Research On Traffic Information Prediction Mechanism Based On Data Mining

Intelligent Transportation System(ITS)is a developing traffic management system with the development of society and transportation.Its main idea is to combine the people,vehicles,roads and real-time situation of traffic,and to make fully use of them.In the field of Smart Travel,a key part of ITS,different traffic information needs to be predicted in advance to help people travel in a planned way.With the continuous development of data mining and big data,the focus of this thesis is to apply data mining technology into prediction mechanism study of short-term traffic information.This thesis focuses on the study of travel time and taxi demand,so that the existing traffic resources and vehicle resources can be utilized more effectively.For the study of travel time,we first propose the concept of segment speed and a scheme for prediction.Then considering the contribution of real data set,an improved adjacent similarity imputation algorithm is proposed for the scheme.Finally,we apply several machine learning models like KNN,ANN and GBRT in our traffic time prediction scheme.After data cleaning,preprocessing and missing values imputation of segment speed,the prediction results show that the improved imputation algorithm is better than the existing popular imputation algorithms.Results also demonstrate some advantages and disadvantages of different prediction models.For the study of taxi demand prediction,we analyze the problem from two aspects of time series and non-time series.In the aspect of time series,the SARIMA model is seemed as a baseline.Considering the pure linear limitation of SARIMA,a hybrid SARIMA-SVR model is proposed.First SARIMA is applied to time series prediction to extract the linear component of data.Then SVR model is applied to the residuals of SARIMA to extract the non-linear component.Then two parts are combined to offer the prediction result.In the aspect of non-time series,DBSCAN clustering algorithm is first used to mine the different hot index grids distribution.Then the hot index information is seemed as a kind of input feature into the ANN for model training.Finally,two parts are merged with the Ensemble Learning algorithm of Stacking to get the final high accurate prediction model.For both prediction studies,we use the MAPE index to evaluate the accuracy of prediction model.Results show that the proposed algorithms and the improved models have an obvious improvement on the accuracy of base model.This helps to improve the prediction mechanism of traffic information.