Previsão de séries temporais no sistema elétrico brasileiro utilizando preditores baseados em aprendizado de máquina: uma análise empírica

Abstract

The overview of electric energy in Brazil is influenced by a variety of complex factors and nonlinear relationships, making forecasting challenging. With the increasing demand for energy and growing environmental concerns, it is crucial to seek solutions based on clean and renewable energy practices, aiming to make the energy market more sustainable. These practices aim to reduce waste and optimize the efficiency of processes involved in the operation of electricity distribution and generation technologies. A promising approach to enable sustainable energy is the application of forecasting techniques for various variables in the energy market. This thesis proposes an empirical analysis of the use of regressors to make predictions in the databases of the Price of Settlement Differences (PLD) in the Brazilian market and wind speed in wind turbines in Northeast Brazil, through principal component analysis. We aim to provide significant information about machine learning techniques that can be employed as effective tools for time series prediction in the electric sector. The results obtained may encourage the implementation of these techniques to extract knowledge about the behavior of the Brazilian energy system. This is particularly relevant, given that energy prices often exhibit seasonality, high volatility, and peaks, and wind power generation is widely influenced by weather conditions. To model the prediction of these two time series, we use the database on the Price of Settlement Differences (PLD), focusing especially on the average energy price of the Brazilian National System. The most relevant variables are related to hydrological conditions, electrical load, and fuel prices for thermal units. For collecting variables related to wind energy, two distinct locations in the Northeast region of Brazil were considered: Macau and Petrolina. For the prediction study, we use a Multilayer Perceptron Neural Network (MLP), a Long Short Term Memory (LSTM), Auto-Regressive Integrated Moving Average (ARIMA), and Support Vector Machine (SVM) to determine baseline results in prediction. To enhance the results of these regressors, we employ two different prediction approaches. One approach involves combining deep artificial neural network techniques based on the Canonical Genetic Algorithm (AG) meta-heuristic to adjust the hyperparameters of MLP and LSTM regressors. The second strategy focuses on machine committees, which include MLP, decision tree, linear regression, and SVM in one committee, and MLP, LSTM, SVM, and ARIMA in another. These approaches consider two types of voting, voting average (VO) and voting weighted average (VOWA), to assess the impact on the performance of the machine committee.