Bias Correction in Monthly Means of Temperature Predictions of the Dynamic Extended Range Forecast Model

Bias analysis of hindcast data from the monthly Dynamic Extended Range Forecast model (DERF2.0) indicates that the monthly mean temperature prediction performance is not good enough to be used operationally and needs to be improved. Quantile mapping of the non-parameter method is applied to correct the DERF2.0 model bias of monthly mean temperature. The first key step of this method is to utilize deterministic model output and calculate the cumulative density function (CDF) of hindcast data. Then, we can obtain the quantile mapping of the deterministic model output on the CDF of hindcast data. The second step is to calculate the CDF of observation data and map the quantile result of model output to the CDF of observational data. The model bias can be reduced to a certain extent after the above procedures. Hindcast verification shows that the method can significantly reduce the root mean square error (RMSE) of model output and improve the predictive skill of spatial distributions of monthly mean temperature anomalies (1983-2012). Prediction skill (PS) and anomaly correlation coefficient (ACC) scores between model output and observations with different lead times have been improved, and this improvement remains stable with different magnitudes of model bias. Comparison of model prediction skills of different months shows that the enhancement of prediction performance in summer is greater than in winter.