Interpreting, Improving, and Augmenting Multi-Model Ensembles

Abstract : Multi-Model numerical weather prediction (NWP) ensembles outperform even the world's most sophisticated single-model ensemble prediction system. An ensemble whose members are the control forecasts of a number of operational NWP centers, the poor-man 5 multi-model (PM MM) ensemble, samples from both initial condition space and model space. This proposal addresses two methods for intelligently adding new ensemble members to multi-model ensembles. The first method utilizes multi-model ensemble forecasts in a lagged average manner. Ensemble sizes are increased by combing ensemble forecasts launched at different times but valid at the same verification time. To avoid the limitations inherent in combining ensemble forecasts that utilize different numbers of observations, an ensemble transform Kalman filter (ET KF) is utilized to incorporate observations into existing ensemble forecasts without re-running any NWP models. The transformation conditions all ensemble members on the same number of observations. The second method is to use a simplified model to mimic the behavior of the PM MM ensemble. Parametric perturbations are used to produce states for the simplified model that populate the tails of the PM MM ensemble increasing the PM MM ensemble spread.