Short-Term Wind Power Forecasting Using CEEMDAN-CNN-BiLSTM Based on MIC Feature Selection

To address the issue of insufficient accuracy in wind power forecasting arising from intermittency and volatility, this paper proposes a short-term wind power prediction model integrating MIC (Maximal Information Coefficient) feature selection with adaptive noise-complete set empirical mode decomposition, convolutional neural networks, and a bidirectional long short-term memory network hybrid architecture. The main innovations of this work lie in the following: Firstly, MIC quantifies the strength of the nonlinear correlation between meteorological features and the MAE (Mean Absolute Error) in power generation, thereby enabling the identification of highly correlated features to reduce the input dimensionality. Secondly, CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) performs adaptive modal decomposition on raw power sequences. Combining sample entropy with K-means clustering reconstructs IMFs (Intrinsic Mode Functions), while the introduction of VMD (Variational Mode Decomposition) for quadratic optimisation significantly improves the quality of signal decomposition, enabling a more refined separation of fluctuation characteristics across different time scales. Finally, the optimised meteorological features and reconstructed components are input into a CNN (Convolutional Neural Network)-BiLSTM (Bidirectional Long Short-Term Memory) module. Power regression prediction is achieved through the synergistic effect of spatial feature extraction and bidirectional temporal dependency modelling. Case study results demonstrate that compared to the TCN (Temporal Convolutional Network)-Transformer, the proposed method achieves a 0.4022 improvement in the coefficient of determination R2, a 13.2598 reduction in MAE, a 19.864 decrease in RMSE (Root Mean Square Error). At the same time, it maintains stable performance even when faced with unreliable data scenarios involving random missing features, demonstrating excellent generalisation ability. Furthermore, the model training time has been reduced to 77.6469 s, with a single prediction response time of just 0.0659 s.