Developing a robust wind power forecasting method: Integrating data repair, feature screening, and economic impact analysis for practical applications

Advances in wind power forecasting technology, critical for energy conversion rates and grid stability, are central to modern power management systems and play a key role in mitigating global climate change. However, practical wind power forecasting faces challenges such as large amounts of anomalous data, stochastic noise, redundancy in meteorological forecasts, and historical underestimation of wind power costs. To address these issues, we introduce a comprehensive wind power forecasting framework, ARFEAT, which integrates Anomaly Repair, Feature Enhancement, Asymmetric loss optimization, and Transformer-based modeling. First, an anomaly detection and repair module based on semi-supervised learning is introduced to guide subsequent model learning and improve noise immunity. Then, an efficient feature screening module is designed to identify high-value features from hundreds of features, reducing meteorological feature redundancy. Finally, we introduce a practical wind farm economic cost assessment index and a transformer model based on asymmetric functions, which can effectively learn wind power trends and significantly reduce costs. Applying ARFEAT to two commercial wind farms, inland and coastal respectively, improves prediction accuracy by an average of 19% and reduces economic costs by 36%, demonstrating robust performance. These results highlight the utility of ARFEAT in improving wind farm profitability and renewable energy grid integration.