Study On Seasonal To Interannual Climate Prediction With Multi-Model Ensemble Method

Climatic disasters such as frequent floods and droughts often cause significant damage to human life,national economy,and public safety.In response to disaster prevention,it is quite important to make effective seasonal-to-interannual prediction.However,errors are always inevitable even in the most exceptional climate models in the world.Errors or uncertainties grow with the increase of forecast time which declines the seasonal-to-interannual climate prediction skills.How to improve the accuracy of seasonal-to-interannual climate prediction is an important issue in scientific research and operational forecast due to prominent socioeconomic demands.Statistical error correction methods is an effective way to improve prediction skills.This study mainly focuses on the assessment and new correction methods of seasonal-to-interannual prediction with a set of current operational climate prediction models.Prediction skills of the monsoon,precipitation and sea surface temperature(SST)on seasonal-tointerannual timescale are comprehensive evaluated.Based on the idea of dynamicstatistics correction and the coupled ocean-atmosphere general circulation models,we construct a variety of dynamic-statistics prediction schemes from the angle of stepwise pattern projection,analogue correction,linear inverse model and deep learning technique.These error correction methods are also combined with multi-model ensemble(MME)technique to improve precipitation and SST prediction on seasonalto-interannual scale.The main conclusions are as follows:1?A comprehensive evaluation of the monsoon,precipitation and SST forecasts on seasonal-to-interannual timescales is performed.In reference to different start seasons,Season-reliant Empirical Orthogonal Function(S-EOF)method is applied to capture the seasonal independent modes of the Asian-Australian Monsoon(AAM)interannual variability.The predictions generally perform better for the SEOF1 than SEOF2 in terms of both spatial and temporal variations.The predictability of models is closely related to ENSO(El Ni?o-Southern Oscillation).The first SEOF mode concurs with the turnabout of warming to cooling in the ENSO no matter which season it started.The second mode starting from JJA(June-July-August)and SON(September-OctoberNovember)leads the warming/cooling by about 1 year,signaling precursory conditions for ENSO.Whereas the second mode starting from DJF(December-January-February)and MAM(March-April-May)are during the decaying phase of ENSO.The evaluation also shows that the prediction skills of precipitation and SST forecast are limited with increasing lead months.2?The prediction skill of two types of ENSO have been improved through the stepwise pattern projection method(SPPM)and development of an analogue-based correction method(ACE).Since ENSO behaviors have become increasingly diverse,the prediction skill after 2000 appeared to be declined.To improve ENSO prediction skill,the stepwise pattern projection method is applied.The result shows that the ENSO prediction is improved after correction.The differences between Eastern Pacific(EP)and Central Pacific(CP)events are always more pronounced in corrected MME forecasts compared with uncorrected ones.The analogue-based correction method is developed,in which forecast errors in models can be corrected empirically using historical forecast errors.The verification shows that forecasts of ENSO indices are improved relative to the uncorrected model forecasts.Compared with the other analogue-based methods,the proposed method in this study is more advantageous than other results.3?An error correction method based on Linear inverse model(CLIM)is proposed,in which forecast errors in models can be predicted empirically using historical forecast errors,which are calculated by the same model.Based on the second generation climate system model of Beijing Climate Center,BCC-CSM1.1M,the correction method is investigated.The independent validation results show that precipitation prediction skill have improved globally after correction.In addition,if SST information is included in the error correction method,the precipitation prediction skill in southern China will be improved more obviously.The error correction methods have also been used to improve ENSO prediction.After correction,the skill over the tropical Pacifica have been significantly improved.The MME experiments are also conducted based on several models of North American Multi-Model Ensemble(NMME)project.It shows that the CLIM can improve the skills for both of the SST and precipitation.After removing errors of each model member by using CLIM,the MME always maintains higher skill.4?The error correction method based on Deep Learning technique(DLEC)is developed.Based on convolutional neural network(CNN)and Deep Residual Network(Res Net),we come up with DLEC to improve the prediction skill of SST and precipitation.The method can significantly improve global precipitation prediction,especially in the tropical Pacific.The method,in which loss function is the RMSE over East Asia,is better at predicting regional precipitation skill in China.The results also show that the MME combining with the DLEC improves the precipitation skill obviously,especially over northeast China,and the maximum improvement is over 0.4.After correction,the SST prediction skill is improved.The improvement gets more obvious with increasing lead months.The corrected MME has high Ni?o3.4 index skill,with a correlation coefficient of 0.9 at 6-month lead and over 0.7 for up to a 11-month lead time.