Previsão da variabilidade do gelo marinho no setor antártico sul-americano com o emprego de redes neurais artificiais
Resumo
Resumo: O objetivo deste trabalho é avaliar o desempenho da Rede Neural Artificial (RNA) de função de base radial (RBF) na previsão da extensão do gelo marinho no Setor Antártico Sul-Americano. Para isso, levou-se em consideração a importância da seleção de variáveis que possuem maior influência sobre a variabilidade do gelo. Sabendo-se disso, através de uma revisão sistemática, foram encontrados possíveis preditores para o modelo de previsão, que passaram por uma posterior seleção, através do método stepwise com passo forward, afim de se obter o melhor conjunto de variáveis de entrada para a rede. Para cada região e cada configuração de previsão proposta, a rede RBF foi desenvolvida e avaliada, apresentando diferentes resultados para cada uma delas. Comparando-se ao desempenho obtido pelo método de previsão por persistência, os resultados das previsões através da rede RBF semostraram satisfatórios, sobretudo, para horizontes de previsão de até 3 meses para o Setor Antártico Sul-Americano e para a região da Península Antártica, com resultados menos satisfatórios para a região do Estreito de Drake. As variáveis de precipitação e temperatura da superfície do mar no Mar de Bellingshausen foram as que apareceram mais vezes nos experimentos de previsão da rede, indicando maior importância para os resultados. Abstract: The objective of this work is to evaluate the performance of the Radial Basis Function (RBF) of Artificial Neural Network (ANN) in the prediction of sea ice extent in the South American Antarctic Sector. Therewith, the importance of selecting variables that have the greatest influence on ice variability was considered. Knowing that, through a systematic review, possible predictors were found for the forecast model, which underwent a subsequent selection, through the stepwise method with forward step, in order to obtain the best set of input variables for the network. For each region and each proposed forecast configuration, the RBF network was developed and evaluated, presenting different results for each one of them. Comparing to the performance obtained by the persistence forecasting method, the forecast results through the RBF network proved to be satisfactory, especially for forecast horizons of up to 3 months for the South American Antarctic Sector and for the Antarctic Peninsula region, with less satisfactory results for the Drake Strait region. The variables of precipitation and sea surface temperature in the Bellingshausen Sea were the ones that appeared more frequently in the network prediction experiments, indicating greater importance for the results.
Collections
- Dissertações [78]