University of Washington Mesoscale Ensemble system post-processing and verification

The University of Washington Mesoscale Ensemble (UWME) system was designed to create probabilistic numerical forecasts of sensible near-surface weather parameters over the Pacific Northwest. Using a multi-analysis approach to initial condition selection, the eight-member UWME system produces probability forecasts for selected parameters and thresholds [e.g., P(2-m temperature < 0 °C) and P(10-m wind speed > 18 kt)] that are skillful compared to sample climatology through 48 h lead time (Eckel and Mass 2005). Probability forecast skill increases with the aid of an additive bias correction, which is applied to each member individually and trained on the previous 14 forecast cases. Further improvements are realized using a parallel eight-member system (UWME+) with physics diversity and perturbations to surface boundary conditions. New comparisons with NCEP's Short-Range Ensemble Forecast (SREF) system, which uses an error breeding approach to initial condition selection, may also be shown.

Despite being skillful, UWME and UWME+ probability forecasts are uncalibrated, since truth falls outside of the forecast ensemble range more often than expected (evidenced by u-shaped verification rank histograms). Even though bias correction improves probability forecast skill, the under-dispersion problem is not ameliorated. Member-wise bias correction actually reduces the ensemble spread and the ensemble spread-skill relationship in the process. Thus, several different ensemble post-processing techniques are investigated to see whether calibration can be achieved and probabilistic forecast skill improved, including: (1) dressing the ensemble mean with its forecast error distribution from a large sample of historical cases and (2) Bayesian model averaging (BMA) using training data within neighborhoods of similar land-use type and elevation.

The neighborhood BMA implementation produces the most skillful probabilistic forecasts of 2-m temperature at 48 h using a grid-based verification at 20-km horizontal resolution. Reliability diagrams and contingency table scores (hit rate, false alarm rate, relative economic value, etc.) for a specific forecast problem at a sample location are used to demonstrate the relative performance of these competing local forecast methods.