As a part of new energy, wind power generation requires precise wind speed
forecasting to enhance grid reliability. This paper proposes a hybrid
time-series pattern prediction framework designed to continuously forecast wind
speed across multiple wind turbine units. The proposed framework incorporates
three key components.
First of all, a multi-scale temporal pattern extraction module is introduced to
improve the capability of the model to capture time-dependent structures,
thereby enhancing predictive accuracy and robustness. Second, a
three-dimensional adaptive probabilistic attention mechanism is developed to
reinforce temporal feature interaction and fusion, ensuring both efficiency and
performance. Third, a feature-factor pattern fusion strategy is applied to
effectively model complex wind speed variations under diverse influencing
factors, while reducing computational burden during training. Through
comparative experiments, our model has better performance, obtaining an RMSE of
36.3, MAE of 5.34, and MAPE of 1.02%, which confirms its capability in
delivering accurate and stable wind speed predictions.