Estratégias de predição de preços do mercado livre de energia por redes neurais artificiais e filtragem estocástica

Abstract

The Settlement Price of Differences (PLD) is used as a basis for the sale of electrical energy in the short-term market. The knowledge of its future values and trends mitigates insecurity in decision making, allowing market agents to decide the most appropriate strategies and establish successful contracts, maximizing profits and minimizing the risks of their ventures. PLD is, however, influenced by several factors linked mainly to the uncertainties of demands and hydrology as well as social and political factors, which gives it high volatility, seasonality, saturation of values and stochastic behavior. In this way, the present work proposes an investigative approach for the prediction of future trends of PLD through Artificial Neural Networks and Stochastic Filtration, with the motivation to assist the decision process of purchasing energy in the short-term market. Such an approach consists of combining a recurrent neural network with an output filter to attenuate the noise inherent in the data, helping to preserve future data trends. The network was trained using the ADAM algorithm and is applied to the Brazilian energy market. The choice of network input data was made based on statistical analyzes of cross-correlation between the time series of the variables of interest. The results show a good predictive capacity with a good monitoring of the trends of the variable of interest throughout the year, offering its main behaviors and trends to market operators, assisting them in making energy purchase and sale decisions.