Browse

A Machine Learning Approach for Freeway Speed Prediction
고속도로 속도예측을 위한 기계학습 접근법

Cited 0 time in Web of Science Cited 0 time in Scopus
Authors
최윤영
Advisor
고승영
Major
공과대학 건설환경공학부
Issue Date
2018-02
Publisher
서울대학교 대학원
Keywords
Freeway Speed PredictionSupport Vector MachinePrincipal Component AnalysisFeature SelectionPropagation of Traffic State
Description
학위논문 (박사)-- 서울대학교 대학원 : 공과대학 건설환경공학부, 2018. 2. 고승영.
Abstract
Prediction of freeway traffic speed can be used for predictive traffic management to improve the quality of the intelligent transportation system. The data-driven prediction is widely used due to its predictive capability. Recently, the non - parametric method using machine learning shows excellent predictive capability. In these methods, the feature extraction or selection is used to mitigate the overfitting and reflect the congestion mechanism. Although this nonparametric approach can be used as advanced traveler information system due to its excellent capability, it cannot provide any information on the congestion mechanism. Lack of information makes it difficult to establish a strategy for use in operational management. This study proposes a highway speed prediction model based on machine learning approach with a feature selection that provides both high predictive performance and interpretation of traffic flow characteristics. To do this, a supervised feature selection is applied using principal component analysis (PCA) based variable grouping and ordering and support vector machine (SVM) based variable selection. Varimax rotation is also applied to obtain the simple structure. In the variable ranking, the variables in the PC are ranked by using the nonlinear correlation coefficient which implies the predictive capability in the machine learning model. The cross-correlation coefficients were used in this study. With this grouped and ranked variables, the variables are selected by the forward selection method. The machine learning regression model in this study is SVM regression which has excellent generalization performance and low computational cost. Empirical data evaluation was implemented based on the several month's data of Kyungbu freeway in Korea and the interstate (I-880) freeway in the United States. Comparing other approaches, the proposed feature selection approach well captured the characteristics of traffic flow among spatiotemporal variables. In particular, the feature selection performance is somewhat better than that of the artificial neural network feature extraction model, stacked auto-encoder, and the ensemble learning model, random forest. The vector space of the PCA is transformed into the traffic phase diagram between two spatiotemporal variables to obtain the implication of proposed approach in traffic engineering area. Based on the traffic phase interpretation, principal components with some loading of dependent variable can explain the propagation of traffic state. The proposed approach captures the propagation of traffic state well according to prediction step. The proposed approach would be used to establish strategies for avoiding congestion or preventing rear-end accidents because it has advantages in the multi-step prediction on congested areas and in identifying the congestion mechanism.
Language
English
URI
https://hdl.handle.net/10371/140519
Files in This Item:
Appears in Collections:
College of Engineering/Engineering Practice School (공과대학/대학원)Dept. of Civil & Environmental Engineering (건설환경공학부)Theses (Ph.D. / Sc.D._건설환경공학부)
  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Browse