Sampling uncertainties in ensemble weather forecasting

Sammanfattning: The aim of ensemble weather forecasting is to provide probability forecasts for the occurrence of meteorological events. The ensembles are constructed by assembling several forecast realisations, each member of the ensemble being constructed to sample the uncertainties in the forecast. These originate from uncertainties in the initial conditions (the analysis) and imperfections of the numerical model.  In order to sample the initial uncertainties several techniques have been proposed. The singular-vector technique yields perturbations optimised to maximize the perturbation growth over a finite time interval, whereas the breeding method recycles the perturbations from the previous ensemble in order to sample growing modes.  The ensemble-transform method represents a further development of the breeding method. Here, to create initial perturbations independent of the current flow situation of the atmosphere, random perturbations are introduced by using the difference between two randomly chosen atmospheric states (i.e. analyses). The method produces dynamically balanced perturbations denoted Random Field perturbations (RF).  Our results show that the RF perturbations initially have the same dynamical properties as the variability of the atmosphere. After integration over a day the perturbations from all three methods (RF, singular vectors and ensemble transform perturbations) converge. The skill scores indicate a statistically significant advantage for the RF method during the first 2-3 days for most of the evaluated parameters. Over the medium range (3-8 days) the differences are very small. We also discuss the influence of the asymptotic variability of the forecasting model on the ensemble properties.