STATISTICAL POSTPROCESSING OF 1-14-DAY PROBABILISTIC FORECASTS FOR COLD EXTREMES OVER TAIWAN

Two statistical post-processing methods, ensemble Model Output Statistics (EMOS) and Ensemble Kernel Density MOS (EKDMOS), are applied in 20-year reforecasts of the National Centers for Environmental Prediction (NCEP) global ensemble forecast system version 12 (GEFS v12) to produce calibrated and downscaled 1-14-day probabilistic forecasts of cold extremes at specific stations over Taiwan. To generate an EMOS forecast, the MOS equation is built using the ensemble mean, and applied to each ensemble member. The EKDMOS uses a kernel density estimation (KDE) to create a probability density function (PDF) from the EMOS forecasts.

Calibration is performed using a leave-one-out cross-validation procedure, where one winter is used for validation, and the remaining 19 winters are used for training. Forecast evaluation shows that the EMOS is under-dispersive, just like the raw ensemble (RawEns) forecasts, with some bias removed. In contrast, the EKDMOS well represents the forecast uncertainty with most of the bias removed. Compared to the RawEns or EMOS, the EKDMOS obviously improves the reliability and discrimination of probabilistic forecasts. The EKDMOS increases the Brier skill score (BrSS) of the RawEns and EMOS by decreasing its reliability, and increasing its resolution components. For any threshold and any lead time, users with a wider spectrum of cost/loss ratio can obtain more benefit from the EKDMOS as compared to the RawEns or EMOS. The EKDMOS distribution, as expected from a reliable forecast system, necessarily approaches the climatology of the training sample when forecast informativeness is lost beyond 10 days.