Presentations

Abstract

Physics-based wake models produce a wide range of wake loss predictions [1], leading to uncertainties in energy production estimates [2]. Extensive validation against operational data is needed to guide model choice depending on site conditions [3]. While machine learning is beneficial to reduce costs or improve accuracy by combining models and leveraging operational data, current applications usually focus on one aspect. Surrogates of high-fidelity wake models [4] reduce costs under similar accuracy. Ensembles increase robustness and provide uncertainty estimates, although methods to select and combine models are lacking. Finally, site-specific models trained from single site data have been proposed [5] but may lack generalization to unseen sites. This raises the question: leveraging large operational datasets, what would be a suitable machine learning framework to combine the strengths of different models, to improve predictive accuracy?  Recent work [6] demonstrated the use of turbine-level wind speed deficits obtained from engineering wake models as features to predict higher-fidelity simulations. These features encode layout-dependent wake effects at the turbine level, which allows generalization. In the present work, we further develop this concept using year-long operational datasets from 7 wind farms, representing 225 turbines and 1GW of installed capacity. The resulting hybrid wake modeling framework is independent of a specific machine learning or physics-based wake model. In the present case, features are derived from Park2 [7], TurbOPark [8], and Zong Gaussian [9] models, and regression is performed against observed wake deficit timeseries using XGBoost [10].  Adopting a leave-one-site-out cross-validation methodology to assess generalization of the model across sites, we show that the approach improves prediction skill by over 40% compared to engineering wake models. We also observe that the dependency of the prediction on turbulence intensity is consistent with wake physics. We propose avenues to further improve prediction skill and reduce uncertainties in wake modeling.