TITLE: Toward data-driven surrogate models of wind turbine response for digital twin applications AUTHORS: Briac Malbois Le Borgne*, Filippo Gatti*, Rodrigue Desmorat* and Matteo Capaldo† ADDRESS: * Université Paris-Saclay, CentraleSupélec, ENS Paris-Saclay, CNRS LMPS -- Laboratoire de Mécanique Paris-Saclay UMR 9026 91190 Gif-Sur-Yvette, France e-mail: briac.malbois_le_borgne@ens-paris-saclay.fr † TotalEnergies -- OneTech 91120 Palaiseau, France ABSTRACT Key words: SCADA, wind turbine, fatigue, machine learning The transition to renewable energy is progressing rapidly. To maintain economic viability and ensure continued growth, it is crucial to accurately estimate system performance and implement strategies to extend operational lifetimes. In the context of wind turbines, the operating strategy is a key driver of overall performance and directly influences fatigue[1]. Many contemporary turbine simulations rely on algorithms with manually tuned parameters[2]. This requires time and expert knowledge. We propose data-driven approaches, using SCADA (supervisory control and data acquisition) from wind turbines, that aim to estimate the turbine's operational response, enabling faster iteration for digital twin development[3,4]. Wind turbines are continuously regulated to maintain optimal performance across varying conditions. Due to operational constraints, SCADA are usually not sampled at high frequency, and do not capture the full operational dynamics. This limits our ability to characterize the turbine's response[3]. To address the limitations of onsite SCADA, we suggest a deep learning framework using rotor speed to model the turbine's input-output response, in line with recent data-driven surrogate models[4]. Such options allow us to adapt to different scenarios, where available SCADA fields are variable. Our baseline random forest model provides estimates, capturing global response in regions~2 and~3 in synthetic data. For better accuracy, and the complex transition region~2.5, we leverage a deep learning model on synthetic data sampled at both 40\,Hz and 1\,Hz. Results show promising performance in autoregressive mode in synthetic data, using only rotor speed as exogenous variables. This approach represents a step towards a black-box approach yielding