Recorded presentation4.3
Skill-based consolidation of multi-model ensembles using Bayesian Model Averaging (BMA) and ensemble regression
Dan C. Collins, NOAA/NWS/NCEP/CPC, Camp Springs, MD; and D. A. Unger and E. A. O'Lenic
Ensemble regression is used at the NOAA National Centers for Environmental Prediction (NCEP) Climate Prediction Center (CPC) in the experimental North American Ensemble Forecast System (NAEFS) to produce surface temperature probability density functions from bias-corrected model ensemble member surface temperatures. Ensemble regression determines the theoretical value of the standard error of the “best member” from the standard error of the ensemble mean and the standard error of individual ensemble members based on the assumption that each individual member forecast is equally likely to be the “best member”. When the skill varies from one member to another, some members will be more likely to be the closest to the observation for a given case than others. We propose a method of consolidation of multi-model ensemble forecasts using Bayesian Model Averaging (BMA) to obtain weights representing the likelihood that a single-model ensemble forecast is “best”. This technique uses a priori information to assign the likelihood that a model ensemble has the best forecast. These weights are applied to the individual ensemble member forecasts, and ensemble regression is used for consolidation, summing ensemble member kernels to determine the total probability distribution. As a proof-of-concept, this method is applied to consolidation of the Canadian Meteorological Center ensemble (CMCE) and NCEP global ensemble forecast system (GEFS) week-two forecasts from the North American Ensemble Forecast System (NAEFS), using prior CMCE and GEFS forecasts and observations to determine weights, which vary by region. Model skill differences may be relatively small for the week-two forecast; however, the method is applicable to consolidation of lagged-ensemble forecasts such as the NCEP Coupled Forecast System (CFS) monthly forecasts, assigning weights for each model lag time, which would have large differences in skill as lag time increases. Recorded presentation
Session 4, Ensemble Forecasting Including Post Processing IV
Monday, 21 January 2008, 4:00 PM-5:30 PM, 219
Previous paper Next paper