Estimativa da probabilidade de ocorrência da precipitação, a partir de métodos estatísticos não paramétricos aplicados a simulações numéricas de um sistema de previsão por conjunto

Abstract

In this paper was used kernel density estimation (KDE), a nonparametric method to estimate the probability density function of a random variable, to obtain probabilistic precipitation forecasts from an ensemble prediction with the Weather Research and Forecasting (WRF) model. The nine members of the prediction system (EPSm) were obtained by varying only the convective parameterization of the model. The cases of study corresponded with heavy precipitation events in southern Brazil. In spatial assessment of the results, probability estimates obtained for 24 h periods for different rainfall thresholds were compared with precipitation estimated by the MERGE product, which combines precipitation estimates from the Tropical Rainfall Measuring Mission (TRMM) with surface observations. The performance of these probabilistic forecasts obtained with KDE did not prove to be better to the individual deterministic forecasts of the EPSm, or than the mean precipitation product (MPP) or the basic probability (PS) of the EPS. However, the local evaluation of the KDE product in places with observations of ANA stations obtained better results, and in general KDE forecasts with >25% and >50% of probability had better values of the skill scores (PC H, TS, F, B, PSS, CSS and HSS) that the EPSm individual forecasts and the MPP. These results seem to indicate that some of the KDE deficiencies found in the spatial assessment can be consequences of the comparison to the MERGE product, since is the local assessment against real observations, KDE results clearly superior to EPSm and MPP.