Enhancing wind speed forecasting accuracy: comparative insights into recurrent neural networks for short-term prediction

Abstract

Short-term wind speed forecasting is essential for maintaining grid stability and supporting the integration of renewable energy, yet the strong variability of wind makes accurate prediction difficult. Sudden fluctuations and nonlinear atmospheric behavior often reduce the performance of conventional artificial intelligent models. To address this challenge, this study evaluates three forecasting methods which include a gated recurrent unit (GRU) model, a temporal convolutional networks (TCNs) model, and a hybrid GRU–TCN design that enables prolonged term forecasting while enabling quick identification of localized weather changes across various meteorological parameters. The researchers used Laayoune, Morocco data to build their model training process. The hybrid method exceeded all other models because it achieved an R² value of 0.99 and a root mean square error (RMSE) of 0.16 m/s and a mean absolute error (MAE) of 0.03 m/s. The system successfully manages sudden shifts in wind patterns while maintaining accurate site-specific physical behavior. The hybrid GRU–TCN design functions as a dependable and expandable system, which delivers real-time wind forecasting capabilities that enable effective smart grid operations and facilitate the growth of wind energy systems.