Interpretability has become a critical requirement in modern deep learning applications for renewable energy forecasting, especially in complex and safety-critical contexts such as offshore wind power systems. To simultaneously improve predictive accuracy and model transparency, this study proposes an explainable hybrid deep learning framework – VFE-IVYA-CNN-BiGRU – for offshore wind speed forecasting. The model begins with feature selection via Pearson Correlation (PCs) to identify the most relevant meteorological variables from a full set of candidates, thereby enhancing input quality and reducing redundancy. The selected features are then passed through a robust preprocessing module (VFE), which integrates Variational Mode Decomposition (VMD) and Fuzzy Entropy (FE). VMD decomposes the original wind speed sequence into intrinsic mode functions (IMFs), capturing multi-scale temporal structures, while FE quantifies the complexity of each IMF to filter out noise-dominated components. The reconstructed sub-sequences are first processed through a Convolutional Neural Network (CNN) to capture temporal local dependencies, and then passed into a Bidirectional Gated Recurrent Unit (BiGRU), which effectively learns time dependencies in both forward and backward directions. To further enhance model performance, the Ivy Algorithm (IVYA) is employed to optimize hyperparameters adaptively, improving convergence and generalization. To improve interpretability, SHapley Additive exPlanations (SHAP) are utilized to quantify the contribution of each meteorological feature to the model's output, revealing both dominant drivers (e.g., gust speed) and interaction patterns across seasons. The proposed framework is evaluated using seasonal offshore wind datasets (spring, summer, autumn, and winter) sourced from a wind power site along the Guangdong coastline, China, and contrasted with six leading benchmark models. Empirical findings reveal that the proposed VFE-IVYA-CNN-BiGRU consistently outperforms existing methods in terms of accuracy, robustness, and interpretability. The integration of SHAP-based explanations ensures model transparency, making the approach a reliable tool for intelligent control and decision support in offshore wind farm operations.
Liu et al. (Sat,) studied this question.