Recurrent neural networks for rainfall-runoff modeling of small Amazon catchments

The insufficiency of hydrological time series prejudices the management of water resources in the Amazon, especially in small catchments. Thus, for the first time, through the Nonlinear Autoregressive Recurrent Neural Network with Exogenous Inputs (RNN-NARX), an attempt is made to simulate daily streamflow time series with a temporal resolution of 365 days, for five small Amazon catchments. A sensitivity analysis of the models was also performed to identify the lowest temporal resolution of the input to obtain satisfactory results. Since it is data-driven model, it is expected that these models have the ability to reproduce learning characteristics with less temporal variability, and make it possible to estimate daily streamflow time series with 365-day temporal resolution. For this objective, daily lagged rainfall and streamflow data were implemented with the support of the Cross-Correlation Function (CCF) and partial autocorrelation function (PACF) at the 5% significance level. Based on five statistical criteria, satisfactory results were obtained with supervised training based on 2 years of rainfall and streamflow data in four of the five analyzed basins (Igarape of Prata, Piranhas River, Caete River and Capivara River). According Garson's algorithm, lagged rainfall is important for these simulations. In general, there are lower percentage errors in the dry periods, and overestimation of floods. In the practical context, the models developed and analyzed are applicable, mainly, for the simulation of average and minimum daily streamflows of small catchments in the Amazon, becoming a tool that should be used for sustainable evaluation purposes of the water availability of these catchments. However, in the case of simulating floods, it is necessary to apply hourly and lagged rainfall and streamflow data to the models.