Abstract: Unplanned icing-induced outages of wind turbines during low temperature and cold wave weather lead to severe fluctuations in wind farm power generation, posing significant challenges to the prediction accuracy of traditional models under such extreme operating conditions. To address this issue, this paper proposes a short-term power forecasting method based on heterogeneous ensemble learning. Firstly, it constructs a wavelet transform-driven deep feature extraction network, in which wavelet transform decouples meteorological data into detail components and trend components. These components are processed separately, where Convolutional Neural Networks enhance the capture of spatial local features, while Long Short-Term Memory networks model temporal dependencies, followed by adaptive feature fusion via a cross-attention mechanism. Subsequently, a heterogeneous ensemble strategy builds a Stacking framework combined with Light Gradient Boosting Machine, Extreme Gradient Boosting, and Support Vector Regression serve as diverse Base Learners to fully exploit feature heterogeneity, while linear regression (LR) acts as the meta-learner to optimize prediction accuracy and robustness. Based on the real-world data from a wind farm in Jiangxi province during winter, the proposed method achieves a 4-hour-ahead prediction with mean absolute error, mean squared error, and coefficient of determination of 0.028, 0.119, and 0.618, respectively. The experimental results demonstrate that the proposed model significantly enhances forecasting accuracy under low temperature and cold wave weather conditions compared with both single models and conventional ensemble approaches.

Key words: short-term wind power forecasting, low temperature and cold wave weather, ice coating, heterogeneous ensemble learning, stacking ensemble