Premium
Ensemble prediction in a model with flow regimes
Author(s) -
Trevisan Anna,
Pancotti Francesco,
Molteni Franco
Publication year - 2001
Publication title -
quarterly journal of the royal meteorological society
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.744
H-Index - 143
eISSN - 1477-870X
pISSN - 0035-9009
DOI - 10.1002/qj.49712757206
Subject(s) - ensemble forecasting , statistical physics , degrees of freedom (physics and chemistry) , range (aeronautics) , nonlinear system , infinitesimal , mathematics , phase space , consistency (knowledge bases) , physics , mathematical analysis , meteorology , materials science , geometry , quantum mechanics , composite material , thermodynamics
The primary goals of ensemble prediction are the identification of particularly unpredictable situations, and the prediction of forecast error. In the medium range, major forecast failures are associated with regime transitions. The predictions of linear theories on error dynamics are inadequate to describe these situations, characterized by the rapid growth of errors which are far from being infinitesimal. Little is known about the efficiency of ensemble forecasting methods when applied to highly nonlinear situations like those associated with a change of regime. This problem is examined in an atmospheric model with two regimes and a small number of degrees of freedom by comparing current ensemble forecasting methods with ensemble forecasts that span all the degrees of freedom of the system. The study, performed in a perfect model environment, shows how the statistical properties of errors are affected by regime transitions. Results obtained selecting initial perturbations in the direction of the leading Lyapunov vectors (LVs) and singular vectors (SVs) are compared with those obtained with initially random vectors (RVs) spanning the whole phase space. It is found that the LV ensembles do not necessarily capture those particular perturbations that eventually lead to a regime transition, but closely reproduce the average RV distribution. On the other hand, the SV ensembles reproduce very closely the distribution obtained by selecting the maximum amplitude of random perturbations, although their consistency with the distribution of the full RV ensemble is poorer than in the LV ensembles. When the spread of individual RV ensembles is analysed by an empirical orthogonal function (EOF) decomposition, one finds that the variations in the amplitude of the first EOF (which carries the bimodal signature of regime transitions) is better estimated by the SV than by the LV ensembles. In verification‐based contingency tables, incorrect categorical forecasts of transition probability are more frequent in the LV than in the SV ensembles.