Estudo de métodos de previsão de séries temporais aplicados ao preço da energia elétrica no mercado de curto prazo brasileiro

Abstract

This work contributes to a comparative study of the performance of time series predictive methodologies in forecasting the price of electricity in the Brazilian short-term market. Exponential smoothing methods, ARIMA and SARIMA, models of artificial neural networks Multilayer Perceptrons (MLP) with univariate and multivariate inputs were applied in predicting the PLD (settlement price for the differences) of the energy submarket in the Southeast/Center-West region of Brazil. For the multivariate Multilayer Perceptrons model, the most relevant predictor variables for PLD prediction were selected. The performance of the models was quantified through prediction error evaluation metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE) and by evaluating the performance in relation to the naive prediction by half of Theil's U coefficient. The results show the prediction superiority of the multivariate MLP model in relation to the others. This model also has a reasonable ability to forecast the trend of PLD values for more advanced periods.