Research On Methods For Traffic State Identification And Prediction Based On Machine Learning

With the rapid development of society and economy,the demand for travel continues to grow,contradiction of road supply and demand has become increasingly prominent,resulting in the number and mileage of congested road increases year by year.Accurate and real-time traffic state identification and prediction play an important role for the intelligent management and control of traffic congestion.With the gradual implementation of ITS,all kinds of traffic detection equipment provides a large number of traffic data with different accuracy,breadth and depth.However,how to effectively analyze traffic data has become a huge challenge.Traditional data analysis methods often have a specific model structure and too many assumptions,so that cannot meet the needs of the analysis of various types of data.Therefore,it is necessary to study and explore new methods for traffic state identification and prediction,in order to fully tap the abundant traffic information contained in traffic data,and further improve the accuracy and reliability of traffic state identification and prediction.This study aims to improve the performance of traffic state identification and prediction,based on the analysis of various traffic data,advanced machine learning methods are used,and organically combine feature selection,swarm intelligence algorithm,time series analysis and other theories or methods,to conduct in-depth study for the traffic state identification and prediction methods.The main research contents and results are as follows:(1)A imputation methods for traffic flow missing data based on FCM optimized byPSO-SVRBased on the analysis of the temporal and spatial correlation of traffic flow data,a missing data imputation method using FCM optimized by PSO-SVR is proposed.Fuzzy C means clustering(FCM)is the basic algorithm,which is optimized by a combination of particle swarm optimization(PSO)and support vector regression(SVR).Firstly,analyzing traffic flow data missing patterns and the expression of missing data.Then,the spatial and temporal correlation of traffic flow data is analyzed by using urban expressway and arterial road data,respectively;and on this basis,the input of the proposed method is determined.Finally,the experimental scheme is designed from three angles,and the performance of the proposed method is verified.(2)A traffic incident automatic detection method based on variable selection and KELMBased on the analysis of the variation law of traffic flow during the incident,15 variables are selected to constitute the initial set for incident detection.Then,the important variables are selected from the initial variables by the random forest-recursive feature elimination method.The KELM model is trained with the important variables as input,and the model parameters are optimized by GSA.In addition,the SMOTE method is used to balance the incident samples and incident-free samples for dealing with the imbalance of these two kinds of samples.(3)A traffic incident duration prediction method based on NCA-BOA-RFBased on the analysis of the influencing factors of traffic incident duration and the characteristics of the selected data set,18 influencing factors were selected as the relevant variables for incident duration prediction.Through neighborhood component analysis(NCA)to identify the 6 key influence factors as the feature variables.Then the training set is constructed by using the feature variables,the random forest algorithm is trained and the Bayesian optimization algorithm(BOA)is used to optimize the algorithm parameters.In addition,in the process of testing the performance of the algorithm,the case of missing variables is considered.(4)A traffic state identification model based on spectral clustering and RS-KNNCombining the advantages of supervised learning and unsupervised learning algorithm,a traffic state identification model based on spectral clustering and RS-KNN is proposed.Based on the spot traffic parameters data,according to the operation characteristics of traffic flow and combined with the level of service(LOS)criteria for Chinese road,the spectral clustering is used to divide traffic state into four categories.Then,the classified traffic flow data are used to train the RS-KNN model.Finally,the validity of the model is verified by the measured data.(5)Two traffic state prediction models based on time series analysis and machine learningBased on the machine learning method,this paper builds two short-term traffic flow prediction models by combining two time series analysis theories,respectively.Firstly,according to the traffic flow time series has chaotic characteristics,combining the machine learning method and chaotic time series analysis theory,builds a short-term traffic prediction model based on multivariate phase space reconstruction(MPSR)and the combined kernel function-least square support vector machine(CKF-LSSVM).The empirical analysis shows that considering the chaotic characteristics and using the multivariate as the model input,which is helpful to improve the prediction accuracy.Secondly,in order to solve the problem that the observed traffic parameter time series contain noise components,combine the machine learning method with the time series denoising theory,a short-term traffic prediction model based on singular spectrum analysis(SSA)and kernel extreme learning machine(KELM)is proposed.The noise components of the original time series are filtered by SSA,then the KELM model is trained using denoising data and the model parameters are optimized by GSA algorithm.The empirical analysis shows that the denoising data are used as the model input,which is helpful to improve the prediction accuracy.