The Regime Dependence of Optimally Weighted Ensemble Model Consensus Forecasts

Previous methods for creating consensus forecasts weight individual ensemble members based upon their relative performance over the previous n days without regard to the underlying flow regime. A post-processing scheme in which model performance is linked to underlying weather regimes could improve the skill of deterministic ensemble model consensus forecasts. Here, Principal Component Analysis of several synoptic and mesoscale fields from the North American Regional Reanalysis dataset provides a means for characterizing atmospheric regimes objectively. A clustering technique that includes a genetic algorithm is developed that uses the resulting principal components to distinguish among the weather regimes. Consensus forecasts are created based upon 48-hour surface temperature predictions produced by the University of Washington Mesoscale Ensemble, a varied-model (differing physics and parameterization schemes) multianalysis ensemble with eight members. Different optimal weights are generated for each weather regime to complete the new method. Consensus forecasts obtained by the regime-dependent scheme are compared using cross-validation with traditional n-day ensemble consensus forecasts for locations in the Pacific Northwest.