Research On Methods And Applications Of Traffic Data Imputation And Predictability Analysis

The rapid development of the social economy has brought about a rapid increase in the demand for transportation.However,the limited capacity of the transportation system restricts the development of the urban economy.The intelligent transportation system(ITS)can effectively alleviate traffic congestion based on sufficient infrastructure and data mining technology.This paper focuses on the state prediction of the transportation system and applies the concepts and methods of machine learning and information theory to three key steps of data mining to improve the accuracy and reliability of the traffic state prediction.These three steps are data processing,data analysis,and data application.First of all,in the data processing stage,missing value is an important problem,and it is necessary to build an efficient missing value imputation model.In the data analysis stage,predictability determines the degree to which the traffic state can be accurately predicted,and reflects the regularity and complexity of the data distribution,which are of great significance to the prediction of traffic states.In the data application stage,predicting recurrent and non-recurrent traffic data is crucial to the decision-making of traffic managers and travelers,thereby alleviating traffic congestion.The main research content and results of this article are as follows:(1)Imputation of missing data: This paper establishes a hybrid spatiotemporal model based on residual analysis to fill the spatiotemporal data matrix.We select the expressway and urban road data sets,and summarize four patterns of missing data.The model is divided into two parts: time dimension and space dimension.The former is used to restore the overall trend of missing data;the latter imputes the time series residuals from the space dimension to improve the imputation accuracy.Numerical experiments verify the accuracy and robustness of the model under various missing rates and missing patterns.(2)Predictability of traffic data: Considering the regularity and complexity of traffic data itself,this paper establishes the predictability measure of discrete and continuous traffic time series based on the concept of Lempel-Ziv entropy in information theory.We excavate the temporal and spatial distribution characteristics of predictability measure,discuss the relationship between the predictability measure and other classical statistical measures,and reveal how predictability measure reflects the inherent uncertainty of time series.Based on the electrical signal processing method of discrete Fourier transform and wavelet transform,we disclose the relationship between the predictability measure and the recurrent or non-recurrent fluctuations of the time series.It is useful for the overall research on the regularity and complexity of data and is important for the prediction of traffic states.(3)Prediction application of data: Based on the machine learning model,we predict recurrent and non-recurrent traffic states.We use the traffic flow speed of urban roads as the recurrent data and correlate the prediction results with the predictability measure to verify the explanatory power of the predictability measure in the accuracy of prediction.We use the urban traffic incidents and their traffic flow data as non-recurrent data.With the data,we build a hybrid model which can comprehensively handle the "imputation plus prediction" problem and identify the factors that affect the duration and predictability of traffic incidents.We also study machine learning models that further improve the accuracy of prediction,which can support the decision of traffic management and optimization control.