Extended Range Probabilistic Forecast Of Surface Air Temperature And Precipitation Using Bayesian Model Averaging

Based on the 10-15 days extended range ensemble forecasts of ECMWF, NCEP and UKMO in the TIGGE dataset, the probabilistic forecasts of 2 m surface air temperature and 24 hour accumulated precipitation were conducted by using BMA (Bayesian Model Averaging). In order to filter out the shortwave disturbance and further improve the prediction skills of extended range prediction, the moving average method was used both in forecasting data and observed data so as to research the change trend of 2 m surface air temperature and 24 hour accumulated precipitation of several days average in the extended range.The results showed that the dispersion of each single forecast system of surface air temperature was lower than ideal distribution. BMA could provide full PDF (Probability Density Function), and describe the forecast uncertainty quantitatively. The uncertainty and error of forecast on the land (higher latitude) was larger than that on the sea (lower latitude). Moving average methods improved the forecast skill of surface air temperature and the longer of moving step was, the better of forecast results were.Similar to the temperature forecasting, ensemble spread of precipitation was also too low. The Talagrand distribution of ECMWF showed "L" type roughly because of high alarm rate. BMA could provide probability of precipitation and full PDF of rainfall. The forecast error on the sea of low latitude and coastal areas was large, while the forecast error on the land of northwest was small. Moving average methods also reduced the forecast error and uncertainty of precipitation. In terms of different levels of precipitation, BMA could improve the forecast skill of light to moderate rain, but was insufficient to forecast the heavy rain.To make up the shortage of BMA method in the forecast of precipitation, "Frequency Matching" method, which used observed precipitation frequency to calibrate the precipitation frequency of each ensemble member, was applied to remove systematic forecast bias caused by model deficiency. Results showed that frequency matching method could effectively reduce the empty alarm and significantly improve the forecast skill of light rain and heavy rain, just was complemented with Bayesian Model Averaging.