Redes neurais artificiais BIGRU_CNN aplicadas à previsão de demanda de energia elétrica de curto prazo

• Ano de defesa: 2021
• Autor(a) principal: Soares, Lucas Duarte
• Orientador(a): Franco, Edgar Manuel Carreno
• Banca de defesa: Machado, Renato Bobsin , Müller, Marcos Ricardo
• 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: Inteligência artificial; Aprendizado de máquinha; Redes neurais recorrentes; Séries temporais
• Palavras-chave em Inglês: Artificial intelligence,; Machine learning,; Recurrent neural network,; Time series
• Área do conhecimento CNPq: ENGENHARIA ELETRICA::TELECOMUNICACOES
• Link de acesso: http://tede.unioeste.br/handle/tede/5712

The present work analyzed the comparison between feedforwards, recurrent, convolutional and bidirectional artificial neural networks based on different layers architectures as a predictive tool for short-term load forecasting. These forecasting models can serve as a support instrument related to the decision making of companies in the energy sector, as the demand for energy is requested one day before its transmission in much of the world. The code of the artificial neural networks was programmed in Python using the Keras package. Forecasts for all networks have been performed 10 times until an acceptable statistical sample is reached so that future values demand for energy are as close as possible to reality. The best forecasting model was the BiGRU_CNN network where the average errors attributed to its predictions in a 24-hour horizon was 3.42% for the MAPE error, 100.75 MW for the MAE accuracy metric and 122.2 MW for the RMSE error.