Postprocessing of precipitation forecasts for new configured NCEP short-range ensemble forecasting (SREF) system

Since December 2005, the NCEP SREF system has been updated by constructing new ensemble components, including 21 members from WRF-based models (two dynamic cores – ARW and NMM), ETA, and RSM models. This system contains perturbed initial/boundary conditions, multiple physics, and multi-model. Results of precipitation forecasting from this suite of new configured operational SREF system were verified compared to NCEP Stage IV precipitation analyses. Calibration of probabilistic quantitative precipitation forecasts (PQPF) was conducted over 212 AWIPS grid domain (~ 40 km), covering the CONUS. According to hydrological and geographic characteristics, bias correction of PQPF was performed for three major regions – western, central, and eastern U.S. The tool used to conduct probabilistic postprocessing is linear regression method and a feed-forward artificial neural network. Improvements by two calibration techniques are comparable in the bias correction processes of the SREF PQPF. The coefficients/weights in the linear regression model and the neural network were trained for warm season (April-September) and cool season (October-March) respectively, and were applied to calculate PQPF during validation periods. Cross-validation was used to compute verification scores and attributes diagrams. The reliability of PQPF was improved after calibration for low thresholds. The sample size for heavy precipitation events poses great challenge to achieve effective calibration and stable weights, especially for high thresholds over the relatively dry western region.