Assessing Electricity Prices and Their Driving Mechanisms in Brazil with Neural Networks

Abstract: In general, electricity prices are very volatile and derive from many external variables. In Brazil, this price is determined by computer models developed and operated by government organizations. The supply and demand relationships are not enough to determine prices in Brazilian submarkets. Due to the particularities of the predominance of hydroelectric production in the country and many regulatory factors, electricity prices in Brazil carry a high level of uncertainty to be managed by market participants. The Brazilian electricity Settlement Price is defined by the composition of three models: NEWAVE, DECOMP and DESSEM; for long-, mid- and short-term predictions, respectively. The prices are based on the Operational Marginal Cost, which those models aim to minimize especially by outputting the cheapest hydrothermal operational settings that can attend the electricity demand. To minimize the prices uncertainty, this research proposes investigating the feasibility of developing a predictive model supported by the time series Machine Learning technique, the Long Short-Term Memory (LSTM). This tool is part of a theoretical framework called Recurrent Neural Networks (RNNs). The raw material for this work is the combination of literature on the history of the Brazilian Energy Market and its particularities, in addition to studies on Neural Network technologies and LSTM applications, as well as real historical data related to electricity price in the country. Accordingly, this work compiles data from June 2001 to April 2023, weekly and by submarket, which represents the input variables of the proposed model. The product of this work revolves around a predictive model programmed in Python with support from the Keras library, capable of predicting 4 weekly prices ahead. In addition, a comparative analysis is registered between the results of the LSTM and DECOMP models, which is the one already widely used on the Brazilian market. For this evaluation, performance indicators were used on the assertiveness of the predicted absolute values, the direction of the predicted price, and the predicted volatility. The results show that the LSTM model was significantly more accurate with respect to direction and volatility and less accurate with respect to the absolute values of the predicted prices.