Ensemble tests of a regional climate model using a perturbed cumulus parameterization scheme

The purpose of this dissertation is to implement the original ensemble technique sampling initial condition problem into the regional climate model and to extend integration into the seasonal period. The seasonal integration has its own characteristics: not only the assumption of perfect model is no longer proper in the regional climate integration, but also sampling initial conditions is trivial when integration is beyond a certain period.; We perturb the model by varying two parameters in the convective scheme in the regional climate model in order to investigate the impact of model errors on the regional climate forecast. The results show that directly perturbing the model indeed establishes the relation of some qualities of ensemble forecast and model errors. It allows us to track which aspects of ensemble forecast the model errors impact. Decomposition of mean square error shows that the perturbed model has less influence on the overall average precipitation over a larger area. The perturbed model may affect the strength and intensity of local precipitation and account partially for the incorrect location of maximum Precipitation. It implies that either lateral boundary conditions or other uncertainties associated with model errors, such as PBL scheme, or both, may responsible for the inaccuracy. The results also suggest that probably it is not enough to compare with the results from other models by using only one deterministic RegCM2 run. The distribution of outlier statistic shows that nearly 50% of days are located outside of the range of the ensemble forecasts. The results may suggest that perturbing the PBL scheme simultaneously be necessary. The ROC curve shows that ensemble forecasts have skill to predict the annual precipitation anomaly. The results indicate as well that the probabilistic forecasts may have a more valuable than ensemble mean and are superior to the reference run.; In addition, analysis shows large values of both parameters will be safe to tune RegCM2. Model solutions will be more stable and have small root mean square errors. However, such cases show more systematic errors and relate closely to the occurrence of precipitation events. Our suggestion is to survey more uncertainties associated with model errors. The perturbed model can help us to narrow down the source of model error.