Abstract: This paper proposes a combined forecasting model based on improved CEEMDAN and BiLSTM to enhance wind power forecasting accuracy and robustness. The model uses CEEMDAN decomposition to break down complex signals into multiple IMFs to address challenges from non-stationary signals and high-frequency noise. Innovatively, AFR calculates the average fluctuation range of IMFs for frequency division, and EWT filters high-frequency components, significantly reducing noise. BiLSTM models are then trained on these components with optimal parameters. Validation across different seasons and datasets demonstrates the model's generality and robustness. This approach effectively improves prediction accuracy and stability, addressing traditional methods' limitations in handling non-stationary signals and noise.