A short-term wind power forecasting method based on sequence decomposition and high-dimensional feature aggregation

To ensure the stable operation of the power grid and mitigate the impact of wind power integration on the grid, more accurate wind power forecasting techniques are needed. Establishing a precise wind power forecasting model remains a significant challenge in this field. In response to this, this paper constructs a composite forecasting model based on a "decomposition-aggregation-prediction" framework, focusing on key steps in the prediction process. First, the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is used to decompose the wind power sequence, with the resulting modal components forming the dataset. Then, a prediction model (CACNNLightGBM) is constructed using Channel Attention, CNN, and Light Gradient Boosting Machine (LightGBM). Stacking Cross Validation (SCV) is used for model training. Simulation validation using actual power data shows that the proposed model has higher accuracy across various prediction scenarios and demonstrates better generalization capabilities compared to conventional forecasting models.