of useless information on the prediction results.Compared with the prediction method using ordinary attention modules,the prediction accuracy is similar,but the calculation time is shorter.Considering the positive effect of fully utilizing the interactive information between the spatio-temporal correlations of three traffic parameters on improving the accuracy of traffic flow prediction,a single road traffic flow prediction method based on the grammar gated convolutional neural network with analogous self-attention(ASA-GGCNN)is proposed to further improve the prediction accuracy.Firstly,the hidden features of three traffic parameters are extracted using a grammar module combined with two-dimensional gated convolutional network(2D-GCNN),which contains the spatio-temporal correlation features of each traffic parameter and the interactive information between them.Secondly,a analogous self-attention(ASA)module is designed to weight the hidden features,which further reduces the calculation time.Considering that the spatio-temporal correlation of traffic parameters in the road network is not only influenced by the spatio-temporal node information but also by the edge information between adjacent nodes,a road network traffic flow prediction method based on the grammar spatio-temporal synchronization graph attention network with adaptive edge information(AEI-GSTSGAT)is proposed,which has high prediction accuracy.Firstly,a sliding time window is used to reconstruct the input features of historical traffic data.Then,the AEI-GSTSGAT is used to extract the features of the reconstructed input features.When extracting hidden features,not only the information of each spatio-temporal node is considered,but also the adaptive edge information between adjacent nodes is considered,making the spatio-temporal topology information of the road network more fully and accurately expressed and utilized.Considering the long computation time required for feature extraction using graph neural networks on road network data with too many spatial nodes,a road network traffic flow prediction method based on the graph segmentation transer grammar spatio-temporal synchronous graph transformer(GS-TGSTSGFormer)is proposed,which has the advantages of high accuracy and short computation time on road network data with too many nodes.Firstly,the entire road network data is segmented into multiple local spatio-temporal subgraph data using the Louvain algorithm and sliding time window.Then,the grammar spatio-temporal synchronous graph Transformer(GSTSGFormer)is used to train these subgraph data in two stages: pre-training and transfer learning.In the transfer learning stage,only fine-tuning some parameters is required to train the accuracy of the prediction model,which improves training efficiency.Firstly,the entire road network data is divided into multiple local spatio-temporal subgraph data using the Louvain algorithm and sliding time window.Then,the grammar spatio-temporal synchronous graph Transformer(GSTSGFormer)is used to train these subgraph data in two stages:pre-training and transfer learning.In the transfer learning stage,only fine-tuning some parameters is required to train a high-precision prediction model,which improves training efficiency.Using Python language and Pytorch deep learning framework to validate the single road section and road network traffic flow prediction methods designed in this paper,and comparing the simulation results with existing traffic flow prediction methods,the advantages of the designed traffic flow prediction method in accuracy,rapidity and adaptability are verified.