Multi-scale Decomposition and Hybrid Deep Learning CEEMDAN-VMD-CNN-BiLSTM Approach for Wind Power Forecasting

To address the challenges posed by the volatility and uncertainty of wind power generation, this study presents a hybrid model combining Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Variational Mode Decomposition (VMD), Convolutional Neural Network (CNN), and Bidirectional Long Short-Term Memory (BiLSTM) for wind power forecasting. The model first employs CEEMDAN to decompose the original wind power sequence into multiple scales, obtaining several Intrinsic Mode Functions (IMFs). These IMFs are then classified using sample entropy and k-means clustering, with high-frequency IMFs further decomposed using VMD. Next, the decomposed signals are processed by a CNN to extract local spatiotemporal features, followed by a BiLSTM network that captures bidirectional temporal dependencies. Experimental results demonstrate the superiority of the proposed model over ARIMA, LSTM, CEEMDAN-LSTM, and VMD-CNN-LSTM models. The proposed model achieves a mean squared error (MSE) of 67.145, root mean squared error (RMSE) of 8.192, mean absolute error (MAE) of 6.020, and a coefficient of determination (R2) of 0.9840, indicating significant improvements in forecasting accuracy and reliability. This study offers a new solution for enhancing wind power forecasting precision, which is crucial for efficient grid operation and energy management. Future work will focus on optimizing the model structure and parameters to reduce computational costs and improve robustness, further advancing the application of hybrid models in wind power forecasting.