Simple Wind Speed Descriptors Enough for Accurate Turbine Fatigue Load Models, Preprint Finds

Wind turbine fatigue loads are central to how wind farms are designed, optimized and controlled, but calculating them with high-fidelity aeroelastic simulations is computationally expensive. Load surrogate models have emerged as a promising shortcut, promising fast predictions at a fraction of the cost. Yet, according to a new preprint posted in Wind Energy Science Discussions, most previous work has focused on proving that such models can be accurate, while far less attention has gone to how they should actually be designed and deployed in practice.

That gap is the target of a study by Adrien Guilloré, Andreas Vad, Filippo Campagnolo and Carlo L. Bottasso of the Wind Energy Institute at the Technical University of Munich. Their preprint, under review for the journal Wind Energy Science, systematically investigates the design and implementation of location-agnostic, control-oriented surrogate models for wind turbine fatigue loads, asking not just whether a model can predict loads but which ingredients it genuinely needs.

To answer that question, the authors built a comprehensive database of aeroelastic simulations spanning free-stream, partially waked and fully waked operating conditions. Against this database they tested alternative inflow representations, different control descriptors, varying numbers of turbulent realizations and different training dataset sizes, effectively sweeping the design space of the surrogate model rather than evaluating a single configuration.

The study’s headline finding is that there exists a finite level of physical complexity sufficient to achieve converged performance for a general wind turbine load surrogate model. In other words, adding more physical detail beyond a certain point stops buying predictive accuracy, a result with direct implications for how much information such models need to carry.

Among the options investigated, the simplest formulation capable of reproducing converged predictive accuracy turned out to be an inflow representation the authors denote SAWS 4A + RATI: four sector-averaged wind speed quantities together with one rotor-averaged turbulence parameter. That compact, physically meaningful set of descriptors was enough to match the performance of more elaborate representations in their tests.

The control side of the model proved similarly economical. Steady-state control setpoints were shown to provide a sufficient description of the turbine’s control actions, meaning the surrogate did not require richer dynamic control information to reach its converged level of accuracy.

The analysis also quantified how much simulation data the models need. Roughly 10 to 12 turbulent realizations were required for robust convergence with respect to turbulence variability, and surrogate performance converged with a relatively modest number of training cases, suggesting that building such models need not demand enormous simulation campaigns.

Because the model is intended to be location-agnostic, the authors also examined how it behaves when deployed differently from how it was trained. They assessed the surrogate formulation using multiple inflow-estimation approaches, including observer-based inflow reconstruction and steady-state wake models. The results showed that significant performance degradation occurs when training and deployment rely on different inflow-description spaces, but that high predictive accuracy is consistently recovered when the surrogate is trained and deployed using the same inflow representation.

Notably, the same design recommendations held across all investigated implementation environments, indicating that the guidelines are not artifacts of one particular setup. For practitioners, the study offers a practical recipe: a small set of reduced-order descriptors, a consistent inflow representation between training and deployment, steady-state control setpoints, and a moderate number of turbulent realizations and training cases.

As a preprint currently under open review, with discussion open until 20 November 2026, the findings are still subject to peer scrutiny. Even so, the work demonstrates, in the authors’ framing, that accurate, robust and computationally efficient fatigue load prediction can be achieved with a small set of physically meaningful reduced-order descriptors, and it provides concrete design guidelines for the next generation of load surrogate models used in wind farm design and control.

Subject of Research: Design and deployment of location-agnostic, control-oriented surrogate models for wind turbine fatigue loads

Article Title: Design and implementation of a location-agnostic and control-oriented load surrogate model for wind turbines

Article References: Guilloré, A., Vad, A., Campagnolo, F., & Bottasso, C. L. (2026). Design and implementation of a location-agnostic and control-oriented load surrogate model for wind turbines. https://doi.org/10.5194/wes-2026-191

Image Credits: AI Generated

DOI: 10.5194/wes-2026-191

Keywords: wind energy, wind turbines, fatigue loads, surrogate models, aeroelastic simulations, wake conditions, inflow estimation, control setpoints, reduced-order models, wind farm design, computational efficiency, preprint