Jared A Lee, Haupt SE and Young GS
The goal of ensemble down selection is to retain the subset of ensemble members that span the uncertainty space of the forecast while eliminating those that are most redundant. There are hundreds of combinations of physics schemes that can be used in typical numerical weather prediction (NWP) models. Limited computational resources, however, force us to constrain the size of NWP ensembles, and to choose what combinations of physics schemes to use. Ensemble down selection can help guide those choices, and also yield information about how many ensemble members are necessary. In this study we examine the use of hierarchical cluster analysis (HCA) as an objective down selection technique. To test the performance of HCA across multiple seasons, a 42 member multi physics ensemble is configured and run, with 48 h forecasts initialized every fifth day for twelve months. HCA is performed on forecast errors of low level temperature and wind components over training periods of one, two, and three months. How the ensemble members cluster is found to change by season. The full and subset ensembles are then calibrated using Bayesian model averaging (BMA). The uncalibrated and calibrated ensembles are verified over one month periods. Statistical tests indicate a likelihood that the subset ensemble comes from same distribution as the full ensemble, and have verification scores nearly the same as the full ensemble. Furthermore, intelligently down selecting a subset ensemble with HCA outperforms random down selection.