Avaliação do desempenho de redes neurais artificiais na previsão de vazões críticas a partir de séries curtas de dados

Abstract

Hydroelectric facilities operation requires the use of models to predict streamflow. Several phenomenological models have been used, however they require many inputs and long historical series to be calibrated. Artificial Neural Networks (ANN) are an alternative to these models. The present study aims to evaluate the performance of neural networks which were calibrated using short time series in order to predict streamflow peaks. It was used a daily historical streamflow data from Chavantes Dam ranging from 1931 to 2007, totaling 77 years. A total of 15 ANNs were trained, varying the percentage of the training groups and forecasting horizons. The percentage of the data used in the calibration of the networks were 10, 20, 30, 40 and 50% and predictions were made for three distinct forecasting horizons, 1, 3 and 7 days. The performance of the ANN models were evaluated using 3 performance metrics: the root mean square error (RMSE), the mean absolute relative absolute error (MARE) and the coefficient of efficiency (CE). It was possible to realize that the best results were obtained when using smaller time horizons, however none of the networks were able to satisfactorily determine the streamflow peaks.