Wind Speed Forecasting for Wind Turbines Using Attention-Based Machine Learning Models

Renewable energy is essential for humanity to decrease fossil fuel usage and reduce GHG emissions. There are several Renewable Energy Sources (RESs) available including solar, wind, biomass, tidal, wave, and geothermal; among them, wind energy is regarded as one of the most dependable. Thus, forecasting wind speed is essential for energy production companies or prosumers to determine potential wind energy production, however, this forecasting process is significantly complex. Several approaches have been found in the literature, including Temporal Convolutional Network (TCN), GRUs, and LSTM. Also, the self-attention mechanism is quite essential to support other models by reducing the dependency on external information and analyzing input and hidden features information provides better result. Therefore, this research aims to forecast wind speed from historical seasonal data by proposing a Temporal Convolutional Attention Network (TCAN) based on self-attention and comparing it with LSTM-Attention (LSTMA) and GRU-Attention (GRUA). More specifically, to evaluate the proposed TCAN model results, LSTMA and GRUA using self-attention are constructed and the efficiency of each model is explained and compared based on MAE, RMSE, and MAPE values. To conduct the analysis, a case study is chosen for the analysis in Bangladesh named Hatia Island which is also a coastal area. NASA POWER dataset is utilized for a period of 12 months. TCAN provides improved outcomes compared to LSTMA and GRUA throughout the summer, fall, late autumn, winter, and spring seasons, whereas GRUA demonstrates somewhat higher performance in the rainy season.