Premium
Statistical postprocessing of dual‐resolution ensemble precipitation forecasts across Europe
Author(s) -
Gascón Estíbaliz,
Lavers David,
Hamill Thomas M.,
Richardson David S.,
Ben Bouallègue Zied,
Leutbecher Martin,
Pappenberger Florian
Publication year - 2019
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.3615
Subject(s) - ensemble forecasting , probabilistic logic , histogram , computer science , downscaling , ensemble learning , weighting , resolution (logic) , brier score , grid , statistics , precipitation , range (aeronautics) , meteorology , mathematics , artificial intelligence , geography , image (mathematics) , medicine , materials science , geometry , composite material , radiology
This article verifies 1‐ to 10‐day probabilistic precipitation forecasts in June, July, and August 2016 from an experimental dual‐resolution version of the European Centre for Medium‐Range Weather Forecasts (ECMWF) ensemble prediction system. Five different ensemble combinations were tested. These comprised subsets of the 51‐member operational ECMWF configuration (18‐km grid) and an experimental 201‐member lower‐resolution configuration (29‐km grid). The motivation of the dual‐resolution ensemble forecast is to trade some higher‐resolution members against a larger number of lower‐resolution members to increase the overall ensemble size at constant overall computational cost. Forecasts were verified against precipitation analyses over Europe. Given substantial systematic errors of precipitation forecasts, both raw and post‐processed dual‐resolution ensemble predictions were evaluated. Postprocessing consisted of quantile mapping, tested with and without an objective weighting of sorted ensemble members using closest‐member histogram statistics. Reforecasts and retrospective precipitation analyses were used as training data. However, the reforecast ensemble size and the dual‐resolution ensemble sizes differed, which motivated the development of a novel approach for developing closest‐member histogram statistics for the larger real‐time ensemble from the smaller reforecast ensemble. Results show that the most skilful combination was generally 40 ensemble members from the operational configuration and 40 from the lower‐resolution ensemble, evaluated by continuous ranked probability scores, Brier Scores at various thresholds, and reliability diagrams. This conclusion was generally valid with and without postprocessing. Reliability was improved by postprocessing, though the improvement of the resolution component is not so clear. The advantages of many members at higher resolution was diminished at longer lead times; predictability of smaller scale features was lost, and there is more benefit in increasing the ensemble size to reduce sampling uncertainty. This article evaluates only one aspect in deciding on any future ensemble configuration, and other skill‐related considerations need to be taken into account.