Clustering Methodologies Applied To Short-Term Ensemble Forecasting

Ensemble analysis is increasingly becoming a part of the operational weather forecasting. In this paper we are interested in the analysis of short-term ensembles. We present the results of applying two basic techniques from multivariate data analysis namely cluster analysis and principal component analysis to short-term ensemble forecasting. In particular we analyze the clustering tendencies in ten field variables – precipitation, pressure, temperature, and convective available potential energy at the surface level, height at 500 mb, wind at 250 mb, absolute vorticity at 250, 500, and 850 mb, and pressure vertical velocity at 700 mb – which are the output of the Eta model for ten different initial conditions. The aim of this work is to analyze the trajectories from these initial conditions and see how they cluster during evolution. In order to achieve our goal, we apply the rotated principal component analysis (RPCA) and several hierarchical and non- hierarchical clustering algorithms. Based on our analyses of these meteorological fields, we observe that the rotated principal component analysis and similarity-based cluster analysis do not provide useful clustering tendencies in all cases. They are very useful only when variance in the data set is not concentrated in one or two dimensions. However, dissimilarity or distance based cluster analysis provides useful information in all cases.