Key Technology Research On Ensemble Prediction Of Climate System Model

Parameter uncertainty quantification approaches are used to reduce parameter uncertainty and improve the simulation skill of climate system model.However,the application of current popular evolutionary algorithms in complex climate system models requires long time and high computational cost.For cost expensive climate system models,fast and effective parameter optimization methods need to be further studied.This paper proposes a set of parameter optimization methods based on multilayer perceptron surrogate model targeting single objective optimization,multi-objective optimization and constrained optimization.The evaluation results of the proposed optimization algorithms and the commonly used optimization algorithms with complex mathematical functions and single-column atmospheric modes show that the proposed algorithms have overall advantages in accuracy and convergence.With the complex single column atmospheric model,the convergence rate of the proposed multi-objective optimization method can be improved by more than 5 times compared with the known NSGAIII method.The initial condition perturbation method of ensemble is of great significance for the study of reducing the initial uncertainty.However,the currently used Lagged Average Forecasting(LAF)method lacks a strong theoretical foundation.This paper proposes a Breeding of Growing Mode(BGM)method for climate prediction,based on the BGM method for weather prediction.And the key techniques in BGM method,such as initial perturbation generation and breeding cycle length,are reconstructed.The reconstructed method is more adaptable to obtain the fastest growing perturbation in climate prediction.This paper compares the proposed BGM method with the LAF method currently used by the National Climate Center in the BCC-CSM climate system model.15-years hindicast results show that the BGM method is significantly better than the LAF method in the prediction of most climate variables in the first month.The 500 hpa potential height is improved by 10% relative to the LAF method in term of RMSE(root mean square error).The improvement effect of some variables can be extended to four months.Finally,this paper designs a new climate ensemble method(BGMOPT)for climate prediction,by integration of the proposed parameter optimization method and the BGM,which is compared with the LAF method in the BCC-CSM model.With BCC-CSM model,the Madden-Julian oscillation(MJO)and the East Asian summer monsoon(EASM)are taken as the objectives,the radiation balance at top of model is used as the constraint.The improved BCC-CSM model with optimized parameters is used for ensemble prediction.The results show that the BGMOPT method performs well in the climate simulation experiments.The four-month global precipitation is about 15% better than the LAF method under the mean square error.Furthermore,this paper proposes a machine learningbased ensemble result improvement method to generate optimal deterministic forecast results for important forecast indicators.This method combines the characteristics of observation and model output data to correct and integrate the ensemble prediction results of the climate system model.In the prediction of the El Ni?o/Southern Oscillation(ENSO),this method can improve the sea surface temperature RMSE by 32% relative to the ensemble average method.