GUEDES, R. V. S.; http://lattes.cnpq.br/7184807680256037; GUEDES, Roni Valter de Souza.
Résumé:
Drought affects the natural environment of an area when it persists for a longer period. Thus, the prediction of drought plays a major role in planning and resource management systems and water resources in a river basin. In the last decade, the methodology based on Artificial Neural Networks (ANN) has shown great skill in modeling and forecasting time series nonlinear and nonstationary. This work applies the methodology of ANNs for forecasting time series of Standardized Precipitation Index (SPI) in the Epitácio Pessoa river basin dam - PB and evaluates their effectiveness. The study area (Epitácio Pessoa river basin dam) is located in the semiarid region of Paraíba State. In the modeling process, subsequent changes were made to the Neural Network configuration in order to obtain a model with the smallest possible error. The feed-forward back propagation Neural Network had one of the best performances, with a two-layer structure and learning algorithm of Levenberg-Marqualdt. Of the 26 gauge stations studied within the basin, the proposed model for prediction of regression showed values above 88% and mean square error below 0.223. The forecasts are more efficient for larger time scales of SPI, in the short term. It was found that as an increase the time horizon reduces the accuracy of the forecast. The results indicate that forecast less than three months is considered satisfactory. For longer terms is necessary to improve the learning process of the Neural Network.
