Objective Developing a high-precision short-term wind power forecasting method can ensure the safe and stable operation of the power grid and provide technical support for enhancing wind power utilization efficiency.

Method Based on the temporal correlation characteristics of wind speed and output series, a combined forecasting framework integrating transfer entropy-based feature selection, ensemble empirical mode decomposition, attention mechanisms, and recurrent neural networks were proposed, and the corresponding deep learning models were established. First, the transfer entropy values between meteorological factors and wind power were compared to determine the direction of information flow and identify the key variable set. Second, ensemble empirical mode decomposition was employed to decompose the wind power series into intrinsic mode functions and a trend term. Third, attention mechanism-based recurrent neural network models were constructed separately, and the forecasting results of the three models were aggregated and reconstructed. Finally, an attention mechanism was introduced to dynamically assign weights to the three models, thereby generating the final combined forecasting output.

Result Comparative case studies show that, compared with the best single sub-model, the combined model achieves maximum reductions of 18.802% in MAE and 17.652% in RMSE, while the maximum increase in R² reaches 3.268%, indicating a significant improvement in forecasting accuracy.

Conclusion The proposed combined forecasting method can effectively enhance the accuracy of short-term wind power prediction and holds significant engineering application value. Future research can further conduct cross-scenario verification under various wind farms and diverse meteorological conditions to assess the generalization capability and stability of the model.