Multi-stage wind speed prediction with CEEMDAN-SE-IDBO-LSTM based on rolling decomposition

Abstract

To address the problem of inaccurate wind speed prediction for wind turbines, this paper proposes a multi-stage coupled “Decomposition-Reconstruction-Optimization-Prediction” approach named CEEMDAN-SE-IDBO-LSTM (CSIL). The method first combines Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and rolling decomposition strategy to prevent information leakage in traditional decomposition. The wind speed components are standardized into high, medium, and low-frequency categories through Sample Entropy-based reconstruction. Subsequently, an Improved Dung Beetle Optimizer (IDBO) combining Halton sequence initialization, mirror reflection obstacle avoidance, and hybrid Gaussian-Cauchy mutation perturbation strategies is applied to optimize LSTM hyperparameters. The optimized LSTM models then process different frequency components for enhanced prediction. Experimental results indicate that the proposed multi-stage method achieves MAE, RMSE, and MAPE values of 0.4342 m/s, 0.5647 m/s, and 3.8726 % respectively, with R²of 0.8003. This represents significant performance improvements over baseline models and outperforms existing mainstream wind speed prediction methods in both accuracy and stability.