Travel Time Prediction Considering Uncertainty And Individual Differences

Reliable prediction of travel time is an important part of travel service.Accurate travel time prediction can help travelers reasonably choose departure time and travel path,reducing traffic load and pressure of urban transport network during peak hours.This help improve the service level of traffic network and is of great significance to the development of urban intelligent transportation.However,the existing travel time prediction methods are mainly point prediction,which only provide a predicted value,unable to reflect the uncertainty and randomness of travel time.These methods fail to provide time reliability,travel time in extreme cases and other important information of travel time to travelers.Therefore,this thesis studies the prediction method of travel time distribution under the background of big data.Firstly,the framework of multi-source traffic data processing and fusion is established.This thesis summarizes the data sources of the existing transportation system and analyzes the necessity of data fusion.On this basis,the license plate recognition data processing and feature extraction method based on license plate matching,track data processing and feature extraction method based on map matching and road weather data extraction method are proposed.The proposed multi-source traffic data processing and fusion framework takes advantages of the multi-source traffic data in the existing traffic system and provides a reliable data basis for travel time prediction.Secondly,in order to predict travel time distribution,a deep leaning based mixture density temporal convolutional network(MDTCN)model was proposed.MDTCN consists of two modules,including temporal convolution module and mixture density module.Temporal convolution module time extracts information from traffic data inputs using hidden layers with convolution structure and the information would be fed to mixture density module.Mixture density uses the gaussian mixture model to represent the distribution and output the probability distribution function of travel time.MDTCN aims at maximizing the logarithmic likelihood function of observed travel time and predicted distribution.The parameters of the neural network are modified by back propagation algorithm,and the model error is iteratively reduced to accurately predict travel time distribution.Thirdly,in light of the “black box” characteristics of complex machine learning models,the prediction rules of these models can be hardly understood.Therefore,different methods for interpreting machine learning models are proposed in this thesis.The methods for linear models,decision tree model and deep learning model of travel time prediction are presented to improve traffic management policies which are aimed at reducing travel time and improving travel time reliability.Finally,the prediction accuracy of MDTCN model was verified by numerical examples.The numerical examples show that the proposed model can predict the travel time distribution with high accuracy.In addition,the interpretation methods of machine learning models can well explain the prediction rules of machine learning models.