Time series clustering using the the total variation distance with applications in Oceanography

Abstract

A time series clustering algorithm based on the use of the total variation distance between normalized spectra as a measure of dissimilarity is proposed in this work. The oscillatory behavior of the series is thus considered the central characteristic for classi cation purposes. The proposed algorithm is compared to several other methods which are also based on features extracted from the original series and the results show that its performance is comparable to the best methods available and in some tests it outperforms the rest. As an application the algorithm is used to determine stationary periods for random sea waves, both in simulations and on a real data set, a problem in which changes between stationary sea states are usually slow.