Assessing combinations of artificial neural networks input/output parameters to better simulate daily streamflow: Case of Brazilian Atlantic Rainforest watersheds

Abstract

This paper presents an assessment of the feasibility of using artificial neural networks (ANNs) with different input and output configurations to simulate streamflows for Brazilian Atlantic Rainforest basins. Hydrological data consisted of potential evapotranspiration (PET), rainfall (P), and streamflow (including runoff and baseflow) daily series (32 years in extent) of 12 watersheds of the Itapemirim River Basin (IRB). As a novelty, the input parameters for the ANNs were chosen basing on the statistical correlation between rainfall and flow data. Daily rainfall amounts until the third antecedent day were strongly correlated to runoff and streamflow. Accumulated rainfall until the 90 antecedent days were strongly correlated to baseflow and streamflow. Also as novelty, the ANNs were trained, for the downstream watershed, considering three different input and output configurations. The 1st configuration (ANNP) simulates daily streamflow considering input variables only related to rainfall. The 2nd (ANNP-PET) also simulates daily streamflow, but considers the differences between precipitation and potential evapotranspiration as input variables. The 3rd (ANN*P_streamflow) also considers input variables only related to rainfall, but the outputs were daily runoff and daily baseflow, which were used to later calculate total streamflow. Better streamflow simulations were provided by ANNP_streamflow, followed by ANNP-PET and ANNP. However, the three ANNs performances were quite similar. Thus, we indicate the use of ANNP for regions with lack of meteorological data (such as IRB) due to the need for fewer input parameters (only rainfall) and no need of splitting runoff and baseflow. After, the trained ANNs were applied to simulate the streamflow of each watershed. The results show that all the ANNs were able to satisfactorily simulate the streamflow of 11 watersheds, with better performances for the largest ones.