Probabilistic Forecast Of The Surface Air Temperature And Precipitation Of East Asia Based On Bayesian Model Averaging

Owing to the nonlinear effect of atmospheric motion, the uncertainty of initial field and the numerical model, it is necessary to provide probabilistic forecast which provide the full information of the future state of the atmosphere. It is all this uncertainties that request the transformation from single deterministic forecast to probabilistic forecast. Therefore, it is key to quantify the uncertainty to improve the forecast skill. By using the TIGGE data from European Centre for Medium-Range Weather Forecasts (ECMWF), National Centers for Environmental Prediction (NCEP), United Kingdom Meteorological Office (UKMO) and Japan Meteorological Agency (JMA) and based on the Bayesian Model Averaging (BMA) method, a probabilistic forecast of the surface air temperature and24hours accumulated precipitation of both single model and multi-model of east Asia (85°-145°E,10-45°N) is carried out. The data used here is daily forecast product of four ensemble prediction systems from June1st,2007to September30th,2007, and the resolution of the data is1°×1°. BMA is a post-processing method to combine the individual models to provide a more skillful results of the forecast. The computation of the parameters is done by the Expectation-Maximization algorithm. A normal distribution is applied to the forecast of surface air temperature, and the distributions of the precipitation occurrence and the cumulative precipitation amount was represented by Gamma distribution simultaneously. The results shows that BMA method can produce probabilistic forecast effectively and provide the quantitative uncertainty information. By applying BMA to the interest variables could reduce the uncertainty. A comparison of BMA and ensemble is also represented, which shows that the probabilistic forecast of BMA shows better result than that of ensemble mean.The probabilistic forecast of surface air temperature based on BMA of single and multi-models, shows that the multi-model probabilistic forecast is more skillful than that of single model. And the precipitation probabilistic forecast shows that the best training period length is about30days. The comparison of BMA and ensemble mean have their own advantages respectively, and both of them could be referred while forecasting.