Three improvements on KNN-NPR for traffic flow forecasting

Abstract

Research has shown nonparametric regression to hold high potential to accurately forecast short-term traffic flows. However, many fundamental questions remain regarding the ability of KNN-NPR(K nearest neighbor nonparametric regression) to meet real-time system requirements and adequate accuracy requirements. So this paper puts forward three improvements which are: effective traffic state vector selection method based on self-association analysis and association analysis; improved variable K search method based on "dense degree"; and advanced data structures based on a dynamic cluster method and hash-function transformation. A field test fully proves that with three improvements, KNN-NPR can adequately meet real-time system requirements and accuracy requirements.