Comparison of LSTM, GRU and Transformer Neural Network Architecture for Prediction of Wind Turbine Variables

Abstract

To deal with climate change and global warming, several countries have taken steps to reduce greenhouse gas emissions and have begun to switch to renewable energy sources. Wind energy is one of the most profitable and accessible technologies in this sense, and among the alternatives to manage its variability is prediction, which has become an increasingly popular topic in academia. Currently, there are several methods used for prediction, the methods based on artificial neural networks (RNA) are the ones that have aroused the greatest research interest. In this sense, this study presents the training of different ML algorithms, using deep learning models for the prediction of the power generated by a wind turbine. In addition, a new filtered signal is generated from the wind signal that is integrated into the set of input signals, obtaining better training performance. These models include long-short-term memory (LSTM), recurring rental unit (GRU) cells, and a Transformers-based model. The results referring to the trained models show that the technique that best fits time series I and time series II is GRU. This study also allows us to analyze different AI techniques to improve performance in wind farms.