Abstract

Wind energy prediction plays a crucial role in renewable energy planning and policymaking. This study leveraged the Darts time-series framework to compare 10 different machine learning algorithms to investigate the use of solar radiation data for accurate wind energy prediction and compare with previous study results. Evaluation was conducted using two performance metrics: Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) to inform on the prediction's accuracy. The best-performing model for wind speed prediction was the CatBoost model with a 12-hour lag, achieving an RMSE score of 0.877, while for wind direction prediction, CatBoost also achieved the best performance with an RMSE value of 93.8, but with a 2-hour lag, both with the inclusion of covariates. In comparing the performance of the top model trained using Darts in this study with previous research conducted without Darts but utilizing similar models available in Darts, the use of Darts yielded comparable or even superior results with better computational efficiency. Classical models, particularly CatBoost, outperform neural network models in terms of accuracy and computational time, providing valuable insights for investment decision-making. Also, the inclusion of covariates significantly enhances the performance of wind energy prediction models. Covariates such as gust, global horizontal irradiance (GHI), and relative humidity demonstrate a strong influence on wind speed and wind direction predictions. With efficient and accurate prediction of wind speed and wind direction using the Darts Model framework, optimal utilization of wind resources can be achieved, ensuring a sustainable and reliable energy supply.

View Github Repo