Forecast and analysis of rainfall over south and southeastern Brazil using artificial neural network / Previsão e análise sobre as regiões sudeste e sul do Brasil utilizando redes neurais artificiais

AUTOR(ES)
DATA DE PUBLICAÇÃO

2003

RESUMO

This study uses an Artificial Neural Network (ANN) technique to establish a non-linear relationship between the large scale atmospheric circulation and local surface rainfall. The method involves the use of statistical downscaling applied to outputs from Eta model. In this sense, prognostic equations were developed for 18 locations using the ANN. This method uses as predictors numerical weather products from the Eta model and surface rainfall as predictand. The objective is to generate site-specific quantitative forecasts of daily rainfall. Also, prognostics models are constructed to delineate rain areas having as predictors outputs of the global model T126 and as predictand maps of rain derived from the Tropical Rainfall Measuring Mission (3B42/TRMM). The selection of ANN input variables is based on the prevailing synoptic weather conditions over southeastern Brazil. It is shown that during the austral summer the main weather systems related to rainfall are: cold fronts, and the South Atlantic Convergence Zone (SACZ) interacting with cyclonic vortices at upper tropospheric levels (CVULs). During the austral winter, cold fronts and midlatitude upper levels cyclonic vortices are the main weather systems associated to rainfall. Several statistics are calculated to examine the performance of the models. It is found that during summer periods the skill score indicates an ANN improvement over Eta model by 50 %. In the winter period ANN improves RMSE in 80% respect to Eta model. Overall, ANN is efficient in predicting continuous rainfall periods associated to cold fronts and SCAZ during the summer and rainfall events associated with cold front and CVUL originating from middle latitude in winter. Also during winter, the ANN is more efficient, because the synoptic systems are better defined by the variables derived from Eta model. For the area forecast the images are pre and post processed with a wavelets transform, in order to minimize training time. The area forecasts utilizing ANN show in summer a pattern similar to that observed the 3B42/TRMM. In spring ANN shows centers of rain over the continent that are not observed by the 3B42/TRMM. On the other hand, the forecast of T126 has a tendency for generating areas of rain in the oceanic which are not depicted by 3B42/TRMM. When the ANN is not able to reproduce a pattern similar to the one obtained by 3B42/TRMM, the results reflect a combination of the 3B42/TRMM and T126.

ASSUNTO(S)

wavelets meteorologia precipitação ondaletas redução de escala precipitation(meteorology) redes neurais artificiais previsão do tempo inteligência artificial artificial neural networks metereologia sinóptica satélite trmm previsão de chuva trmm satellite rainfall forecasting (weather forecasting) backpropagation synoptic meteorology retropropagação meteorology weather forecasting artficial intelligence downscaling

ACESSO AO ARTIGO

Documentos Relacionados