Previsão de velocidade do vento com dados reais de um parque eólico localizado no Ceará utilizando redes neurais artificiais: um estudo comparativo das redes MLP, LSTM, GRU e CNN

Abstract

The present work conducts a comparative study of the performance of different Artificial Neural Networks for predicting future wind speed based on the historical wind speed series read from an anemometric measurement tower located in the city of Quixaba-CE. It highlights the importance of this type of forecasting to improve the operational planning of wind farms, promote the integration of this energy source into the National Interconnected System (SIN), and enhance the stability of the power grid. The work discusses the various types of wind speed and wind energy forecasting methods that are utilized, as well as the related studies on the neural network architectures that were defined for constructing the predictive models (MLP, LSTM, GRU, and CNN), demonstrating the significance and effectiveness of these models for the evaluated time horizon. The work also provides a brief introduction to Artificial Neural Networks, describing the structure of each explored architecture and explaining the learning process based on the backpropagation algorithm, which adjusts the network weights. The wind speed data were processed and used to optimize the hyperparameters through the cross-validation technique applied to time series. With the hyperparameters defined, the models were built and trained with wind speed data at 10-minute intervals from 2021 to 2022 and subsequently tested with data from 2023. The models were evaluated using the metrics MAE, RMSE, MAPE, R², and through variance analysis (ANOVA), and were considered effective in predicting wind speed for the locality of Quixaba-CE. Among the implemented models, the model based on the GRU network obtained the best results, with a MAE of 0.47 m/s, a RMSE of 0.64 m/s, a MAPE of 7.55% (indicating that, on average, the model's prediction diverges 7.47% from the real value) and an R² of 0.90. With the results analyzed, it was possible to conclude that, considering the cost of the computational resources involved, the models based on the MLP and CNN networks could be used in production for short period forecast horizons. However, it highlights that if the aim is to seek maximum performance, models based on LSTM and GRU networks are more recommended.