Due to the growing importance of renewable energy in the energy industry, accurate wind power forecasting has become increasingly crucial. The deep learning forecasting method based on MLP (multi-layer perceptron) is prone to overfitting and struggles to capture long-term dependencies. Traditional solutions mainly involve increasing the dataset size, adjusting model parameters, or using more complex neural networks, but their effectiveness has been limited. To address these issues, the feature extraction capability of MLP in the frequency domain was studied, and a strategy combining frequency domain feature extraction with an adaptive frequency selection algorithm was proposed. This strategy dynamically adjusts the weights of the frequency components during model training to select the features that contribute most to the prediction. It effectively mitigates overfitting, enhances the model's ability to capture long-term dependencies, and significantly improves both the prediction accuracy and efficiency. Finally, two sets of case studies were conducted, and the results demonstrate that the proposed method outperforms traditional deep learning forecasting models in terms of prediction precision.