GALDINO, J. F.; http://lattes.cnpq.br/6553559281081240; GALDINO, Jurandir Ferreira.
Résumé:
This work aimed to model and generates monthly average streamflows data from monthly rainfall based on the mathematical technique of Artificial Neural Networks. The methodology was applied to data – pluvio-hydrometric of Paraguaçu Basin River - BA, located in the state of Bahia in northeastern Brazil. Were used 14 rain gauges distributed spatially within the basin and one stream station called Argoim. The backpropagation feed-forward network, chosen as the best to represent the rainfallrunoff modeling, used in its architecture the algorithm Backpropagation of Levenberg - Marquardt with 30 neurons in the hidden layer, and transfer function hyperbolic tangent sigmoid (tansig) in the intermediate layers and output of the model. Of the results obtained in this study, the best correlation was the function (tansig1) 0.999, with regression coefficients of 0.99, 0.997 and 0.996. In the forecast for two years, the average regression was 0.999 in the training phase, in the validation phase 0.982 and 0.995 in the test phase. Already the average MSE of the 15 stations was 25.519, considering the higher peak flows. When assessing the performance of the whole training process of ANNs with the lowest MSE was possible to identify the best performance for RNA 4. This RNA has the highest coefficient of efficiency 0.99 and lower REMQ equal to 5.051 m³/ s.
