Previsão da velocidade do vento na escala do parque eólico utilizando o modelo WRF e rede neural artificial

• Ano de defesa: 2019
• Autor(a) principal: Santos, Iwldson Guilherme da Silva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Alagoas; Brasil; Programa de Pós-Graduação em Meteorologia; UFAL
• Palavras-chave em Português: Ventos; Energia eólica; Weather Reaserch and Forecasting (Modelo atmosférico); Redes neurais (Computação); Winds; Wind energy; Weather Reaserch and Forecasting (Atmospheric Model); Neural Networks (Computation); CNPQ::CIENCIAS EXATAS E DA TERRA::GEOCIENCIAS::METEOROLOGIA
• Link de acesso: http://www.repositorio.ufal.br/handle/riufal/5769

The objective of this work is to improve the prediction of the wind speed with 24 hours in advance, using forecast of the atmospheric model WRF (Weather Research and Forecasting) every 10 minutes. In addition, these outputs are refined through an Artificial Neural Network (ANN) nonlinear autoregressive with external inputs (NARX) to improve the WRF prognosis. This is a study with many tests and computational simulations. The OBS data were measured in an anemometric tower with measurements of wind speed and direction in 50, 70 and 100 meters. The tower was located at Craíbas (Alagoas State, Brazil). The configuration of the WRF model to generate the simulated data was based on recent studies for tropical regions. The ANN training was done with the WRF and OBS initial series. Then, the algorithm makes the forecast of the wind velocity from the training output. The results showed that the WRF prognoses made at 10-minute intervals were much better than those obtained in several 60-minutes studies. The use of ANN-NARX using these prognoses to forecast the next day’s wind speed proved to be a viable option. The mean velocity OBS was 5,30 m/s while the mean WRF e ANN were 5,20 m/s and 5,32 m/s respectively. In the comparison between ANN and WRF the following were observed: Mean deviation almost null (-0,01 m/s versus -0,35 m/s); Lower REQM (1,18 m/s versus 1,24 m/s); higher correlation coefficient (0,70 versus 0,61). It was evident that the chosen period for training (3 days) implies large errors when there is lot variability in the OBS and/or WRF series.