On clustering of non-stationary meteorological time series. submitted to the Journal of Climate, (available via biocomputing.mi.fu-berlin.de

• Main Author: Illia Horenko
• Other Authors: The Pennsylvania State University CiteSeerX Archives
• Format: Text
• Language: English
• Published: 2008
• Subjects: North Atlantic
• Online Access: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.576.3058; http://biocomputing.mi.fu-berlin.de/publications/Ho08b.pdf

A method for clustering of multidimensional non-stationary meteorological time se-ries is presented. The approach is based on optimization of the regularized averaged clustering functional describing the quality of data representation in terms of K regression models and a metastable hidden process switching between them. Proposed numer-ical clustering algorithm is based on application of the finite element method (FEM) to the problem of non-stationary time series analysis. The main advantage of the presented algorithm compared to HMM-based strategies and to finite mixture models is that no a priori assumptions about the probability model for hidden and observed processes are necessary for the proposed method. Another attractive numerical fea-ture of the discussed algorithm is the possibility to choose the optimal number of metastable clusters and a natural opportunity to control the fuzziness of the result-ing decomposition. The resulting FEM-K-Trends algorithm is compared with some standard fuzzy clustering methods on toy model examples and on analysis of multi-dimensional historical temperature data in Europe and a part of the North Atlantic.