Efficiency of MLP architecture on closed-loop mode for the simulation of a hydrological system

ABSTRACT Estimatives of hydrological responses are needed for the watershed planning. The aim of this study was to evaluate the hydrological behavior simulation of the Upper Canoas basin using artificial neural networks Multi Layer Perceptron (MLP) method, as well as to analyze the contribution of the input variables for modeling. It were tested 12 treatments with combinations of variables such as precipitation, evapotranspiration (ET0) and discharge, as well as transformations and temporal displacements of these variables, in order to determine the variables that promoted the better performance on discharge modeling. The MLP was trained in open-loop mode using part of the observed discharges. The discharges for the whole series were simulated in closed-loop, using the discharge simulated on the previous time step as input. The learning algorithm used was the Levenberg-Marquardt. The treatment with the best performance (NS = 0.9119, RMS = 14.29 m3/s) employed the daily precipitation of the four rainfall stations (Urubici, Vila Canoas, Lomba Alta e Anitápolis), precipitation of the four stations with -2 days of response time, and simulated discharge from the previous day. Despite the low RMS, the modeled discharge using MLP was generally overestimated.