2025-03-072025-03-072021-10-06FIGUEIREDO, Yann Fabricio Cardoso de. Predição de séries temporais da velocidade do vento no Brasil. Orientador: Lídio Mauro Lima de Campos. 2021. 83 f. Trabalho de Conclusão de Curso (Bacharelado em Ciência da Computação) – Faculdade de Computação, Instituto de Ciências Exatas e Naturais, Universidade Federal do Pará, Belém, 2021. Disponível em:. Acesso em:.https://bdm.ufpa.br/jspui/handle/prefix/7785Historically, Brazil has its electricity production sustained by two main generation sources, hydroelectric and thermoelectric, and the second, in many of the plants, is used more when the first has a low period. In recent years, Brazil has been increasing research and consequently the use of another type of energy generation, wind, which is currently the second largest source of energy generation in the country. The advantage of wind power generation is that it is a type of clean energy, thus helping to preserve the environment. Despite being beneficial and cheaper than the other energy sources mentioned, the generation by wind turbines faces a problem of unpredictability regarding the main resource, the wind. An increasing need then arises to research ways to predict wind availability based on certain variables, such as wind speed. This work aims to develop wind speed prediction models, using artificial intelligence techniques applied together with Deep Artificial Neural Networks (RNAPs) through direct (Deep Feedforward) and recurrent (LSTM) network architectures ), in order to have a more complete notion of the availability of wind energy in the studied regions. The research was based on meteorological variables from the INMET (National Institute of Meteorology) repositories, in the case of the Macau (RN), and SONDA (National Environmental Data Organization) repositories, in the case of the Macau wind farm. Petrolina (PE). Both databases contain information for the period from January 1, 2004 to May 31, 2017 in the database in day format, June 1, 2016 to May 31, 2017 in the database in time format, and May 27, 2017 to May 31, 2017 with the base in minute format. For a single specific prediction model, based on days format, the time series runs from January 1, 2015 to July 31, 2018. In total, 15 models were generated, created from the application of Ensemble Learning Methods, in the case of voting and bagging, in various sub-models created for each database and data term. The best model was using the short-term Petrolina base, with records available in the format of hours, obtained by using the bagging method to create the final model. The best model obtained MAE of 0.0036, MAPE of 0.0012% and RMSE of 0.0143.Acesso AbertoEnergia eólicaPrediçãoRedes neurais artificiaisRede deep feedforwardRede recorrente (LSTM)Velocidade do ventoSéries temporaisWind energyPredictionArtificiais neurais networksDeep feedforward networkRecurring network (LSTM)Wind speedTime seriesCNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAOTrabalho de Curso - Graduação - Monografia