Artificial Neural Networks–Based Model Parameter Transfer in Streamflow Simulation of Brazilian Atlantic Rainforest Watersheds

Abstract

This paper presents an assessment of the calibration and transfer of artificial neural networks (ANNs) to simulate streamflow at Brazilian Atlantic Rainforest basins. Primary data consisted of rainfall and a streamflow daily series (32 years in extent) of 12 subbasins of the Itapemirim River basin (IRB). First, data from three subbasins were used to adjust three ANNs to estimate daily specific streamflow from input parameters related to rainfall. After, the ANNs were applied to simulate the flows in all other IRB subbasins. The ANNs were able to reproduce the subbasin discharges for which they were adjusted. They also reached satisfactory performance when applied in most of the other subbasins. The obtained results demonstrate that the ANN technique is a viable alternative for simulating flows in regions lacking primary data for hydrological modeling. Besides, calibrating ANNs with subbasin data of an intermediate size or position tends to present a better overall performance than calibrating for the smaller (upstream) or the larger subbasins (downstream).