Combined Prediction of Wind Power in Extreme Weather Based on Time Series Adversarial Generation Networks

The current global climate is complex with an increasing frequency of extreme weather events. The randomness, variability, and intermittency of new energy sources pose significant challenges for the balance between power generation and consumption in electricity grids. This challenge is particularly pronounced during extreme weather events such as cold waves, which aggravate the challenges of power supply stability. Accurate prediction of the wind power output can reduce the need for system backup capacity, ensuring the stable operation and reliability of the power system. However, current prediction models do not effectively consider the impact of extreme weather, leading to low prediction accuracy and large deviations. Additionally, the mechanisms by which extreme weather affects wind power differ from those under normal weather conditions. Extreme weather events such as cold waves are rare, and sample data are scarce, making it difficult to establish precise prediction models. To address these issues, this study proposes a wind power output prediction correction method based on time-series generative adversarial networks. First, a time-series generative adversarial network algorithm was used to generate samples from meteorological and power data, constructing an extreme weather meteorological and power database to effectively address the issue of sample scarcity. Second, the improved particle swarm algorithm (IPSO) and Bayesian optimization (BO) algorithm were used to optimize the parameters of single-prediction algorithms such as eXtreme gradient boosting (XGBoost) and least absolute shrinkage and selection operator (LASSO). Subsequently, by combining the algorithm performance and principle of maximum diversity, the optimal algorithm combination was determined using the evaluation indicators of the prediction error and Pearson correlation coefficients to construct an extreme weather power prediction model under the stacking ensemble learning framework, overcoming the limitations of single algorithms and effectively improving the prediction accuracy. Finally, Support Vector Regression (SVR) was used to construct a power prediction correction model for similar days, correcting the predictions. Through verification using an actual cold wave case study, the proposed method demonstrated significantly improved prediction accuracy and performance compared to single algorithms and conventional prediction models.