Research On Analysis And Prediction Method Of Short-Time Traffic Flow Based On Deep Learning

In recent years,with the development of social economy and transportation,people's quality of life has been gradually improved.At the same time,due to the increase of private vehicles,the number of trips and the length of trips,serious traffic congestion frequently occurs in most cities,which not only extends people's travel time,but also increases the consumption of fuel by stop-and-go vehicles on highways,thus causing environmental pollution and damage.In view of such problems,a set of advanced and self-learning intelligent transportation system can play an important role in the transformation of urban traffic.For example,the prediction of traffic flow can indirectly guide and control the traffic flow,thus ensuring the safety of key transportation hubs.Among them,the accuracy and stability of the short-term traffic flow prediction results can play a key evaluation basis for the performance of the system.Therefore,it is of great significance to analyze and predict the short-term traffic flow.Due to the complexity of traffic flow data and its time-varying and nonlinear characteristics,it is difficult to make accurate predictions with conventional or single-feature models.In addition,many methods are not sufficient in time feature extraction,such as failing to consider the possible impact of weather conditions,holidays and other factors on actual traffic flow prediction.In view of this,this paper proposes two prediction models based on deep learning to model,analyze and summarize the short-term traffic flow.(1)Applying the principle of discrete wavelet decomposition into the neural network structure,a dynamic neural network structure(DWNN)model based on discrete wavelet was proposed,which was used to construct the deep learning model of frequency perception for traffic flow data analysis.Compared with the traditional stacked fusion model and the single deep learning model,DWNN model not only retains the advantages of multistage discrete wavelet decomposition in frequency learning,but also can fine-tune the wavelet basis matrix coefficient and all other neural network parameters under the framework of deep neural network.The trained forecasting model can predict the actual traffic flow information and finally obtain more accurate traffic flow information.In the experimental study,the real data set is preprocessed by missing value repair,noise reduction and normalization to improve the accuracy of prediction.The accuracy and validity of the model proposed in this paper are verified by comparing the prediction evaluation indexes of several models proposed in this paper.Experimental results show that the model is simple and efficient,and has better prediction accuracy to some extent.Compared with the results of traffic flow prediction using a single feature model,the accuracy of traffic flow prediction using multiple features learning proposed in this paper has been greatly improved.(2)Using DWNN model for short-term iterative training and prediction of traffic flow can be highly accurate,but it is found through experiments that the prediction accuracy begins to decline slowly as the prediction step size increases.In order to further optimize the predictive ability of DWNN model,a dynamic wavelet transform network predictive model based on attention mechanism was proposed.By combining dynamic wavelet neural network module and two-way short-and longterm memory model(Bi LSTM)codec,use of attention mechanism weighted and storage operations repeated training model of context information,the high correlation between the output and input sequence of the sequence distribution of high weight,then extract the useful information,the output stream sequence to forecast the future traffic flow data.Finally,the model is evaluated according to the evaluation index proposed in this paper.The results show that the optimized model can increase the predicted step size while maintaining the accuracy and robustness.To sum up,in the face of short-term traffic flow prediction problems,the two models proposed in this paper respectively solve the defects of traditional or single feature prediction models in complex data processing and the influence of model prediction step size on the prediction results.Combined with the experimental results,the effectiveness and robustness of the two proposed schemes are verified.