A Study On Short-term Traffic Volume Forecasting Based On Non-Parametric Regression

Short-term traffic volume forecasting is one of key technologies of the Intelligent Transport Systems (ITS). Perfect performance of forecasting and meeting real-time requirement concern the effective realizations of traffic control and transportation induction system.Based on the analysis of the properties of traffic flows, this dissertation begins to recognize Non-Parametric Regression (NPR) from two different angles: deductive-inductive method and theories about non-linear time-variant system. NPR is suitable for short-term traffic volume forecasting theoretically. The main steps and influencing factors are discussed in applying NPR.NPR as a new intelligent method has many shortcomings restricting its applications which focus on: ill-suited database structure, low searching efficiency, open loop structure etc. This dissertation improves the method to advance forecasting accuracy and meet real-time requirement.The main improvements include:(1) The original volumes and searching data are separated to be stored in two databases. The databases are based on unidimensional and multi-dimensional structures and searching strategies. The applications of balanced binary tree and R tree as logistic structures, static chain as physical structure reduce the searching time and meet the real-time requirement.(2) A closed feedback loop is added upon the most important step--pattern matching. The matching results are amended by forecasting errors to improve forecasting accuracy.(3) Analysis the factors of NPR affecting robustness and focus on the contradiction between the collection of original data and real-time forecasting when rebuilding the system. The coefficient database and the idea of batch forecasting are put forward to solve the problem.NPR needs the data centers and K nearest neighbors round every center. But the original data does not possess these. Furthermore, it has features of high- dimensions and large superfluous data. So the pretreatment operations to original data are necessary. In this dissertation principal component analysis is adopted to bring down input variable dimensions and eliminate the relativities among them. Superfluous data is rejected by cluster analysis. Typical road network is simulated by traffic simulation software and the volumes data is gotten by all kind of simulation conditions. To the data for creating database, represented by volume modes with which the network has been simulated, the simulation parameters are set by large spans to get the marginal states of volumes. To the data for testing, the simulation parameters are set by small spans to study the mode evolutions. The experiments are focused on the comparisons of the two data structures at the respects of forecasting accuracy and time consumption. The comparison results show that the unidimensional structure is superior to multidimensional one.