A Novel Numerical Approach for Fatigue Load Emulation of Offshore Wind Turbines Using Machine Learning

Abstract

During their lifetime, offshore wind turbines are subjected to complex loading conditions. Understanding the loading history of the support structure of these structures and the long-term implications on the structural integrity renders structural health monitoring (SHM) imperative. In that direction, it is common practice in the wind turbine industry to use only a few physical sensors in a small number of intelligent turbines per fleet to directly measure the load levels in the support structure, implying SHM insights limited to only the instrumented turbines. This often leaves stakeholders with insufficient information on a farm level, not facilitating important decisions regarding project lifetime or adjustments to the trade-off between energy harvest and fatigue damage accumulation through upgraded controller behavior. To bridge this gap, machine learning models have been used as surrogate models with the purpose of virtually sensing fatigue loads through a process known as load emulation. However, previous applications have used site-specific, position-specific training data to develop the machine learning models. As such, the resulting surrogate model decreases in performance when asked to extrapolate to a different site, or even to a different position of the same wind farm. This research presents a novel methodology for constructing a single machine learning load emulation object that is useful for emulating fatigue loads in a generic concept, applicable to any wind turbine within a wide, pre-defined solution space without loss in accuracy. The methodology introduces the use of simplified structural models which preserve key degrees of freedom of the geometric and dynamic properties to represent potential offshore wind sites that cover the desired solution space. These structural models are used to run 10-minute, hydro-servo-aeroelastic numerical simulations and the results are used to train the machine learning model. Load emulation is performed at the foundation-tower interface level. Finally, the surrogate model is validated by comparing the emulated loads from the novel simplified approach against the detailed loads from the numerical assessment of site-specific turbine models at two offshore sites in the North Sea.