Highway Short-term Traffic Flow Clustering And Forecasting Based On Similarity Measurement

Short-term traffic analysis is one of the most essential part of Intelligent Transport System.Short-term traffic clustering can identify the similar pattern of different route in different time and support traffic management.Short-term traffic forecasting can provide prediction of traffic condition in the future and support traffic control and guidance.This thesis uses the traffic volume of a certain traffic survey station near a scenic spot on holidays and normal days as the experimental data.The main strategy of innovation is to enhance the similarity measurement of traffic volume.Two research gaps,which were rarely focused on but of great pratical significance,have been solved.The one is the limitation of classical similarity measurement and its improvement in short-term traffic clustering.The other is using KNN method in short-term traffic forecasting with asymmetric loss.First,the basic information of data,exploratory data analysis and preprocessing methods are introduced.The differences of traffic pattern on holidays and normal days are discussed.Time series analysis is used to identify the generative mechanism of traffic flow series.Second,based on the different traffic patterns in holidays and normal days,hierarchical method,K-Means method and K-Medoide method are used to test the feasibility.LOESS smoothing technique is used to improve the effect of clustering and different parameter of LOESS are tested to find the different influence.Other classical similarity measurement except for Euclidean distance are tested.The limitation of Euclidean distance is discussed and enhanced by adding the punishment of stationarity of traffic flow series differences.Third,the baseline ARIMA model is builded firstly.The static version of KNN forecasting is transformed to dynamic version and is tested both on raw data and on smoothed data.The corresponding residual error diagnosis is also conducted.Then the sensitivity of three important parameters of KNN method are tested to find the most suitable value.At last,the problem of forecasting with asymmetric loss is proposed.Three possible methods which use the change of similarity measurement are proposed.The criterion index and KNN algorithm are transformed and tested in corresponding experiment.The feature of this thesis is enhancing the similarity measurement of traffic flow series to solve problems.The enhancement version of Euclidean distance and forecasting with asymmetric loss are proposed as innovation point.This research expands the applicability of KNN method in short-term traffic forecasting and get certain theoretical and pratical significance.However the method of forecasting with asymmetric loss is relatively rough and is sensitive to outliers.It needs to be improved by further corresponding research.