Mitigating stochastic model error effects in ensemble forecasts: A post-processing method based upon filtering

A method of ensemble post-processing designed to counter the effects of stochastic model errors on ensemble covariances is investigated. The method performs a series of filtering experiments with the operational ensemble members, obtaining a set of non dynamical samples of the forecast space which is less corrupted by stochastic errors. It then uses some number of the samples to complement or supplant the operational members, forming a so-called hybrid ensemble. The hypothesis is that the composition of the hybrid ensemble will provide it with covariances that are improved relative to those of the operational ensemble. With improved covariances, and mean and variance comparable to that of the operational ensemble by design, the hybrid ensemble should yield multi-dimensional probabilistic forecasts that are better than those derived from the operational ensemble.; The thesis has two principal objectives. The first is to evaluate the method's ability to sample states in the forecast space which are better representations (in a root-mean-square sense) of possible forecast alternatives than are the operational members. Composite and ensemble-by-ensemble comparisons are made between the root-mean-square error characteristics of the distribution of samples and those of the operational ensemble. These comparisons are based upon a year's worth of global ensemble data, and suggest that the samples are in fact consistently better representations of possible forecast alternatives. The second principal objective is to assess whether versions of hybrid ensemble can be found that yield multi-dimensional probabilistic forecasts that perform better on a systematic basis than corresponding forecasts yielded by the operational ensemble. Simple versions of hybrid ensemble are constructed for a substantial number of 192h 500 hPa geopotential height ensembles. The performance of probabilistic forecasts derived from each of the hybrid ensembles is compared to that of similar forecasts derived from the operational ensemble using the Brier Score and Relative Operating Characteristic verification measures. The findings suggest that there are indeed ways of using the samples in a hybrid ensemble configuration that afford improved multi-dimensional probabilistic forecasts. These findings encourage continued investigation and refinement of the post-processing methodology.