Research On Short-term Travel Time Prediction Method For Segments Of Expressway

With the rapid development of the regional economic integration,the motorization level is increasing.The traffic congestion and traffic safety problems on expressways become more serious.The Intelligent transportation system is an effective way to solve these problems.As an important part of the intelligent transportation system,travel time prediction has attracted many attentions from the research field and many applications in traffic management centers.Real-time and accurate travel time prediction can support the decision of the manager and provide better service for travelers.From the previous literature,it can be found the current studies focus on the prediction methods.However,the study of the harmony between the travel time prediction part and the other part is less and the travel time prediction theory is still incomplete which will restrict the operation of the intelligent transportation system.Especially,in our country,the intelligent level of the expressway is developing whose information collection system,data processing system cannot meet the information and data requirement of the travel time prediction methods.To fill this gap,this paper studies the travel time prediction method and proposes a complete travel time prediction theory based on Key International(Regional)Cooperation and Exchange Projects of the National Natural Science Foundation of China “Research on multitensor networks for coupled high-dimensional multi-modal big data and its empirical study” as well as Ministry of Communications Science and Technology Demonstration Project “The intelligent platform for Jiangsu Expressway operation and information service”.Using the realworld data collected from China and America,the paper introduces the methods of data collection,data processing,feature extraction and model building.First,the paper introduces how to match the data from different detectors or database.Then,the methods to repair the traffic-related time series and panel data are proposed.Third,the traffic congestion detection and congestion duration time are studied.Finally,to improve the accuracy and consider the need of the travelers,a multiple steps prediction model and a model considering external variables are built.From a theoretical perspective,the paper can help us learn the temporalspatial rules of the traffic flow especially the travel time,understand the traffic congestion detection method and know the effect of external variables on travel time.From a practical perspective,the results of the paper can improve the intelligent level of the management and operation on expressways.Moreover,it can reduce traffic congestion and decrease traffic accidents on expressways.The release of the traffic information can enhance the travelers' experience.In summary,the main contents and findings of the paper are as follows:(1)The related literature about travel time prediction on expressways,including traffic flow data collection,traffic condition classification and prediction,is concluded.The shortage of previous studies is pointed.First,the collection of traffic flow,especially the travel time,is difficult.Second,the repair of the different types of data and panel data is insufficient.The studies of the traffic-related traffic flow data focused on single temporal-spatial correlation.Third,the problem of the sample in the traffic congestion detection model still remains.Moreover,some attention should be paid to the variable selection in model developing.Finally,the travel time prediction focused on the single step.The consideration of external variables in the prediction model is not sufficient.(2)The multiple types of data used in this paper are introduced.First,the position and characteristics of the two expressways are presented.Second,the characteristics of highresolution traffic flow data,traffic incident data and meteorological data are described.The new techniques to collect travel time is proposed.The two types of speed data are compared.Finally,the method to match the data in temporal and spatial dimensions is proposed.The methods in this section can provide data support for the following studies.(3)The methods to repair the data for travel time prediction is introduced.First,the correlations in traffic flow data is classified as local,global,temporal and spatial views.The global temporal correlation is modeled by Long-Short Temporal Memory(LSTM)network.The global spatial correlation is modeled by Support Vector Regression(SVR).The local views are modeled by collaborative filtering models.Then,four views are summarized by kernel regression to estimate the missing values.Second,the temporal-spatial features are extracted by auto-encoders,and then the features of different types of data are fused in the fusion layer.Based on the fusion layer,the missing values of different types of data are estimated.Finally,the variables in the traffic incident samples are defined as continuous and categorical variables that are modeled by linear regression and logit regression.To improve the robustness of the method,the traditional Multivariate Imputation by Chained Equations(MICE)is modified with which a complete dataset can be obtained.(4)The methods to detect traffic conditions and predict the duration time of congestion are proposed.First,traffic incident detection is defined as a binary classification problem.A temporal and spatial correlation variables selection method is introduced to select different types of variables.To improve the real-time characteristic of the detection model,a new sample selection method is proposed.Moreover,the generative adversarial network(GANs)is used to increase the number of samples.Second,using the data collected from microwave detectors,a traffic condition classification model is built based on k-means and traffic flow theory and evaluated by data from the video.Then,a traffic condition prediction model is proposed by modifying the traditional neural network.Finally,a Restrict Boltzmann Machine(RBM)and a Gaussian Bernoulli RBM(GBRBM)are applied to extract features from the continuous and categorical variables,respectively.A joint layer is built to combine the extracted features.Using the traffic accident data,the selection of parameters in the method is introduced and the accuracy is evaluated.(5)A travel time prediction method is proposed based on a deep learning model whose parameters are optimized.First,the temporal correlation of travel time data is analyzed using time series autocorrelation coefficient and partial autocorrelation coefficient.The spatial correlation of travel time data in the expressways is analyzed by global Moran's I index.Following,a variable section method is proposed.Second,the traditional Particle Swarm Optimization(PSO)is modified based on the knowledge of the experts which can rapid the running speed.Compared with statistical models,hybrid models and neural network models,it can effectively increase the prediction accuracy.Finally,the PSO is extended to multiple objective PSO(MOPSO)to optimize the deep belief network to achieve multiple steps prediction.In the hybrid model,the MOPSO is solved by Pareto Optimality.The hybrid model can reduce the error accumulation in the stepwise iterative algorithm.(6)A travel time prediction method considering the external variables is proposed.First,the influence factors are divided into emergent factor and non-emergent factor.A paired-sample t test is used to test the significance of the two types of factors.It proves the consideration of the external factors is necessary.Second,a travel time prediction method is proposed considering external factors.It includes three parts: variable selection,prediction method and evaluation.In the first part,the internal and external variables are selected by autocorrelation function and random forest.In the second part,the travel time series is decomposed into simple series by Ensemble Empirical Mode Decomposition(EEMD)method.Then,the simple series are predicted by random weight neural network.By combining the simple prediction results,the final prediction results are obtained.Based on the point results,a quantile regression is used to estimate the prediction interval.In the third part,the built model is evaluated by point evaluation criteria and interval evaluation criteria.