Neural networks input-based models to calibrate the mean precipitation of an ensemble prediction system

• Ano de defesa: 2016
• Autor(a) principal: José Roberto Motta Garcia
• Orientador(a): Rafael Duarte Coelho dos Santos, Christopher Alexander Cunningham Castro
• Banca de defesa: Marcos Gonçalves Quiles, Nandamudi Lankalapalli Vijaykumar, Carlos Henrique Quartucci Forster, Fábio Dall Cortivo
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Instituto Nacional de Pesquisas Espaciais (INPE)
• Programa de Pós-Graduação: Programa de Pós-Graduação do INPE em Computação Aplicada
• País: BR
• Link de acesso: http://urlib.net/sid.inpe.br/mtc-m21b/2016/08.05.13.14

Although ensemble forecasting systems provide richer forecasts by adding probabilistic concepts to single deterministic forecasts, they have intrinsic shortcomings caused by the lack of full comprehension of the relationship between meteorological variables. It is especially noticed in medium and large-scale forecasts, whose effects of chaotic behavior of the atmosphere drastically increase as the target forecasting date is lengthened. Improvements on weather forecasting systems can be done either by the meteorology staff concerning physical aspects of weather behavior as well as by implementing computational statistical methods in order to tune the weather forecasting model output. The purpose of this work is to compute, along the forecast horizon, a more accurate precipitation value than the ensemble mean precipitation by post-processing INPE/CPTEC's ensemble prediction output. To achieve the goal, some prognostic fields and derived data are combined and submitted as explanatory variables to an artificial neural network system. Experiments were guided in an exploratory way such that several computational models were generated and thereafter assessed. The study was individually performed at some grid points located within the boundaries of La Plata Basin. Results indicate that the application of this methodology presented values closer to actual values when compared to the ensemble mean precipitation. It also shows that the inclusion of the ensemble mean precipitation itself, as well as data from adjacent grid points, improve the calibration process of the target grid point. In addition, the exploratory approach detects different artificial network models to fit specific location and lead-time. Although this input-driven system computes less than ideal forecasting values, it performs better than the mean output of the ensemble model, which is widely used in various weather forecasting products.