An Ultra‐Short‐Term Wind Power Forecasting Method Based on Adaptive Cleaning of Streaming Data and Differentiating of Input Feature Contributions

ABSTRACT Ultra‐short‐term wind power forecasting plays a crucial role in real‐time dispatching, frequency regulation, and intraday electricity market transactions. Forecasting accuracy heavily depends on data quality and feature informativeness. However, most existing studies conduct data cleaning offline, with limited attention to real‐time data quality during forecasting. Moreover, they often use historical power and NWP data uniformly, neglecting the time‐varying importance of input features. To address these issues, this paper proposes an ultra‐short‐term wind power forecasting method based on dynamic cleaning of streaming data anomalies and adaptive processing of input feature contributions. Firstly, similar samples of the current wind process are retrieved online via time series similarity matching, enabling real‐time anomaly detection in streaming data. Secondly, anomalous power sequences are reconstructed using a theoretical restoration model based on wind speed fluctuation identification. Finally, a forecasting architecture with personalised encoding and dynamically fused decoding is designed to enhance prediction accuracy. The proposed method has been successfully applied to a wind‐solar‐storage power station in Inner Mongolia, supporting both grid dispatching operations and daily maintenance. Compared to baseline methods, it achieves average reductions in forecasting errors of 0.59–9.99 percentage points for RMSE and 0.62–8.49 percentage points for MAE.