Ajay H. Kumar, Arjun Singh, Sanjay Kumar
Accurate wind speed forecasting impacts grid stability, energy pricing, and ability to replace fossil fuels with clean power, however, stochastic nature and nonlinearity of wind make prediction a challenge. To address these challenges, this work explores a range of forecasting strategies like statistical, machine learning (ML), and deep learning (DL) methods, for short-term (ST) and mid-term (MT) horizons. The experimental results across forecast horizons of 3, 6, 12, and 24 steps demonstrate that statistical models perform well at very short horizons (RMSE ≈ 0.55–1.14, R 2 up to 0.40) but degrade rapidly as the forecast length increases, with RMSE above 1.30 and R 2 near 0.00 at a forecasting horizon of 24. ML model XGBoost offers high accuracy and stability in comparison to RF, having an RMSE value of 0.43 and R 2 value 0.78 at step-3, remaining outstanding even compared at step-24, which offered RMSE = 1.00 to 1.18 and R 2 = 0.25 to 0.40, respectively. Transformer consistently performs better in having RMSE = 0.55–0.70 and R 2 up to 0.57 at step-3, and maintains better accuracy for longer horizons with RMSE = 0.95 and R 2 up to 0.50 at step-24. To improve the wind speed forecasting (WSF), Variational Mode Decomposition (VMD) with LSTM has been used by decomposing wind speed signals (WSS) into intrinsic mode functions (IMF). The proposed decomposition-driven hybrid model and advanced deep learning techniques help to achieve accurate and better forecasting and can be implemented in various applications of forecasting. • Traditional statistical models work well for very short forecasts but quickly lose accuracy over longer periods. • Machine learning, especially XGBoost, offers more stable and reliable predictions. • Deep learning models like Transformers deliver the best performance across short- and long-term forecasts. • Using VMD with LSTM further boosts accuracy by filtering noise and capturing hidden wind patterns.