Improvement of Multi-model Ensemble Forecasting using Recursive Bayesian Model Process

The advantage of using multi-center, multi-model ensemble forecast system has been demonstrated by many previous works. The North America Ensemble Forecasting system (NAEFS) is a successful example (Candille, 2009). The NAEFS simply combines the two ensemble forecasting systems (Global Ensemble Forecast System (GEFS) of the National Weather Service (NWS) and Canadian Meteorological Center Ensemble (CMCE) of the Meteorological Service of Canada (MSC)) using an equal weight for each model. A post-processing algorithm named as Recursive Bayesian Model Process (RBMP) has been developed to apply multi-model ensemble forecasts. The method is mainly based on Bayesian Model Averaging (BMA) (Raftery et al. 2005). We adopted the station-based BMA codes developed in MDL to global grid-based codes to calibrate probability distributions for each grid in the global ensemble forecasts. The method of BMA is to estimate the weights for each model and a uniform variance based on most recent model performance. In the RBMP, The two parameters are recursively updated using the decaying averaging technique in order to reduce storage space in operational forecasting. We also apply a 2nd moment adjustment technique to the BMA-calibrated forecasts in order to improve over/under-dispersive of ensemble forecasts.

The RBMP was applied to NUOPC forecasts of 2-m temperature and 10-m winds for the summer and fall of 2013. NUOPC combines the three global ensemble forecast systems (GEFS) from the NWS, MSC, and FNMOC. For 2-m temperature, the method is the efficient which improves ensemble forecast skill for all lead time with a maximum improvement for short lead-time forecasts. For 10-m winds, the RBMP only slightly improve the first moment adjustment for short-lead forecast although the improvement for the second moment adjustment is perfect. The RBMP is being applied to other seasons (winter and spring). The result will be also presented.