A hybrid BiGRU-Transformer approach for short-term wind power prediction based on secondary decomposition

Accurate and reliable wind power prediction is vital for grid integration and stability. However, the complexity and strong non-stationarity of wind power sequences limit feature extraction and model generalization. To overcome these challenges, this study proposes a short-term hybrid forecasting framework based on feature selection and multi-scale decomposition. The Shapley Additive Explanations (SHAP) method evaluates feature importance and reduces dimensionality, ensuring highly relevant and non-redundant inputs. Considering the multi-scale characteristics , a hybrid decomposition strategy is introduced: Season-Trend Decomposition using Loess (STL) separates seasonal, trend, and residual components, and the fluctuating residual is further decomposed by Variational Mode Decomposition (VMD) into intrinsic modes, thereby alleviating non-stationarity. For prediction, a Bidirectional Gated Recurrent Unit (BiGRU) captures temporal dependencies, while a Transformer extracts global and multi-scale features. To enhance adaptability, an Improved Crested Porcupine Optimizer (ICPO) optimizes key parameters of the BiGRU-Transformer model. Simulation experiments on seasonal datasets verify that the proposed method outperforms benchmark models, reducing MAPE, MAE, and MSE by 7.6%, 10.4%, and 11.8%, respectively, while exhibiting strong stability and generalization. These results confirm the superior forecasting accuracy and robustness of the proposed model.