A method for short-term wind power forecasting under extreme weather conditions based on meteorological factor interpretability and hybrid deep learning algorithms

Accurate forecasting of renewable energy generation is the foundation for all renewable energy consumption technologies. Currently, wind power forecasting techniques have become relatively mature under normal weather conditions. However, under extreme weather conditions, the difficulty of research is caused by data scarcity and significant weather fluctuations. Here, this paper proposes a wind power forecasting framework that considers small sample expansion and hybrid deep learning algorithms. First, the forecasting framework selects sensitive meteorological factors through the shapley additive explanations (SHAP) theory, reducing the redundancy of model input features. Simultaneously, a conditional generative adversarial network based on discriminant weights is employed to expand small samples of extreme weather data, overcoming the issue of data scarcity and improving model training efficiency. Finally, a hybrid deep learning model, the convolutional neural network (CNN)-bidirectional long short-term memory (BiLSTM) network-attention mechanism (AM), is constructed. This model combines the CNN and BiLSTM network to capture local features and long-term temporal trends complementarily and further enhances the learning of key temporal features through an AM, finally outputting the forecasting results. Through case analysis on a wind power dataset in Liaoning, China, the experimental results show that compared with traditional forecasting methods, the proposed framework reduces the mean absolute error (MAE) by 52.29% and increases R2 by 0.0625 under windy conditions, and reduces the MAE by 64.29% and increases R2 by 0.0462 under low-temperature weather. This indicates that the proposed wind power forecasting framework can effectively improve the forecasting accuracy under extreme weather conditions.