Predictability Of The Madden-Julian Oscillation In The Subseasonal-to-seasonal Prediction Models

In this study,We compute Bivariate Anomaly Correlation Coefficient(ACC)and root-mean-square error(RMSE)to estimate the prediction skill of the subseasonal-to-seasonal(S2S)database and we calculate the amplitude error((?))and phase error((?))of each model.In addition,the sensitivity of MJO prediction skills of S2 S models to the season of initial conditions and the intensity of MJO signals is analyzed.And two multimodel sets are constructed by using these models,and the prediction skills of multi-model ensemble prediction are analyzed,which shows the advantages of multi-model ensemble prediction.And the reason of the difference of MJO prediction skill of each model is explored.Finally,the potential predictability of MJO is estimated using both the signal-to-noise ratio(SNR)-based and information-based methods to explore how much room there is for improving the prediction skills of MJO.The main conclusions are as follows:1)We compute Bivariate Anomaly Correlation Coefficient(ACC)and root-mean-square error(RMSE)to estimate the prediction skill for each model of the subseasonal-to-seasonal(S2S)prediction project,the MJO prediction skill of the individual models in the S2 S database is a range of 8 to 32 days using 0.5 as a threshold of correlation coefficient for the skillful prediction.Most of the amplitude errors((?))and phase errors((?))in S2 S are almost negative,indicating that the amplitude of the mode prediction is generally smaller than the actual situation and the model underestimate MJO's moving speed.The prediction skill is also sensitive to its initial conditions.The skill is better when the prediction is initialized in winter or with strong MJO amplitude.2)The quality of ensemble prediction system will cause the differences of MJO prediction skill.Good ensemble prediction system,i.e.,ensemble spread close to the RMSE,has better prediction skill.The ensemble spread of ECMWF model with the highest prediction skill is almost equal to the RMSE,which indicates that the prediction system of this model is designed best.Through the construction of multi-model set and the comparison of the prediction skills of single model contained in the multi-model set and the multi-model set,it is found that the construction of multi-mode set can effectively improve the prediction skills of MJO.3)To explore the capacity and room of improvement of the MJO prediction,we investigate the MJO potential predictability and its features.The potential predictability of S2 S models is analyzed using SNR-based and information-based method respectively.It is found that the potential predictability differs a lot between different models.The potential predictability is much larger than actual prediction skill,for instance,S2 S models still have a lot of room of improvement for MJO prediction.Compared with the information entropy method,the SNR method underestimates the potential predictability of MJO.The strength of MJO signal has little effect on the potential predictability,which is different from the prediction skills discussed above.The possible reason is that the difference of initial conditions has little effect on the model imperfection.