Predicting the time evolution of OD matrices is a critically important topic for many applications within thetraffic domain, ranging from ex ante evaluation to real-time prediction and control. Since OD matrices are highdimensional multivariate data structures, the specification and estimation of such OD prediction models is bothmethodologically and computationally cumbersome. In this paper we demonstrate that by significantly reducingthe dimensionality of the OD data, in such a way that the structural patterns are preserved, we can reduce thecomputational costs dramatically, without significant loss of accuracy.In this paper we explore the application perspectives of principal component analysis (PCA) for this purpose.First, using PCA we find that the dimensionality of time series of OD demand can indeed be significantlyreduced. Moreover, we show how the results from the PCA method can be used to reveal structure in theunderlying temporal variability patterns in dynamic OD matrices. The results indicate that we can distinguishbetween three main patterns in dynamic OD matrices that follow structural, structural deviation and stochastictrends. We provide insight into how these trends contribute to each OD pair and how this information can beused further in predicting dynamic OD matrices on the basis of a set of dynamic OD matrices obtained from realdata.
展开▼
