Previsão regionalizada de vazão sazonal utilizando redes neurais artificiais.

Abstract

Many river basins have discontinuous and/or small hydrological series, raising the hydrometric data demand in these basins. There are lacks of information in the semiarid reservoirs related to seasonal forecasts of outflow or drained volume. This dissertation contemplates Piranhas- Acu and Apodi river basins, located in the northeastern Brazil with great importance for the region, with the use of the seasonal precipitation forecast. The basins with hydrometric data were used in the establishment of a regional model to estimate the seasonal daily average outflow in the basins without hydrometric data. The precipitation forecast and the physiographic characteristics of the basins (draining area, length of the main river, mean slope of the river, density of draining) were used in this process. The Artificial Neural Nets (ANNs) technique was used and its results were compared to a Multiple Regression Model developed in previous researches. The ANN showed a good performance when compared to Multiple Regression Model. Relations between the model adjustment quality and the physical characteristics of the sub-basins were noticed. This result shows the need of a data evaluation study with the use of representative samples of the set of basins for the Neural Net calibration. Both hydrological models showed good performances in the seasonal outflow forecast, despite the fact that the precipitation forecasts still contain high level of uncertainty. It was evidenced that the accumulated uncertainties, propagated to the outflow forecasting stage, had been attenuated in the transformation process of rainfall in outflow through the hydrological basins.