Aplicação de redes neurais artificiais para previsão de demanda e preço de energia elétrica no contexto de cidades inteligentes

• Ano de defesa: 2019
• Autor(a) principal: Laboissiere, Leonel Alejandro
• Orientador(a): Fernandes, Ricardo Augusto Souza
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Engenharia Urbana - PPGEU
• Palavras-chave em Português: Redes neurais artificias; Previsão de séries temporais; Mercado de energia elétrica; Cidades Inteligentes
• Palavras-chave em Inglês: Artificial neural networks; Forecasting of time series; Energy market; Smart cities
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO; CIENCIAS EXATAS E DA TERRA; ENGENHARIAS
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/11559

This research proposed a very short-term forecasting framework electricity price and demand based on Artificial Neural Networks (ANN). Effectiveness forecasting tools are essential to facilitate the decision making process of the stakeholders in the deregulated electricity market. Besides, accurate short-term load forecasting (STLF) and electricity price forecasting (EPF) play a significant part for controlling and scheduling of smart grids, consequently, to ensure effectiveness energy resources of smart cities. For case study, a dataset from Australian National Electricity Market was used. The dataset is formed by historical from climate variables, demand and prices series. It should be mentioned that all of these variables were preprocessed using the Weighted Moving Average (WMA) to minimize the effect of noise on the data and help identify trends. Therefore, ANN input set are made by 66 variables/attributes. Correlation-based Feature Selection (CFS) algorithm was applied to form the most relevant variable set to STLF and EPF. As a consequence, reduction of 84 to 90% of the number of variables considered. Moreover, WMA of meteorological variables were selected applying CFS. In sequence, 20 executions of training and validation of Multilayer feedforward ANN were made. The best results have mean absolute percentage error (MAPE) from 2.68% to 4.84%, for STLF, and MAPE from 7.06% to 19.01%, for EPF.