A Novel Interpretable Deep Learning‐Based Wind Speed and Power Generation Forecasting Using Multiscale Attention and Post Hoc Feature Importance Mechanism

ABSTRACT Accurate and efficient wind speed forecasting can enhance the scheduling of wind farms and ensure the stable operation of power grids. However, the inherent stochastic variability and complex fluctuation patterns of wind speed sequences increase the difficulty of forecasting, and existing deep learning‐based forecasting methods struggle to provide interpretable results. This study proposes an interpretable wind speed forecasting method based on deep learning. This method integrates two‐stage decomposition, time series embedding, a dual‐channel hybrid neural network, advanced attention mechanisms, and meta‐heuristic algorithms to achieve precise and efficient wind speed predictions. In addition, this study introduces a model‐agnostic post hoc feature importance ranking method for interpretability, which enhances the interpretability of the forecasting model by processing test data to output feature importance rankings. After wind speed predictions are completed, this research incorporates real wind turbine data to perform wind power conversion for enhancing its practical value. The designed ablation experiments and multiple comparative experiments in this study validate the comprehensiveness and advancement of the model. The interpretability results and wind power conversion outcomes also provide additional analytical perspectives for related decision‐making processes.