Study Of The Algorithm For Super-ensemble In Short-range Precipitation Forecast

With the developing of weather forecasting model, the uncertainty of the forecasting model raises an increasing attention. An ensemble forecast developed by using multiple parameterization schemes can eliminate the model's uncertainties effectively. While based on the different performance of each ensemble member, the super-ensemble algorithm may avoid reducing the forecast skill as what ensemble mean does by giving a different coefficient of each member.As it is difficult to make a good prediction of heavy rain in summer, especially in Meiyu season, a trial of building an effective ensemble product for short-term precipitation using the super-ensemble technology in different parameterization schemes was brought out. Making the assumption that the difference in short-range forecasts of precipitation arise largely from differences in the cumulus parameterization, we get an ensemble group with five members using WRF's different cumulus parameterization schemes. Four kinds of ensemble products have been conducted by utilizing the ensemble mean, the bias-removed ensemble mean, super-ensemble and the Empirical Orthogonal Function(EOF) super-ensemble algorithms. Both ensemble members and ensemble products were tested in a precipitation case in Meiyu season of 2013. To achieve an optimal ensemble product, the length of running training period has been tested for three kinds of super-ensemble algorithms. Compared with the single ensemble member, the ensemble products were verified by examining the root mean square errors, mean errors and correlation coefficient. The results show that forecasting skill of bias-removed mean, super-ensemble algorithm and EOF super-ensemble algorithm may improve by extending the training periods and the bias-removed mean ensemble has the highest quality. Meanwhile, the EOF super-ensemble algorithm seems to be more stable than the super-ensemble algorithm.Analyzing the deficiency of the super-ensemble algorithm, a stepwise regression was used to reject the unimportant ensemble member and an ensemble mean based forecast algorithm was used to reduce the dependency on training period. The performances influenced by training periods of both new ensemble algorithms were tested, which seem to be promoted by extent running training period. Compared with the super-ensemble algorithm, the performance of updated algorithms was proved better.