| El Ni(?)o-Southern Oscillation(ENSO)is an interannual signal of air-sea interaction in the tropical Pacific Ocean,and the strongest climate variability in the global climate system.Accurate prediction of the El Ni(?)o-Southern Oscillation(ENSO)is crucial to climate change research and disaster prevention and mitigation.The First Institute of Oceanography,Ministry of Natural Resources developed the second-generation shortterm climate prediction system FIO-CPS v2.0(First Institute of Oceanography-Climate Prediction System version 2.0).The prediction skill(the Anomaly Correlation Coefficient,ACC)can exceed 0.75 at a 6-month lead time,but the prediction skill fell rapidly for 7-to 9-month lead times,and even below 0.45 at one-year lead time.At the same time,the prediction skill of ENSO has declined significantly since 2000 compared with the end of the 20 th century.Machine learning can find bias rules from a large number of model and observation data,construct bias correction model,and improve model simulation and prediction skills.The use of machine learning methods is expected to play an important role in correcting the bias of ENSO in FIO-CPS v2.0.Based on the ENSO prediction results from FIO-CPS v2.0 and the machine learning method—XGBoost(e Xtreme Gradient Boosting),this paper proposed the idea of correcting ENSO from the two sources of the initial bias(the bias caused by inaccurate initial values)and the intrinsic bias(the bias caused by the numerical model itself)of the prediction system,The dynamical and statistical hybrid prediction model FIO-CPS-HY(First Institute of Oceanography-Climate Prediction System-Hybrid)was established.Considering the deficiencies of the whole hindcast performance of FIOCPS v2.0 in the past 40 years and after 2000,the correction validation experiments of the whole hindcast results of FIO-CPS v2.0 and the independent test correction experiments were carried out,respectively.The main conclusions are as follows:1.An ENSO hybrid prediction model FIO-CPS-HY was established to conduct research ideas’ validation on FIO-CPS v2.0 whole hindcast and bias correction experiments on prediction after 2000,respectively.The validation experiments of the whole hindcast results of FIO-CPS v2.0 were conducted.The 40-year(starting from1982 to 2021)Ni(?)o3.4 Sea Surface Temperature(SST)hindcast results by FIO-CPS v2.0 were randomly divided into a training set(70%)and a test set(30%),combined with the initial bias and the intrinsic bias as input features of the XGBoost model,the Ni(?)o3.4 observational SST was used as the truth to train a hybrid prediction model and ultimately used in the whole hindcast results correction.The independent test correction experiments of FIO-CPS v2.0 were conducted.Aiming at the shortcomings of ENSO prediction skills after 2000,the FIO-CPS v2.0 hindcast results were divided into the training set(starting from 1982 to 2006)and the test set(starting from 2007 to 2021)in chronological order.To avoid the impact of climatology selection on the test set,the SST anomalies,the initial bias anomalies,the intrinsic bias anomalies in the Ni(?)o3.4area of the prediction system were input into the XGBoost model for training and used in a subsequent independent test.2.The validation experiments of the whole hindcast and correction experiments of independent test of FIO-CPS v2.0 were carried out by using the established ENSO hybrid prediction model FIO-CPS-HY,and showed good ENSO prediction skills.The latest 40 years(starting from 1982 to 2021)of hindcast results showed that ACC and root mean square error(RMSE)of the Ni(?)o3.4 index using FIO-CPS-HY could be greater than 0.96 and within 0.24℃ for 1-to 13-month lead times,respectively.And the ACC and RMSE from April to June exceeded 0.9 and was within 0.24℃,respectively.FIO-CPS-HY also showed the ability to predict ENSO events based on the further evaluation of the 40-year hindcast results.The independent test correction results show that the ACC and RMSE of the Ni(?)o3.4 index from FIO-CPS v2.0 to FIOCPS-HY for 7-to 13-month lead times could be increased by 57.80%(from 0.40 to0.63)and reduced by 24.79%(from 0.86℃ to 0.65℃),respectively.FIO-CPS-HY can improve prediction skills to a level comparable to the early stage of FIO-CPS v2.0(starting from 1982 to 2006),which proved that the hybrid prediction method was feasible and effective.3.In order to understand the physical features of hybrid prediction models,the input features of the hybrid prediction system are further analyzed.The results showed that the prediction results of FIO-CPS v2.0 are the most important,followed by the intrinsic bias,and the initial bias’ contribution is the least.As the lead time increased,the correction of ENSO by the hybrid prediction model required the contributions of the intrinsic bias and the initial bias,which also confirmed the importance of improving the physical process of the dynamical model,the parameterization schemes and the assimilation prediction technologies.Although FIO-CPS-HY still has some biases in real-time prediction,this study provides possible ideas and methods for enhancing short-term climate prediction capabilities,and demonstrates the potential of integrating machine learning and numerical models in climate research and applications. |
