Research On Short-term Traffic Flow Prediction And Data Completion Technology Under Deep Learning Framework

In order to alleviate the pressure of road traffic caused by the rapid development of society and economy,intelligent transportation systems are highly integrated in the traffic control process.Accurate prediction and analysis of road flow in a congested state plays a very important role in realizing traffic-induced diversion,bringing down security risks,and improving operating efficiency of the transportation system.However,current short-term traffic flow forecasting mainly performs single prediction of time series feature information,and lacks the extraction of high-dimensional feature information of traffic flow to reduce the accuracy of prediction.In view of the above problems,this paper uses the idea of deep architecture to predict the traffic flow during peak commutes.Through "differential data map" processing,the Deep Belief Networks(DBN)and Convolutional Neural Networks(CNN)that are widely used in image processing are used to extract the spatiotemporal characteristics of traffic flows.Support vector regression is connected at the top of these two networks to optimize the network structure.The completion of missing data further improves the accuracy of traffic flow prediction.The main contributions of this article are as follows:(1)Based on the construction of the "differential data graph" concept for normal working days traffic flow data,this paper proposes the application of DBN and CNN networks to traffic flow prediction,and the top network of DBN and CNN is individually connected to the SVR layer for regression prediction to improve Prediction accuracy.The experimental results show that the convolutional neural network models connected to SVR on top of the two deep networks have significantly improved the accuracy of traffic flow prediction,and the mean square error has been reduced by 0.003 and 0.036 respectively,and the prediction accuracy has been increased by 3% and 19% respectively.(2)Considering the completion of the original missing time series,a missing data prediction model is designed based on LSTM.This model learns the potential correlation and influence between historical data,and gets a good prediction effect for the prediction of missing data,and integrates the model with the most significant prediction effect from the previous step to make the final traffic flow prediction.The data set evaluation results show that the prediction accuracy of the LSTM model for missing data is improved by 3%.Combining the data obtained in(2)with the DG-CNN-SVR model in(1),the prediction accuracy is increased by 0.5% compared with that in(1).This paper proposes a combination model prediction method based on a deep frame that can accurately predict short-term traffic flow,and provide theoretical support for reasonable and efficient road system management.