An attention-based multi-input LSTM with sliding window-based two-stage decomposition for wind speed forecasting

Accurate forecasting of short-term wind speed is essential for enhancing the grid integration safety and economic efficiency of wind farms due to the inherent volatility and stochastic nature of wind speed. An improved hybrid model for wind speed forecasting is proposed in this study. The model incorporates three novel parts: (1) a sliding window-based two-stage decomposition (SWTSD) that leverages the strengths of improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and variational mode decomposition (VMD) is designed for effective feature extraction while preventing data leakage; (2) an attention-based multi-input long short-term memory (MI-LSTM) with additional input gates is developed to extract valuable information from each sub-series, capture the relationships between the sub-series, and mitigate noise impact; (3) a simultaneous Bayesian optimization strategy is applied to optimize hyperparameters for both the decomposition technique and the forecasting model. To verify the effectiveness of our proposed model, the wind speed dataset covering four seasons at the Dong Xinzhuang wind farm in Shaanxi, China is collected. Experimental analysis demonstrates that our proposed model is superior to benchmark models in all four comparative experiments, confirming the effectiveness of our designed decomposition, forecasting, and optimization modules. The results suggest that our proposed hybrid model could be a viable option for future short-term wind speed forecasting practices.