Calibrating 2 m Temperature Forecast for the Regional Ensemble Prediction System at NMC

It's known that ensemble forecasts provide a flow-dependent sample of the probability distribution of possible future atmospheric states instead of the single and deterministic prediction. Ideally, the probability of any event could be skillfully estimated directly from the relative event frequency in the ensemble. Unfortunately, although the quality of ensemble prediction systems (EPS) has been improved greatly, the direct output of EPS generally is subject to the systematic deficiencies, especially for surface variables. They are under-dispersive and lack of reliability. Therefore, statistical post-processing methods have been developed to improve direct model output. The nonhomogeneous Gaussian regression (NGR) is used to calibrate 2 m temperature forecast of the regional EPS at NMC/CMA. The NGR is the statistical correction method with the first and the second moment (mean bias and dispersion) for Gaussian-distributed continuous variable. This method is based on the multiple linear regression technique and provides a predictive probability density function (PDF) in terms of the normal distribution. The method of minimum continuous ranked probability score (CRPS) estimation is used to fit the regression coefficients of PDF. It can be found that NGR method can greatly improve 2 m temperature forecast compared with the raw ensemble output, and the improvement is as follows: The mean bias is reduced and the spread of ensemble members is increased reasonably; the L-shaped Talagrand diagram of the direct ensemble output has been improved and the calibration reduces the number of outliers, especially in the 9th bin; the probabilistic scores (the brier score, continuous ranked probability score, area under relative operating characteristic curves) all show the significant forecast skill improvement in the calibrated forecasts.In addition, the sensitive study is performed to investigate the effect of the training length, and the results show that the training length plays a minor role, at least for the chosen verification period. Finally, the comparison by using the time-decaying average bias correction method and NGR is performed, showing that NGR not only has advantages in reducing ensemble mean bias and increasing ensemble spread, but improves the forecast skill in terms of probabilistic scores.