An evaluation of post-processed TIGGE multimodel ensemble precipitation forecast in the Huai river basin

SummaryThis paper evaluates how post-processing can enhance raw precipitation forecasts made by different numerical weather prediction (NWP) models archived in TIGGE (THORPEX Interactive Grand Global Ensemble) database. Ensemble Pre-Processor (EPP), developed at U.S. National Weather Service, is used to post-process raw precipitation forecasts. EPP involves several major steps: (1) deriving the joint distributions of raw forecasts and observations corresponding to different canonical events; (2) obtaining the probability distributions of observations given the raw forecasts; and (3) constructing ensemble forecasts from the conditional probability distributions given the raw forecasts. Raw precipitation forecasts from five NWP models (CMA, ECMWF, JMA, NCEP and UKMO) during the summer-fall period (rainy season) from 2007 to 2011 were evaluated over the Huai river basin in China. The lead time for the precipitation forecasts is set to 9days, which are divided into 11 canonical events (defined as daily precipitation events or aggregate precipitation events over a period of several consecutive days). Our experiments show that post-processed precipitation forecasts shows substantial improvement over the raw forecasts. Post-processing reduces both the biases and the root mean squared error of the raw forecasts significantly. In terms of ensemble spread, both the Brier skill scores and continuous ranked probability skill score are improved appreciably after post-processing. Reliability diagrams and rank histograms also confirm that post-processed ensemble forecasts possess better ensemble spread property compared to the raw forecasts. Among the five NWP models, ECMWF and JMA have the best overall performance in both raw and post-processed forecasts. The raw and post-processed UKMO and NCEP forecasts outperform other models in certain events. Post-processing can improve the CMA raw forecasts substantially, but still its performance is consistently worse than that of the other models