SOUSA, K. S.; http://lattes.cnpq.br/3131994564250882; SOUSA, Katherine da Silva.
Resumo:
The development of water-related engineering projects generally requires the existence of reliable hydrological data. In ungauged watersheds, obtaining flow data becomes a challenge that may be overcome by the use of flow regionalization techniques. In this study, models based on multilayer perceptron artificial neural networks (ANN) were applied for the regionalization of average monthly flows in river basins without systematic hydrological monitoring. The methodology consisted of estimating monthly average flows based on hydro-meteorological and physiographical data. Data from 20 watersheds located in Rio Grande do Sul State, Brazil, were used for calibrating the ANN models. The model with the best calibration performance, considering the adjusted coefficient of determination, was tested in five new watersheds located within the study area. The best performance was achieved by the ANN model with 16 neurons in the hidden layer. When analyzing the results of Nash-Sutcliffe coefficient for the test data set, it was observed that the estimates for all assessed watersheds were considered acceptable. The results found in this case study suggest that this methodology has potential to transfer information from one watershed to another within a region with similar hydrological behavior.