Urban Traffic Flow Prediction Based On Spatiotemporal 3D Convolution Neural Network

Traffic flow prediction is an important part of smart city construction.Efficient dynamic analysis and prediction of traffic flow is of great significance for solving congestion,road planning and intelligent transportation construction.The goal of urban regional traffic flow prediction is to predict the future traffic information by analyzing the historical traffic GPS data,so as to help people make better travel decisions.In the real world,accurate traffic forecasting is a huge challenge,which is affected by many complex factors.Early studies mainly used statistical learning methods and machine learning methods to solve the traffic flow prediction problems of time series and road networks.However,these two kinds of prediction methods can not simultaneously model the more complex and huge spatiotemporal correlation and other spatio-temporal characteristics,which limits the expression ability of the model.Although some traffic flow prediction methods based on deep learning can automatically learn the spatiotemporal characteristics from the trajectory dataset and can well model the spatiotemporal periodicity and spatiotemporal complexity,these studies usually destroy the inseparable spatiotemporal correlation in traffic flow and do not pay attention to the spatiotemporal flexibility in traffic flow data.In order to solve the problems existing in current research methods,this paper proposes a spatiotemporal convolution neural network(ST2DNet)to capture the endogenous spatiotemporal correlation in traffic flow.Because it mainly captures the impact of time changes on traffic flow prediction by stacking convolution layers,it fails to deal with the spatiotemporal flexibility in traffic flow data.This paper further proposes a spatiotemporal traffic flow forecasting model based on 3D convolutional neural network,which is called spatiotemporal adaptive 3D convolutional neural network(STA3DCNet),and is used to forecast the traffic flow in urban areas.The traffic flow model uses 3D convolutional neural network to model both temporal and spatial features to capture the inextricable spatial and temporal correlations.An adaptive transform module is proposed to assign different weights to channels to capture the spatiotemporal flexibility.In this study,two open urban area vehicle trajectory datasets,Chengdu and Xi’an,were used to carry out relevant experiments.Experimental results show that the proposed traffic flow prediction model STA3 DCNet is superior to other models.The ablative experiments carried out in this study further proved the superiority of the model.The core innovations of the proposed STA3 DCNet model are as follows:(1)In order to capture spatiotemporal correlation and spatiotemporal flexibility better,an adaptive 3D convolution module is proposed to model spatiotemporal dependency.In order to capture both temporal and spatial features at the same time so as to maintain the spatiotemporal correlation,a new module called as 3DConv SE is proposed by using the residual structure and the 3D convolutional neural network design.In order to capture spatiotemporal flexibility,we assign different weights to different channels through the selection mechanism proposed in our adaptive transform component.Through the ablation experiment,it is proved that these two points are helpful to improve the prediction performance of the model.(2)Considering the spatiotemporal complexity at the same time,with the help of embedding vector and mask matrix,this paper introduces external factors into the urban regional traffic flow forecasting method using integrated model,finally,weighted feature fusion module is used to deal with the spatial-temporal features and external features,which further improves the performance of the model.