Choosing the Optimal Configuration of a Mixed-Physics Ensemble for Quantitative Precipitation Forecasts

Recent papers on ensemble forecast systems point to the need to account for model uncertainty along with initial condition/analysis errors. What, then, is the optimal ensemble configuration for modeling systemic error? One method that has been proposed is an ensemble in which each member uses different physical parameterizations or different combinations of these parameterizations.

Precipitation forecasts to 36 h from a nine-member MM5 mixed-physics ensemble were run during the spring of 1999 for 51 cases with the purpose of addressing the question of an optimal configuration. The ensemble members are constructed from the different combinations of three planetary boundary layer schemes and three convective parameterization schemes, all of which are in common use.

Various measures of ensemble and individual model performance are employed, including bias, equitable threat scores, rank-probability scores, relative operating characteristic curves, root-mean squared differences, and spread ratios. Results indicate that while the behavior of the individual members may follow conventional wisdom, aspects of the behavior of ensembles formed from different groupings of these members are counterintuitive. In particular, our results show that the inclusion of model configurations that produce inferior deterministic forecasts can improve performance of the ensemble.