Ultra-Short-Term Offshore Wind Power Prediction Based on Secondary Decomposition, Multi-Objective Optimization, and Deep Learning

This study addresses the challenges of strong nonstationarity, high noise interference, and inadequate multi-step prediction accuracy in ultra-short-term offshore wind power forecasting, proposing a method integrating multi-step signal decomposition, multi-objective optimization, and deep learning. First, ICEEMDAN decomposition is applied to the original sequence to suppress mode mixing, and sample entropy combined with K-means clustering is used to reduce the dimensionality of Intrinsic Mode Functions (IMFs). Next, Variational Mode Decomposition (VMD) extracts multi-scale features from the highfrequency components of the reconstructed signal. Subsequently, a multi-step deep learning architecture is constructed: a Convolutional Neural Network (CNN) captures local spatial features, a Bidirectional Long Short-Term Memory network (BiLSTM) models forward and backward temporal dependencies, and an attention mechanism dynamically assigns weights to key features, forming the corresponding prediction model. Finally, the Crested Porcupine Optimizer (CPO) is introduced for multiobjective optimization, which automatically optimizes hyperparameters by minimizing prediction error and maximizing stability, enabling 4-6-hour ahead predictions. Data from a Fujian wind farm demonstrate that the model achieves a coefficient of determination of $98.02 \%$, a root mean square error of $13.6 \%$, and a mean absolute error of $16.1 \%$ on the training set, with good generalization performance on the test set, effectively enhancing prediction accuracy and stability.