Redes neurais para a previsão de preços de energia no mercado livre brasileiro
Data
2019-12-02
Autores
Santos, Nathália Rodrigues 
Orientadores
Maciel, Leandro dos Santos
Tipo
Trabalho de conclusão de curso
Título da Revista
ISSN da Revista
Título de Volume
Resumo
A energia elétrica é um insumo básico e essencial no mundo moderno. O Brasil se difere de países europeus devido a sua matriz energética com grande participação de fontes renováveis na geração de energia. Após a liberação do setor elétrico de energia no Brasil, a partir dos anos 90, e da separação das atividades de geração, transmissão, distribuição e comercialização de energia, atualmente, a contratação de energia no país está dividida em dois ambientes, o Ambiente de Contratação Regulada e o Ambiente de Contratação Livre. A presente monografia objetiva propor uma nova abordagem para a predição de preços de energia do Brasil, com base na utilização de modelos de Redes Neurais Artificiais (RNAs). Para isso, foram utilizados os dados de preços do sistema de livre negociação, no qual as cotações flutuam de acordo com as negociações entre os agentes, ou seja, com base na oferta e demanda. Os dados utilizados foram os Preços de Liquidação de Diferenças (PLD) no período desde maio de 2003 até julho de 2019. Como os preços são disponibilizados por submercados do sistema elétrico, foram consideradas quatro séries de preço de diferentes submercados: Sudeste, Sul, Norte e Nordeste. Os modelos competitivos de previsão, ARIMA, SARIMA, HoltWinters (HW) e RNA foram avaliados com base em diferentes métricas de erros para previsões um passo à frente e 19 passos à frente, representando curto e longo prazos, respectivamente. Em todas as regiões o modelo de RNA apresentou resultados mais acurados no curto prazo (um passo à frente), enquanto que no longo prazo (19 passos à frente), nenhum dos modelos resultou em bom ajuste. Entretanto, o modelo de suavização HW apresentou o melhor ajuste aos preços em alguns momentos com preços bem próximos aos reais quando as previsões de longo prazo são analisadas.
Electricity is a basic and essential resource in the modern world. Brazil differs from European countries due to its energy matrix with large participation of renewable sources in power generation. After the deregulation of the electric power sector in Brazil, since the 90's, and the separation of the activities of generation, transmission, distribution and commercialization of energy, the energy contracting in the country is divided in two environments, the Regulated Contracting and the Free Contracting Environment. This monograph aims to propose a new approach to the prediction of energy prices in Brazil, based on the use of Artificial Neural Networks models. To do so we used data from the free trade system, in which prices fluctuate according to the negotiations between agents, in other words, based on supply and demand. The data used were the Difference Settlement Prices (PLD) in the period from May 2003 to July 2019. As prices are available by submarkets of the electric system, we considered four different price series in four different regions: Southeast, South, North and Northeast. Competitive forecasting models, ARIMA, SARIMA, HoltWinters (HW) and ANN, were evaluated based on different error to forecast one-step ahead and 19 steps ahead, representing the short and long term respectively. In all regions the ANN model presented more accurate results in the short term (one-step ahead), although in the long term (19 steps ahead) none of the models had a good adjustment, however the HW model presented the best price adjustment at some times with prices very close to the real ones if we analyze the forecasts of the long term.
Electricity is a basic and essential resource in the modern world. Brazil differs from European countries due to its energy matrix with large participation of renewable sources in power generation. After the deregulation of the electric power sector in Brazil, since the 90's, and the separation of the activities of generation, transmission, distribution and commercialization of energy, the energy contracting in the country is divided in two environments, the Regulated Contracting and the Free Contracting Environment. This monograph aims to propose a new approach to the prediction of energy prices in Brazil, based on the use of Artificial Neural Networks models. To do so we used data from the free trade system, in which prices fluctuate according to the negotiations between agents, in other words, based on supply and demand. The data used were the Difference Settlement Prices (PLD) in the period from May 2003 to July 2019. As prices are available by submarkets of the electric system, we considered four different price series in four different regions: Southeast, South, North and Northeast. Competitive forecasting models, ARIMA, SARIMA, HoltWinters (HW) and ANN, were evaluated based on different error to forecast one-step ahead and 19 steps ahead, representing the short and long term respectively. In all regions the ANN model presented more accurate results in the short term (one-step ahead), although in the long term (19 steps ahead) none of the models had a good adjustment, however the HW model presented the best price adjustment at some times with prices very close to the real ones if we analyze the forecasts of the long term.
Descrição
Citação
SANTOS, Nathália Rodrigues. Redes neurais para a previsão de preços de energia no mercado livre brasileiro. 2019. Trabalho de Conclusão de Curso (Bacharelado em Ciências Econômicas) - Escola Paulista de Política, Economia e Negócios, Universidade Federal de São Paulo, Osasco, 2019.