Previsão de velocidade do vento para produção de energia eólica utilizando redes neurais transformers

Abstract

The growing demand for renewable energy requires accurate wind speed forecasts to optimize wind power production and ensure grid stability. This study investigates the use of Transformer neural networks for long-term wind speed forecasting, exploring their advantages over recurrent models such as Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), which face challenges in capturing long temporal dependencies and making long-term predictions. The main motivation for this study lies in the variability and unpredictability of wind patterns, which hinder the efficient operation of wind farms. In this context, the objective of this research is to enable the use of Transformer neural networks for long-term wind power forecasting. The dataset used in this research consists of four time series representing wind speed in m/s, collected hourly throughout 2018 (a total of 8,760 samples) in the city of Limoeiro do Norte – CE. This dataset was derived from MERRA-2 (Modern-Era Retrospective Analysis for Research and Applications, Version 2), an atmospheric reanalysis system developed by NASA. Additionally, temporal data (such as day, month, and year) were incorporated to increase the dimensionality of the dataset. To mitigate data scarcity, augmentation methods such as time series decomposition into modes (VMD and ICEEMDAN) and Generative Adversarial Networks (TimeGAN) were employed. The Transformer model, originally designed for Natural Language Processing (NLP), was adapted for time series forecasting, utilizing variants such as the Temporal Fusion Transformer (TFT), which incorporates sparse attention mechanisms to improve computational efficiency and reduction on prediction error. The long-term forecasting horizon was defined as 168 hours ahead or a full week, and model evaluation was conducted using Root mean square error (RMSE), Mean Absolute Error (MAE) e Mean Absolute Percentage Error (MAPE). The results highlight significant improvements: the pure TFT model, with RMSE of 1.478 m/s and MAPE of 19.24%, was substantially outperformed by hybrid approaches. The best configuration, combining TimeGAN and VMD with the TFT, achieved an RMSE of 0.625 m/s, MAE of 0.525 m/s, MAPE of 8.26%, and DA of 88.02%, representing reductions of over 55% in error metrics compared to the baseline. These findings suggest that Transformer-based models are highly promising for meteorological time series forecasting when enhanced with decomposition and data augmentation techniques, although future work is still required on computational optimization and broader validation with multi-year datasets.