Ensemble-based online meta-learning approach for ultra short-term wind power forecasting

Real-time power system operation under high renewable energy penetration introduces a great challenge for ultra short-term wind power forecasting (WPF), particularly across multiple lead times. Traditional WPF models are usually trained in the offline mode, which would degrade over time due to the non-stationary characteristics of wind energy resources. Accordingly, recent efforts have introduced online learning to incorporate new data and update the prediction model in real time. However, these methods still suffer from instability caused by frequent updates and limited ability to capture temporal correlations across multiple lead times. This paper proposes an ensemble-based online meta-learning approach that fully utilizes the most recent data across multiple lead times of ultra short-term WPF. A two-loop inner-outer updating strategy is proposed, where the inner loop performs rapid updates for each lead time, and the outer loop updates shared meta-parameters to improve stability across multiple lead times. The proposed framework effectively adapts to time-varying wind patterns and captures dependencies across multiple lead times. In addition, it is computationally efficient and requires little data storage. Experiments on real-world datasets verify that the proposed method significantly outperforms conventional offline and online methods in forecasting accuracy and consistency for ultra short-term WPF.