Risk-Regularization Optimization for Learning Wind Power Curve

Wind power prediction requires an accurate estimation of the wind power curve. However, anomalous data in the SCADA dataset deteriorates the performance of regression models. This paper proposes a novel risk-regularization optimization-based method that formulates a chance-constrained optimization problem for training an interval neural network. This method ensures that normal data is preserved with high probability while anomalous data is removed. This approach overcomes the limitations of existing anomaly detection techniques that rely on hyperparameter tuning and reference datasets. To address the computational challenges of solving chance-constrained programs, we introduce a sample-based risk-regularization method that enables efficient optimization. Experimental validation using real-world wind turbine data demonstrates the superiority of the proposed method. The results show that our approach improves anomaly detection and power curve regression performance over conventional methods. These findings confirm the effectiveness of our approach in enhancing wind power curve estimation.