Previsão de demanda de energia elétrica em microgrid considerando níveis menos agregados por meio da aplicação de rede neural artificial GRNN combinada com o método estatístico SARIMA

• Ano de defesa: 2022
• Autor(a) principal: André Junior, Nelson Nunes
• Orientador(a): Rocha, Carlos Roberto Mendonça da
• Banca de defesa: Conceição, Katiani Pereira da , Maeda, Monara Pereira da Rosa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual do Oeste do Paraná; Foz do Iguaçu
• Programa de Pós-Graduação: Programa de Pós-Graduação em Engenharia Elétrica e Computação
• Departamento: Centro de Engenharias e Ciências Exatas
• País: Brasil
• Palavras-chave em Português: SARIMA; Redes neurais artificiais; Previsão de demanda; GRNN; LSTM
• Palavras-chave em Inglês: SARIMA,; Artificial neural networks; Demand forecast; GRNN; LSTM
• Área do conhecimento CNPq: CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO
• Link de acesso: https://tede.unioeste.br/handle/tede/6542

The growth of electricity consumption in the world requires countries to have a well-structured plan in relation to forecasting the demand for electricity in their most diverse sectors. Several techniques are used to predict electrical loads, such as artificial intelligence models, statistical models and hybrid models. This work aims to present a model based on the combination of a statistical method, the Integrated Autoregressive Model of Moving Averages with Seasonality (SARIMA), and an artificial neural network, the Generalized Regression Neural Network (GRNN), with the objective of improving the accuracy of electricity demand forecasts. The data set used in this work belongs to a group of buildings located in the Itaipu Technological Park (PTI) and was acquired through electronic meters installed together with the transformers that serve each of these buildings, performing data collection every 15 minutes. After treating and refining the database, forecasting techniques were applied, each using a forecasting horizon of 1, 3 and 5 days, the first technique being the combination of GRNN and SARIMA, and the other techniques used were the predictive methods themselves separately, the artificial neural network called Short and Long Term Memory Network (LSTM), widely used for forecasting time series, and a combination between LSTM and the SARIMA statistical method allowing the comparison of their results. The results obtained with the proposed combined model GRNN+SARIMA are, in general, more accurate when compared with the results of the techniques individually, as they combine the advantages of each technique and end up smoothing the negative characteristics of each other, thus causing a balance that is reflected in the generated forecast. The results are similar to the LSTM results in some simulations.