A Comparison of Optimal-score-based Correction Algorithms of Model Precipitation Prediction

Based on data from national meteorological stations, one year quasi-symmetrical mixed running training period (QSRTP), and precipitation prediction from CMA (T639), ECMWF, NCEP, JMA, both optimal threat score (OTS) method and optimal equitable threat score (OETS) method are designed to conduct a comparison experiment on correction algorithms for model precipitation with frequency matching (FM) method. Through classification correction, three methods are used merely to calibrate model precipitation amount with the predicted rain-belt location and shape kept unchanged. The OTS method figures out correction coefficients of different precipitation classes by optimizing threat score (TS) of corrected precipitation within training period. OETS is similar to OTS but achieved by optimizing ETS. Correction experiments are conducted twice a day with forecast time at 0000 UTC and 1200 UTC, respectively. To consider seasonal background, 20 days before the forecast day and 20 days after the same day in the previous year are adopted to constitute training period. For each national meteorological station, there are 80 samples in total. The correction experiment shows that for either precipitation products of ECMWF, JMA, NCEP, CMA, or their ensemble mean, both OTS and OETS show much better performance than FM in 24 h accumulated precipitation classification calibration with different lead time according to traditional verification methods like TS and ETS. In particular, OTS is the best and can improve precipitation prediction in all lead times. After correction, both OTS and OETS incline to forecast larger precipitation area than observation for most classes but less precipitation amounts. Compared to FM, both methods tend to produce a little higher false alarm rates in middle and low classes, which is much less than the reduced missing rate, thereby leading to a higher threat score. In terms of ECMWF correction, OTS and OETS have a relatively stable Bias score of 1.1, although there are much fewer samples in high class. By contrast, FM produces an unstable Bias score, especially in maximum class with score over 2.2, indicating an excessively high missing rate. As for stable equitable error in probability space (SEEPS), OTS has superiorities over all lead times, all single models and multi-model mean. Furthermore, TS of corrected ECMWF precipitation using OTS method in 2015 are also better than subjective forecast from all aspects, with national averaged threat score of 1 d rainstorm forecast reaching 0.194.