An XAI framework for robust and transparent data-driven wind turbine power curve models

Transparency in AI improves model reliability: The paper demonstrates that explainable AI (XAI) methods can reveal whether machine learning models for wind turbines are learning physically meaningful patterns rather than just fitting noise, which is critical for performance in real-world, changing conditions.

Standard error metrics alone are insufficient: The research shows that looking only at test set errors can be misleading—learned strategies and interpretability checks provide better indicators of whether a model will generalize well, particularly when working with limited data.

Physics-informed validation catches problematic models: By comparing machine learning models against simple physics-based baselines, the framework can automatically identify when complex models are learning unreliable patterns, helping practitioners avoid deploying models that may fail in unexpected conditions.

Pipeline decisions matter and should be visible: The paper illustrates how choices made throughout model development—including data selection, preprocessing, and training parameters—significantly affect what strategies models learn, emphasizing the importance of transparency in these decisions for responsible AI deployment.

Explainability enables better decision-making: Understanding why a model makes predictions allows engineers and stakeholders to make more informed choices about which models to trust for critical infrastructure like wind turbines, rather than treating AI systems as black boxes.