Ensembles of numerical weather prediction (NWP) models are used to predict the range of possible future atmospheric states, and the corresponding forecast uncertainty. There are many different ways to configure NWP ensembles. While there is currently no agreement on a single best method to configure ensembles, several studies indicate that an ensemble should account for both initial and lateral boundary condition error and model error. A common way to account for model error is through a multiphysics approach – with ensemble members using different combinations of physical parameterization schemes. The multiphysics approach raises the issue of how best to choose a few combinations of physics schemes from the large number of possible combinations. This study examines objective statistical post-processing methods to down-select a smaller multiphysics ensemble from a large one, for the purposes of adequately representing model error in an ensemble.

We build a 42-member multiphysics ensemble using the WRF-ARW model. Forecasts are run for 48 h every fifth day from December 2009 to November 2010. Because wind energy applications require that ensemble spread should accurately represent uncertainty in low-level mean wind, verification focuses primarily on boundary layer winds. Various approaches for down-selection are examined, including principal component analysis and k-means clustering, both in a univariate and multivariate context. We then statistically dress, or calibrate, the ensemble probability density function using Bayesian model averaging. We also examine the length of training and verification periods required for these ensemble down-selection approaches.