Abstract
Wind energy has become a key answer to the world's energy problems, providing a clean and sustainable option instead of relying on fossil fuels. Enhancing wind energy systems and energy management is essential through efficient wind speed prediction. However, the complex nature of wind speed data contains significant challenges with existing forecasting models for long-term nonlinear forecasting accuracy, and this causes a lack of wind energy predictions, which may cause false distributions of energy. This study proposes a multi-step methodology that integrates Variational Mode Decomposition (VMD) with advanced machine learning like Extreme Gradient Boosting (XGBoost), Adaptive Boosting (AdaBoost), Light Gradient Boosting Machine (LightGBM), K-Nearest Neighbor (KNN), and transformer-based model (Informer) to improve long-term wind speed forecasting. The approach involves data collection from the NASA Power project, which consists of 35k samples of wind speed data, with performance evaluated on R-squared (R²) score and error metrics. The proposed approach demonstrated state-of-the-art performance, with LightGBM achieving the highest R² of 98% and the lowest error metrics. XGBoost and KNN performed slightly lower in R², achieving 97% score. Despite the high performance of the Informer model, it demonstrated the lowest in scores with a 78% R² score. The study's novelty lies in highlighting the effectiveness and efficiency of VMD in addressing the complexities of wind speed data and underscores the potential of combining decomposition techniques with advanced machine learning models for accurate wind speed forecasting.