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

This is the authors' abstract. We don't add key points for this paper.

Accurate wind power forecasting is essential for ensuring power system stability and improving wind integration efficiency. This paper presents a novel ultra-short-term wind power prediction model. First, feature selection is performed on wind power data using Pearson and Spearman correlation coefficients (PCC/SCC) to eliminate redundant features. Then, an Improved Subtraction-Average-Based Optimizer (ISABO) is proposed to optimize the parameters of Variational Mode Decomposition (VMD), decomposing the raw wind power sequences into more stationary subcomponents. Finally, an enhanced Multi-Scale Graph Network (MSGNet) is introduced by incorporating the Dish-TS general paradigm for mitigating distribution shift and a probabilistic sparse self-attention mechanism, resulting in the IMSGNet prediction framework. Experimental results show that the proposed FS-ISABO-VMD-IMSGNet model achieves superior forecasting accuracy, outperforming the state-of-the-art VMD-based DCInformer by reducing the Mean Absolute Error (MAE) by 28.2%. This study provides a more reliable foundation for maintaining power system stability. • Novel ultra-short-term wind power forecasting method combining feature selection, decomposition, and deep learning. • Improved optimization algorithm enhances variational mode decomposition for better signal processing. • Advanced neural network model integrates sparse attention and distribution shift adaptation. • Achieves 28.2% reduction in mean absolute error compared to existing benchmark models. • Demonstrates strong prediction accuracy and robustness under complex data conditions.