Traffic Flow Missing Data Imputation And Short-term Traffic Flow Prediction Method Based On Machine Learning

With the rapid development of the social economy,the number of car ownership has increased explosively,and the imbalance contradiction between road supply and demand has become increasingly acute,resulting in frequent traffic congestion in various cities.Compared with relying solely on the expansion of road infrastructure to alleviate traffic congestion,the establishment of a sustainable intelligent transportation system(ITS)through information technology can effectively assist traffic diversion.At the same time,with the advancement of data acquisition technology and means,a large amount of traffic data has been accumulated,but there is a problem of missing or damaged data.Accurate and timely traffic flow information is the foundation of ITS and plays an important role in the management and control of traffic congestion.Therefore,in order to provide better data support for ITS,it is necessary to study the methods of traffic flow missing data imputation and short-term traffic flow prediction.The purpose of this paper is to improve the accuracy of missing traffic data imputation and improve the accuracy of short-term traffic flow prediction.Considering the nonlinearity and complexity of traffic flow data,this study useed machine learning method to analyze and process traffic flow data.And a Nearest Neighbor Denoising Stack AutoEncoders(NN-DSAE)model was proposed for traffic flow missing data imputation,a Long-Short-term memory neural network(LSTM)and decision tree model was proposed for short-term traffic flow prediction.In the proposed traffic flow missing data imputation model,the principles of AutoEncoder,Denoising AutoEncoder,Denoising Stack AutoEncoders and Nearest Neighbor were introduced.Based on this,an imputation method based on NN-DSAE was proposed.The method firstly improved the given range of the neighborhood.And then according to the Nearest Neighbor knowledge,it found the traffic flow that has high correlation with the traffic flow missing data.The neighborhood matrix was used as the input value of the DSAE network.Finally,the traffic flow data obtained from the California Department of Transportation's PEMS system was used,which is to verify the performance of the NN-DSAE imputation method under different data missing rates.Meanwhile,compared with the historical average method and the traditional neural network method.The comparison results showed that the NN-DSAE-based traffic flow missing data imputation model is better than the other two methods.The proposed short-term traffic flow prediction model is a combined model.Firstly,the optimal wavelet threshold function and threshold selection method in wavelet analysis were determined by experiments.The db3 three-layer wavelet decomposition was used to decompose the traffic flow data,which processed by the missing data imputation,into uniform and random parts.Then,the LSTM network method was used to predict the uniform part of the traffic flow data.And the real traffic flow data collected by the PEMS system was used to verify the validity of the short-term traffic flow prediction model.Meanwhile,it is compared with ARIMA,shallow neural network(SNN),regression tree and traditional RNN method to measure the performance of the model.The results showed that the LSTM network was more stable and the prediction accuracy was better.The random portion of the traffic flow data was used as an error adjustment for short-term traffic flow prediction that used a decision tree approach.Finally,the combination of the traffic flow uniformity,which was predicted by the LSTM network,and the random traffic segment,which was predicted by the decision tree,was used as the final predicted traffic flow value.And the prediction effects of the single predictor LSTM and the decision tree were compared respectively.The results showed the combined model was more stable and better predictive performance.