Research On Real-Time Updating Method Of Traffic Flow Parameter Prediction Model Based On Transfer Learning

When the traffic state changes,if the dataset used by the trained deep learning model does not include new traffic state data,it will lead to a decrease in the prediction accuracy of the model.Therefore,in order to maintain the prediction accuracy of the model,it is necessary to update the model in real-time in order to manage and control road traffic appropriately.This paper combines Transfer learning and deep learning methods to solve this problem,and uses the Online Update Convolutional Neural Network Short and Long Term Memory Model(OUCNNLSTM),which has more advantages than the deep learning model in forecasting when the traffic state changes.OUCNNLSTM will transfer its ability to predict general traffic conditions from historical data to prediction models after emergencies,allowing a small amount of data generated after changes in traffic conditions to be used to train new prediction models,greatly reducing the training time of the prediction model.The update frequency of the prediction model is greater than the update frequency of data collected by traffic sensors,The accuracy of the model in predicting traffic parameters did not decrease after the traffic status changed.The specific research is as follows:Firstly,traffic flow parameters,deep learning model and Transfer learning method are studied.The definition and relationship of traffic volume,traffic density,and traffic flow speed,as well as the analysis of the temporal characteristics of traffic flow and the spatial characteristics of different locations;This paper summarizes the research work of artificial intelligence technology in the field of traffic prediction: from the traditional statistical model,machine learning model,deep learning model to Transfer learning method.Obtained the modeling principles for traffic prediction models.Secondly,carry out traffic data acquisition,data preprocessing,and data feature analysis.The California Department of Transportation Performance Measurement System(Pe MS)dataset was used as the research object of this article.Preprocess the dataset and perform time analysis on traffic data during daily peak hours and rest days;Carry out Spatial analysis of traffic flow data for the traffic data of multiple monitors located upstream and downstream on the same road.Finally,a traffic prediction model that can be updated online was constructed and its prediction performance was verified and analyzed.Based on the BP(Back Propagation)neural network,gradually build the OUCNNLSTM model and set the hyperparameters of the model.According to the classification of "migration content",the model adopts the method of parameter migration to carry out Transfer learning.The OUCNLSTM model has good adaptability to the traffic forecast during the change of traffic state.Through experiments 1,2,3 and 5 on the prediction of traffic data at three consecutive points of a road,the following conclusions are drawn: compared with the CNNLSTM model,the OUCNNLSTM model combined with Transfer learning and deep learning methods has higher prediction accuracy.For the predicted value during the traffic accident,MAE decreased by 14.250% on average,MSE decreased by24.302% on average,and OUCNLSTM can update the model in real time through Transfer learning.Compared to models updated with historical data,the update model rules used in this article reduce the update time by 92.199%.The OUCNNLSTM model achieved higher prediction accuracy using a smaller training dataset and changed traffic state changes by setting the number of lanes to validate the model’s effectiveness,achieving the same results.Therefore,the Mixture model used in this paper has the advantages of shorter training time and higher prediction accuracy compared with the deep learning model in view of the decline in prediction accuracy of the original prediction model after the change of traffic state.