Rainfall-runoff models can help the management of water resources, especially in the Amazon, a region marked by the low density of hydrological monitoring, and thus benefit the multiple uses of water and the adequate use of water resources. This work seeks to simulate daily streamflows of five small catchments in the Amazon, through the Autoregressive Recurrent Nonlinear Neural Network with Exogenous Variable (RNN-NARX). Daily rainfall and streamflow data were used for simulation. The cross-correlation and partial auto-correlation functions helped to determine lagged data, relevant inputs, with a significance level of 5%. In addition, the Levenberg-Marquardt error backpropagation algorithm was used for supervised training of RNN-NARX. Five statistical indices and Garson's relative contribution of each input variable were also used to evaluate the simulations. Thus, the simulated flows were classified between unsatisfactory and very good, in addition to showing a general tendency to underestimate floods. The autoregressive characteristic of each catchment is fundamental for better results, quality attributed to the water storage capacity. A plausible explanation for the main sources of uncertainty is due to the spatial variability of precipitation between monitoring stations and the precipitations occurring in the catchment, meteorological anomalies and discretization aspects. The sensitivity analysis of the models against different training intervals showed that the implementation of 2 years, for the supervised training of the RNN-NARX, is sufficient to obtain efficient simulations in four of the five small Amazon catchment analyzed.
|
