A practical few-shot wind power prediction method for extreme cold waves: Integrating improved ITimeGAN, de-stationary attention, and SHAP interpretability

Cold waves, characterized by abrupt temperature drops and severe icing, trigger drastic, non-stationary power fluctuations in wind turbines—violating the stationarity assumption of conventional short-term wind power forecasting models and causing large prediction errors. Compounding this, the extreme rarity of historical cold-wave samples leads to insufficient training data for data-driven models. To address these challenges, this study first identifies cold-wave events using rigorous thresholds from the China Meteorological Administration to ensure meteorological validity. Then, a hybrid framework tailored to cold-wave scenarios is proposed: an Improved Time-series Generative Adversarial Network (ITimeGAN) is developed to augment scarce cold-wave data, where integrating self-attention and Temporal Convolutional Networks (TCN) allows ITimeGAN to capture the fine-grained temporal dynamics of extreme weather sequences, while Wasserstein optimization stabilizes adversarial training to avoid gradient collapse, generating synthetic samples highly consistent with real cold-wave physical characteristics. Meanwhile, a De-Stationary Attention (DSA) mechanism is introduced to adapt to the strongly non-stationary power patterns during cold waves, and a Bidirectional Gated Recurrent Unit (BiGRU) predictor leverages bidirectional gating to exploit temporal dependencies in both forward and backward directions. Finally, SHapley Additive exPlanations (SHAP) analysis decodes distinct feature interaction patterns between normal and cold-wave scenarios. Experimental results on real-world wind farm data demonstrate the proposed method's superior accuracy, with one-step prediction R2 reaching 0.9663, and enhanced interpretability that guides reliable wind farm management during extreme cold-wave events.