Frequency domain approach has been widely adopted to evaluate fatigue damages of offshore structures due to its efficiency although time domain approach is more accurate. It is assumed that in that approach, the stress response spectra induced by wind or wave load can be expressed by a stress transfer function. However, since structural response of a wind turbine to different wind speeds is non-linear, the stress transfer functions could change with wind speed. This means that repeated simulations are needed in order to calculate the stress transfer function according to wind speed change. The problem, though, is that if the number of simulations is large, prohibitively high computational and time costs probably will be incurred.
In this study, to reduce the number of simulations and, at the same time, increase the accuracy of results, a regression analysis is performed to obtain the stress transfer function induced by wind load by using artificial neural network. Sensitivity analysis was conducted to determine how many sample points are required and how to select them. The total stress spectrum were calculated by summing stress spectrum induced by wind load from the ANN model and induced by inertia load from motion analysis based on linear wave theory. Numerical analysis was performed to verify the performance of the proposed regression model.
A fatigue analysis for offshore floating wind turbines is a highly complex problem due to nonlinearity of excitations, large platform motions and mooring forces. For this structure, time domain approach which calculates hot spot stress in each time step has been widely used to estimate fatigue damages at wind turbine supporting structures. However, it requires a lot of computational time and cost so fatigue analyses should be performed in a limited representative environmental conditions. This makes it difficult to optimize the structures at the initial stage of design so studies have been conducted to apply frequency domain techniques to predict the fatigue life more efficiently (Van der Tempel, 2006; Bachynski and Moan, 2012; Kvittem and Moan, 2015).