Predictability Of Interdecadal Variability Of Surface Air Temperature In Mid- And High-latitude Eurasia And Improvement Of Prediction Skill

In recent years,decadal climate prediction has become a hot topic in the field of climate research.The Eurasian surface air temperature(SAT)experiences multidecadal variabilities under global warming background,and its effective decadal prediction has important social and economic value.Based on latest CMIP6 Decadal Climate Prediction Projection(DCPP)models and ERA5 reanalysis datasets,this thesis evaluates the predictability of the seasonal mean SAT in four seasons with different models,explores the sources of predictability of SAT in Eurasian(SATI),and tries to improves the model data by statistical and different artificial intelligence methods.The main conclusions are summarized as follows:(1)The assessment results show that the skills of SAT are lowest in winter,significant in spring,highest in summer and surprised in autumn.There is almost no skill can be observed over Eurasia for the winter SAT in multi-model ensemble(MME)and rest of the models.The predictive skills of the SAT are mainly derived from the predictive abilities of the trends.All the winter SATIs predicted by models with different advance times show consistent and significant warming trends,but the skills to predict decadal variability are limited.At the same time,all four seasons surface climates are well predicted in broad tropical regions and Atlantic.All models have limited forecast levels in the North Pacific.(2)By exploring the predictable sources of SAT,the predictive skills of SAT linear trend are mainly due to the influence of external forcing.After the removal of linear trend,the regions with skills in models mainly benefit from initialization.The warming trends persist in subsequent seasons,which are beneficial to the prediction of spring and summer,but limit the prediction skills of autumn.After subtracting the winter SAT skill,there is almost no skill left in other season predictions.Therefore,the skills of the SAT in CMIP6 DCPP models generally dependent on the winter SAT prediction.Moreover,it is found that MME have better predictive skills for AMO,which are significantly present in all seasons,but the skill decreases in summer with the increase of lead time.The skills of PDO are concentrated in autumn,and sources of skills are mainly reflected in the prediction of the warming trends.At the same time,the abilities to predict atmospheric circulation are still limited,and the decadal oscillations of winter AO and NAO are not captured.Although the model reproduces the seasonal continuous influence of AMO and PDO well,it is considered that poor prediction quality of PDO is a possible reason that affecting SATI prediction skills,with the contribution of AMO which has high model prediction quality to SATI prediction being not significant.(3)The improved results of variance adjustment and machine learning methods show that,when the key SST significant region is used as the characteristic factor of the machine learning models,the revised results based on LSTM and GRU neural networks can better reproduce the decadal changes and phase turning points that are consistent with observations,especially the SATI in winter has the most obvious improvement,which is of great significance to the improvement of model data.The Random Forest and Light GBM models change the original warming trend of the models,but the predictive skills scores are far less than the model itself.The improved results of the Prophet model are better fitted to SATI variability on the basis of maintaining the trend of model warming,which is conducive to improving the forecasting skills in spring and summer,but the improvement ability in winter and autumn is limited.The modified results based on variance adjustment can improve the prediction skills of SATI in winter to a certain extent,but the improvement effect is not obvious in other seasons.