Investigating Seasonal Impacts on Clustering and Ensemble Down-selection

Numerical weather prediction (NWP) ensembles are valuable tools for a number of probabilistic forecasting applications. Imperfect knowledge of the model physics is one of the sources of uncertainty in NWP models. A common way to account for this uncertainty is to construct a multi-physics ensemble. NWP models typically have dozens to hundreds of possible combinations of physics schemes. NWP ensembles are expensive to run, however, so only a subset of possible combinations can be used. Furthermore, it is desirable to avoid running ensemble members that provide essentially redundant information.

To determine a more efficient method to glean ensemble information, a 42-member multi-physics ensemble is built with the Weather Research and Forecasting (WRF) Advanced Research (ARW) model. Forecasts are run for 48-h periods every fifth day over a span of 12 months. Verification focuses on temperature and wind components at low levels (surface, 925, 850, and 700 hPa) because of relevance to wind energy and atmospheric transport and dispersion forecasting applications.

Hierarchical cluster analysis (HCA) provides an objective way to down-select from a large ensemble to a smaller ensemble, while still retaining most of the forecast information from the large ensemble. By examining how the ensemble members cluster in different seasons, we determine that certain classes of physics schemes contribute more or less to forecast uncertainty in different seasons. These seasonal differences highlight some important considerations for modelers when building ensemble prediction systems that incorporate model physics uncertainty.