Wind Power Forecasting Method Under Extreme Weather Conditions Based on Meta-Learning and Multi-Head Attention Mechanism

Wind power drops sharply under extreme weather conditions, which will lead to severe imbalances between power supply and demand, posing a significant threat to the safe and stable operation of new power system. Accurate forecasting of wind power under extreme weather conditions is crucial. However, extreme weather conditions are small probability events and only have limited representative samples, resulting in poor forecasting accuracy of traditional statistical models that rely on big data. Therefore, a wind power forecasting method under extreme weather conditions based on meta-learning and multi-head attention mechanisms is proposed in this paper. A base learner based on CNN-Attention-GRU is constructed at first. The samples under extreme weather conditions from nearby wind farms are used to pre-train the forecasting model to make up the lack of representative samples. Then, a meta-learner based on XGBoost is developed to perform targeted training using extreme weather data from the target wind farm to improve forecasting accuracy. Furthermore, the multi-head attention mechanism is used to enhance the learning capabilities of base learner and meta-learner. Operational data of eight wind farms affected by cold wave / typhoon weather conditions is used to verify the effectiveness and robustness of the proposed method. The results show that, when root mean square error is used as the evaluation index, the forecasting error can be reduced by 0.13%-4.29% under cold wave conditions, 0.35%-9.84% under typhoon conditions.