Impact Of Initial Perturbation On Meso-scale Model Uncertainty

In recent years, computer power resources have greatly increased, thereby enabling the operational use of numerical forecasts with higher resolution and more detailed description of topography. Nevertheless, it's still difficult to get the satisfactory results about the severe weather and the long-range forecasts. Lorentz's discovery led to the emergency of a new discipline, dynamic theory, and to the realization that many apparently deterministic systems, like the atmosphere and its numerical models, are also chaotic. But if we run an ensemble of forecasts from slightly perturbed initial conditions, then the average of the ensemble forecasts will filter out some unpredictable of forecasts and provide some guide on the real atmosphere. Now the ensemble forecasts have been a very popular technology of weather forecast. Several operational forecast centers like ECMWF, NCEP, CMC, Met-office and JMA et al run their own ensemble forecast systems daily. The products of ensemble forecasts play a very important role in weather forecasts. In CMA, the ensemble forecast is in the initial stage. The weather forecasters are not familiar with the ensemble forecast. And it's so difficult for them to extract the useful information from so many products.The task of this paper is focused on the generation of initial perturbation with two methods. Based on the previous studies, I construct and improve two medium-range limited-area ensemble forecast system—GRAPES-BGM and GRAPES-ETKF. Previous comparison of the skills of the two ensemble generation schemes, ETKF analysis perturbation is able to reflect the density and accuracy of the observations. The ETKF ensemble error variance is spread with comparable amount independent, orthogonal, and uncorrelated directions. BGM methods can only keep the sum of the perturbations equal to zero. But the perturbation is not orthogonal. The computational expense of ETKF is very small, but more than breeding ensemble.It's concluded from the study that:1. From the comparison of the threat skill, brier skill and the other validation, the results of BGM scheme are better than ETKF schemes.2. The spread and talagrand distribution provided that the effect of BGM is better. But it must be connected with the addition of model perturbation.3. I examine the mean eigenvalue spectra of ensemble covariance matrices. The spectrum of the ETKF eigenvalues is flatter than from the breeding methods. The ETKF ensemble would eventually maintain all amplifying normal modes. The breeding methods only maintain error variance in the direction corresponding to the most rapidly amplifying normal mode. This result is the same with the previous study. 4. The ensemble mean and poststamp tell us the two ensemble system improved the intensity and location of the heavy precipitation in different levels. Some members estimate the location of the severe weather well. It is a good guide for weather forecasters.5. The two ensemble system differently increased the level of prediction of intensity and location of heavy rain. Overall, we can see that the breeding method do well in the intensity prediction, however, ETKF schemes predicted the location and range well. And the breeding method is good at improving the range of the heavy rain.6. The addition of the perturbation of boundary conditions does help increasing the spread. It improves the skill of precipitation, too.7. The cluster analysis downscaled 50 ECMWF ensemble members to 16 representative members. The 16 members provide initial and boundary conditions for ECMWF/ALADIN system. The pair and notpair ECMWF/ALADIN system have better results than operational ensemble system which didn't include clustering process.