Learning wake perturbations of floating wind turbines using motion-conditioned neural network

Platform motion in floating offshore wind turbines alters wake recovery and coherent structures, changing the inflow experienced by downstream turbines. High-fidelity CFD can resolve these effects, but repeated simulations across many motion conditions remain computationally expensive. Moreover, in direct full-field surrogate prediction, the motion-induced wake response is learned alongside the dominant fixed-wake background, making motion-related changes less explicit. To address this, wake prediction is formulated here as a perturbation-learning problem, in which the moving-turbine wake is decomposed into a time-resolved fixed-turbine reference wake and a motion-induced perturbation. A motion-conditioned recurrent neural network is trained to predict this perturbation from the recent fixed-wake history, perturbation history, and motion information, including motion type, amplitude, Strouhal number, and phase. The moving wake is then reconstructed by adding the predicted perturbation to the fixed-reference wake. The method is tested on planar wake slices extracted from high-fidelity CFD simulations of the MoWiTO 0.6 model floating wind turbine undergoing prescribed pitch and roll motions at different amplitudes and frequencies. Under the evaluated fixed-reference formulation, the normalized velocity-deficit perturbation RMSE is 49.1% lower than for the direct full-field baseline across the representative pitch and roll cases. The corresponding wake-center error is 43.6% lower. The predicted perturbation fields enable spatial identification of where platform motion modifies the wake deficit, shear-layer structure, wake recovery, and wake displacement. Wake-diagnostic results also show lower errors in downstream velocity profiles, maximum deficit, and wake width. The results indicate that perturbation learning can provide a more interpretable wake-reconstruction approach for floating wind turbine wake modeling.