A multi-dimensional interpretable wind speed forecasting model with two-stage feature exploring

Accurate wind speed forecasting is critical for optimizing wind energy utilization, yet its inherent stochasticity, nonlinearity, and regional variability pose significant challenges. Existing models also lack multidimensional interpretability, limiting their practical utility. To address these issues, we introduce a new short-term wind speed prediction model with an hourly time span that combines state-of-the-art techniques: a two-stage feature selection for meteorological data, snow ablation optimizer (SAO), and temporal fusion transformer (TFT). Meteorological features are processed arithmetically and nonlinearly and combined with statistical feature extraction to form a comprehensive feature set that captures the unique fluctuating features in the wind speed series. A two-stage feature selection strategy ensures valid feature information and controls input quality. Finally, the TFT results and tree SHAP construct a multidimensional interpretable analytical framework for the input and forecasting process. The proposed method shows good prediction performance on all four wind speed datasets of the Williams Wind Farm, with MAPEs of 6.7 %, 6.29 %, 16.22 %, and 12.72 % in spring, summer, fall, and winter, respectively. It also provides decision makers with a clear analysis from feature engineering to predictive modeling, which contributes to orderly energy planning and strategic placement.