Scaled Physical Modelling of Floating Offshore Wind Turbines Using a Neural Network-Based Surrogate Model for Aerodynamic Emulation

Scaled physical modelling of floating offshore wind turbines is a crucial step in the continued development of floating offshore wind to reach Net-Zero goals. However, it is an inherently difficult task due to the incongruity between Reynolds scaling (important for wind effects) and Froude scaling (important for wave effects). One option to overcome this difficulty is to use a real-time hybrid testing approach, whereby a Froude-scaled model is tested in a wave basin with a thruster replacing the wind turbine. Typically, a low-fidelity numerical model is run in real time to determine the appropriate aerodynamic thrust applied by the thruster. In this paper, the real-time hybrid testing approach is extended by using a surrogate model, pretrained on numerical simulations, to determine aerodynamic thrust in the lab. It is found that for less complex wind load conditions, such as constant wind or turbulent wind with the wind turbine controller turned off, multiple linear regressions are satisfactory surrogate models, but for turbulent wind with the wind turbine controller turned on, artificial neural networks offer significant improvement. The development of the appropriate surrogate models is presented, and they are shown to predict thrust on a testing dataset with a mean absolute error over average thrust of (Formula presented.) 2% for operational environmental conditions and (Formula presented.) 6% for the most extreme environmental conditions tested. The methodology is demonstrated by testing a 1:70 scale of the VolturnUS-S platform with the 15MW IEA wind turbine, in the COAST Laboratory, University of Plymouth, in benign and extreme weather conditions representative of a location in the Celtic Sea. The novel methodology represents an improvement on the ability to accurately test floating offshore wind turbines in the laboratory, by enabling aerodynamics to be represented using higher fidelity numerical modelling, physical or field data.