Otimizando a previsão de cheias sazonais para o Rio Negro utilizando redes neurais de propagação direta

Abstract

This paper proposes a new method for forecasting the maximum seasonal amplitude, using feedforward neural networks and, as input variables, climatic indices and the river amplitude measured a few months earlier before the maximum amplitude be verified. A new method for selecting the most relevant prediction variables is proposed. For neural networks training, two methods for improving its generalization are used: early stop and regularization. The best prediction result is obtained with two input variables, resulting in a correlation prediction coefficient of 𝑟𝑝=0,755.