Gaussian process regression for fatigue reliability analysis of offshore wind turbines

The fatigue limit state (FLS) often drives the design of offshore wind turbine (OWT) substructures in European waters. Assessing fatigue damage over the intended design life of an OWT is computationally expensive, primarily as dynamic structural analyses have to be run for a large number of stochastic wind and wave loading conditions. This makes structural reliability assessment for the FLS a challenging task. In addition to evaluating load-induced fatigue damage, simulation-based structural reliability analysis also requires sampling of random variables that model uncertainties in the capacity of OWT structural components. To this aim, we develop and validate a computational framework for OWT fatigue reliability analysis that relies on Gaussian process (GP) regression to build surrogate models of load-induced fatigue damage. We demonstrate that the proposed approach can reduce the computational effort required to evaluate FLS reliability with high accuracy through application to three plausible offshore wind farm sites in Europe. The sensitivity of various goodness-of-fit metrics to different model assumptions is investigated to further reduce the computational effort required to perform GP regression/predictions. The results from this study can provide guidance for practical applications of the proposed framework in OWT projects.

Automated summary

A computational framework relying on Gaussian process regression to build surrogate models of load-induced fatigue damage for offshore wind turbine (OWT) fatigue reliability analysis has been developed and validated in this study. The proposed approach can reduce the computational effort required to evaluate fatigue limit state (FLS) reliability with high accuracy through application to three plausible offshore wind farm sites in Europe. Further investigations on the sensitivity of various goodness-of-fit metrics to different model assumptions aim to reduce the computational effort required for GP regression/predictions.