A Wind Power Prediction Method Based on VMD-MI-HC-CEEMDAN Decomposition and LSTM-SE-MSConvFormer

Abstract

High-accuracy wind power forecasting technology is crucial for enhancing grid operational stability given the high and increasing proportion of renewable energy. Although the existing methods can enhance data feature modeling ability to some extent, they generally fail to handle nonstationary components and capture deep temporal dependencies at the same time, which severely limits further improvement in prediction performance. This study proposes a novel prediction method that integrates a two-stage decomposition strategy with hybrid deep learning to address these issues. First, the original wind power sequence is decomposed using the variational mode decomposition (VMD). The intrinsic mode function components are then reconstructed by the mutual information (MI)-weighted hierarchical clustering (HC). Thereafter, the reconstructed signal is further decomposed by the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to enhance the stationarity and predictability of temporal features. Finally, a parallel forecasting architecture that integrates the long short-term memory (LSTM) network and the multiscale convolutional attention transformer (SE-MSConvFormer) model is constructed, which fully exploits local dynamics and global dependency relations in the wind power time series. The experiments compared with several conventional network models demonstrated that the proposed model achieved the best performance, obtaining the mean absolute error (MAE) 0.0083, root-mean-squared error (RMSE) 0.0127, mean absolute percentage error (MAPE) 60.27%, and R ${}^{{2}}~0.9988$ . Compared with the convolutional neural network (CNN)-BiLSTM model, the RMSE of the proposed method was reduced by 17.5%.