Redes neurais diretas e recorrentes na previsão do preço de energia elétrica de curto prazo no mercado brasileiro

Abstract

There are few articles about short term electricity price prediction in the Brazilian market. Existing works use ARIMA predictors and feedforward neural networks however, without input selection or lag selection for these inputs. Besides, there is no work with use of recurrent neural networks in the Brazilian electricity market. The short term electricity market may show important opportunities for the agents acting as the commercialization in this market is less bureaucratic in relation to the long-term market.. This article shows the use of feedforward and recurrent neural networks (besides comparison with the ARIMA model) to predict short term electricity price with the use of correlation for exogenous input selection for the networks and also for lag selection to these inputs. It is shown that, for one step forward predictions, both implemented networks outperforms the ARIMA model, and in general, feedforward network works better than recurrent network. Besides, lag selection in the input improves feedforward network performance.