Analysis and Forecasting for Traffic Flow Data

Abstract

In this thesis, a number of techniques related to Principal Component Analysis (PCA) are used to derive core traffic patterns from streams of traffic data on a large number of road segments. Using a few number of k hidden variables, we show that the traffic information on the road segments can be captured by k traffic patterns. The dimensionality of the correlated road segments is successfully reduced from n to a much smaller number k by applying techniques related to Principal Component Analysis (PCA), where n is the number of road segments and k is the number of hidden variables. We use the k nearest neighbor(KNN) method to predict the values of the hidden variables over small time windows. As a result, we are able to forecast the speeds for n road segments very quickly. Our results are aimed at network-level and real-time prediction. In general, the computation of PCA is computationally demanding when n is large. A more efficient online version of PCA, called PASTd algorithm is used to reduce the data dimension. As a result, our forecasting method is efficient, flexible, and robust.