Statistical Postprocessing and Multivariate Structuring of High-Resolution Ensemble Precipitation Forecasts

A comparison of statistical postprocessing methods is performed for high-resolution precipitation forecasts. We keep hydrological end users in mind and thus require that the systematic errors of probabilistic forecasts are corrected and that they show a realistic high-dimensional spatial structure. The most skillful forecasts of 3-h accumulated precipitation in 3 × 3 km^2 grid cells covering the land surface of the Netherlands were made with a nonparametric method [quantile regression forests (QRF)]. A parametric alternative [zero-adjusted gamma distribution (ZAGA)] corrected the precipitation forecasts of the short-range Grand Limited Area Model Ensemble Prediction System (GLAMEPS) up to +60 h less well, particularly at high quantiles, as verified against calibrated precipitation radar observations. For the subsequent multivariate restructuring, three empirical methods, namely, ensemble copula coupling (ECC), the Schaake shuffle (SSh), and a recent minimum-divergence sophistication of the Schaake shuffle (MDSSh), were tested and verified using both the multivariate variogram skill score (VSS) and the continuous ranked probability score (CRPS), the latter after aggregating the forecasts spatially. ECC and MDSSh were more skillful than SSh in terms of the CRPS and the VSS. ECC performed somewhat worse than MDSSh for summer afternoon and evening periods, probably due to the worse representation of deep convection in the hydrostatic GLAMEPS compared to reality. Overall, the high-resolution postprocessing comparison shows that skill for local precipitation amounts improves up to the 98th percentile in both the summer and winter season and that the high-dimensional joint distribution can successfully be restructured. Forecasting products like this enable multiple end users to derive their own desired aggregations.