Development of a precipitation downscaling method, and its intended use in the context of ensemble forecast post-processing

Meteorological ensemble forecasts frequently suffer from under-dispersion, and sometimes systematic biases. Statistical post-processing can effectively correct for these deficiencies, producing output forecasts that are reliable. But due to the number of grid points, lead times and weather variables involved, post-processing can only be done in a univariate fashion, ending up with a series of forecast distributions that lack a multivariate structure. In the perspective of hydrological modeling though, the covariability of the ensemble members across grid points, lead times and weather variables is crucial, and must then be restored.

In the first part of this seminar, I will briefly present the state of the art regarding the techniques used for restoring this covariability. In particular, I will focus on a particularly attractive method called Ensemble Copula Coupling (ECC), which restores on the post-processed forecast the same multivariate structure as in the raw ensemble. However, ECC gives poor results in the context where the raw forecasts are at much coarser resolution than the post-processed distributions, because of a missing fine-scale variability. This is particularly problematic for the variable precipitation, whose fine-scale variability plays a crucial role in hydrological modeling. To address that issue, we have developed a statistical precipitation downscaling technique, and intend to apply it on the raw forecast fields prior to ECC.

The second part of the seminar will focus on this downscaling method, which can also be used in a variety of contexts outside of post-processing. It is a stochastic technique based on the Gibbs sampling, an iterative process that is capable of introducing realistic, weather-dependent, and possibly anisotropic fine-scale details, while preserving the precipitation volume at the coarse scale. I will go over the basic principles, and briefly present how it can be calibrated such that the downscaled fields have the same “texture” to that of the analyses. Finally, I will conclude with the on-going work, and the scientific challenges that haven’t been solved yet.