Study Of The Bias-Correction And Multi-model Combine Of Mesoscale Ensemble Forecast

As a new development dynamics stochastic forecast technique, ensemble forecast, especially the super-ensemble forecast, become more and more important in the numerical forecast field. But because of the different model and parameterization schemes of physical processes used in the multimodel ensemble forecast system, the systemic bias of ensemble member of are quite different with each other, which is against with the principle of "equal-likelihood". Bias-correction should be down before they are combined. Based on 2m temperature forecast products from the mesoscale ensemble forecast models of NMC/CMA,MRI/JMA, NCEP,MSC and Meteo -Fr.& ZAMG, a study of the bias-correction(both of system error and ensemble spread correction), and multi-model super-ensemble forecast is conducted.First of all, remove the forecast bias by the way of the adaptive (kalman filter type) algorithm. The result shows that all the forecast skill of these various could be improved, either the mean of ensemble forecast or probabilistic forecast. The PDF of each forecast member are much more similar with the others. The crps of the corrected forecast is smaller than the raw forecast. The spread of forecast system is increased reasonably, and closer to the rms, which means that the forecasts are much more reliable. What's more, the talagrand diagram representing for the reliability of the ensemble probabilistic forecast, the roc and EV representing for the resolution of the ensemble probabilistic forecast and the scores of the ensemble probabilistic forecast such as bs / bss, all of these indexes prove the improvement of the corrected ensemble forecast.Then, two multi-model ensemble forecast with the ensemble member are 70 and 11 respectively are constructed. By contrasting the result with the MRI/JMA, the most excellent of the five single ensemble forecast systems, we found that even with the same number of ensemble member, multi-model ensemble forecast still overmatch the single ensemble forecast. The system error is reduced, and ensemble spread is reasonably adjusted. It proves that multi-model ensemble forecast constructed with the forecast result of different numerical forecast center can make up for the uncertainty of the initial disturbance and numerical forecast model. So, is there a optimal amount of ensemble number exist? The result shows that the forecast ability is enhanced with the increment of the ensemble member obviously at the beginning, and then the increase rate slow down, so much as no improvement anymore. For the sake of economizing compute resource, we set the optimal amount of ensemble number to be 40 in the multi-modle mesoscale ensemble forecast. For the ensemble mean forecast of the multi-modle ensemble system, three kinds of mathematic method are used to combine the single model ensemble mean forecast, which are the arithmetic average, multiple linear regression and BP artificial neural network method. The result shows that the consensus ensemble mean forecasts have the best perform, the one of multi-model take the second place, and the ones of single model are the worst .And among all of the consensus ensemble mean forecasts, the consensus forecast by multiple linear regression and BP artificial neural network are much more precise than the one by arithmetic average.