Artificial Neural Networks for Filling Missing Streamflow Data in Rio do Carmo Basin, Minas Gerais, Brazil

Abstract

Adequate availability of data directly influences the quality of hydrological studies. In this sense, procedures for filling gaps of observations are often applied in order to improve the length of hydrological series. One technique that can be used is the Artificial Neural Network (ANN), which process information from input data creating an output. This study aims to evaluate the application of ANN to fill missing data from monthly average streamflow series at Rio do Carmo Basin in the state of Minas Gerais, Brazil. A 26-years series (from 1989 to 2012) was used for ANN modelling while the two proceeding years, 2013 and 2014, were used to simulate failures pursuant to evaluating the performance of the ANN. The ANN construction was performed by the software WEKA that uses the multilayer perceptron model with sigmoidal activation functions. Four types of ANN were generated: five attributes and two (MLP1) or five (MLP2) neurons; and with three attributes and one (MLP3) or three (MLP4) neurons. The best-fit model to ANN was the MLP1, verified by Pearson correlation coefficients (0.9824), and coefficient of determination r² (0.9646). The model used five attributes, four input data (year, month, streamflow data from Acaiaca and Fazenda Paraíso stations) and one output data (streamflow from Fazenda Oriente station), that considered the temporal variation of streamflow. Hence, the utilization of the ANN generated by the WEKA was adequate and can be considered a simple approach, not requiring great computational programming knowledge.

Keywords:
data consistency; data mining; hydrologic estimation