Ultra-short-term wind power prediction method based on optimized signal decomposition and deep learning

• GBDT-based feature selection removes redundancy and boosts forecast accuracy. • SO-enhanced NGO adaptively tunes VMD modes, improving interpretability and accuracy. • VMD–Informer hybrid effectively models temporal dynamics for ultra-short-term wind power. • RMSprop optimization speeds up training and stabilizes prediction performance. Given the limitations of traditional wind power prediction methods in capturing complex relationships within data, this study proposes a novel ultra-short-term wind power prediction framework that integrates feature selection, adaptive signal decomposition, and an efficient deep learning model. First, a Gradient Boosting Decision Tree (GBDT)–based correlation analysis is employed to identify key meteorological features, reduce redundancy, and improve input data quality. Second, an Improved Northern Goshawk Optimization (INGO) algorithm with a Subtractive Optimizer adaptively determines the number of modes in Variational Mode Decomposition (VMD), enhancing both accuracy and interpretability. Finally, the Informer model with ProbSparse attention and RMSprop optimization is combined with improved VMD to efficiently capture volatility and randomness in wind power data. Case studies on two large-scale wind farms in Inner Mongolia and Northwest China show that the proposed framework consistently outperforms benchmark models in both accuracy and convergence speed. These findings indicate that the joint optimization of feature selection, adaptive decomposition, and attention-based forecasting provides a reliable and effective solution for ultra-short-term wind power prediction.