COMPARATIVE ANALYSIS OF RECURRENT AND TRANSFORMER MACHINE LEARNING MODELS FOR WIND TURBINE ACTIVE POWER GENERATION FORECASTING

Wind turbines are inherently intermittent energy sources that predominantly depend on meteorological conditions. Accurate prediction of wind turbine active power generation is crucial for optimizing the operation of modern power systems rich in renewable energy sources, improving system stability, enhancing the efficiency of production and consumption balancing, optimizing the electricity market, and numerous other aspects. The development and application of advanced artificial intelligence and machine learning (ML) methods can further increase prediction accuracy, enabling the system to better adapt to the variable nature of wind energy. This paper investigates the efficiency of two machine learning approaches for predicting wind turbine active power, focusing on ML models based on recurrent neural networks (LSTM and GRU) and Transformer models adapted for processing and predicting time series data. Transformer models have recently gained prominence in time series analysis due to their ability to effectively recognize long-term dependencies in data through the self-attention mechanism, enabling parallel sequence processing and more accurate identification of relevant patterns in the data. The choice of transformer model architecture and parameter tuning is conditioned by two iterative processes. The validation of the developed ML model is performed using a practical open-source dataset that contains a large number of input features related to meteorological and operational system parameters. The selection of the most relevant features is based on their correlation with the target variable (in this case, active power), reducing the dimensionality of the problem and improving model efficiency. The experimental evaluation includes an analysis of model performance using standard metrics (RMSE and MAEI), while simultaneously examining the transformer architecture and its parameter set. The obtained results provide insight into the advantages and limitations of both approaches in various scenarios of short-term and long-term prediction, with the aim of improving broader prediction strategies and optimizing wind farm operations.