Objective To address the challenges of excessive feature factors, difficulties in effectively mining the mapping relationship between key features and power, and low prediction accuracy in wind power forecasting, this paper proposes a short-term wind power prediction method that incorporates two-stage feature selection.

Method First, the maximum relevance minimum redundancy (mRMR) criterion based on mutual information was used to analyze the correlation among features such as wind speed and direction in wind power prediction data. Initial feature screening was achieved by maximizing the correlation between features and wind power while minimizing inter-feature redundancy. Next, the out-of-bag (OOB) feature importance evaluation method within the random forest (RF) algorithm was employed to assess the significance of the pre-screened features, and a secondary screening was performed to obtain the optimal feature subset. Then, the adaptive parameters of the Newton-Raphson based optimizer (NRBO) were improved, and the enhanced NRBO algorithm was used to optimize the parameters of a bidirectional long short-term memory network (BiLSTM). Finally, the optimal feature subset was input into the NRBO-BiLSTM model to predict short-term wind power.

Result The case study analysis shows that the proposed two-stage feature selection method can effectively eliminate the redundancy among features and low-importance features, and the optimal feature subset significantly improves the model performance; The improved NRBO algorithm has a remarkable optimization effect, and its prediction accuracy is significantly better than that of the comparison models.

Conclusion This study successfully combines two-stage feature selection with an improved NRBO optimization algorithm. The proposed model can accurately capture the fluctuation patterns of wind power and effectively improve the prediction accuracy of short-term wind power.