A Hierarchical Ensemble Learning Approach for Short-Term Wind Power Prediction Based on CEEMDAN-VMD Decomposition

Wind power prediction plays a crucial role in power system dispatch and the secure and stable operation of power grids. To address the challenge of traditional single models in effectively handling the multi-scale characteristics of wind power sequences, this paper proposes a short-term wind power prediction method based on CEEMDAN-VMD decomposition and hierarchical ensemble learning. The method first applies CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) to initially decompose the original wind power sequence, obtaining multiple Intrinsic Mode Functions (IMFs). It then further decomposes the IMF components using VMD (Variational Mode Decomposition), categorizing the results into low-frequency, medium-frequency, and high-frequency components based on their frequency characteristics. For prediction, the method employs specialized models tailored to each component's features: LSTM (Long Short-Term Memory) for low-frequency components, Random Forest for medium-frequency components, and XGBoost for high-frequency components. Finally, a dynamic weighting ensemble strategy based on accuracy and stability is proposed, which adaptively adjusts the weights of each component's prediction results to generate the final power forecast.Experimental results demonstrate that the proposed method outperforms traditional single models and simple averaging ensemble approaches across evaluation metrics including RMSE, MAE, and R2, validating its effectiveness and superiority. Parameter sensitivity analysis and ablation experiments further confirm the necessity of both the dynamic weighting mechanism and dual decomposition strategy, offering a new technical approach for wind power prediction.