Surrogate model uncertainty in wind turbine reliability assessment

Abstract

Lowering the cost of wind energy entails the optimization of wind turbine material consumption without compromising structural safety. Traditionally, wind turbines are designed by the partial safety factor method, which is calibrated by probabilistic models and presented in the IEC 61400-1 design standard. This approach significantly reduces the amount of aero-elastic simulations required to assess the fatigue limit state of wind turbines, but it may lead to inconsistent reliability levels across wind farm projects. To avoid this, wind turbines may be designed by probabilistic methods using surrogate models to approximate fatigue load effects. In this approach, it is important to quantify and model all relevant uncertainties including that of the surrogate model itself. Here we quantify this uncertainty according to Eurocode 1990 for polynomial chaos expansion (PCE) and Kriging using wind data from 99 real sites and the 5 MW reference turbine designed by NREL. We investigate a wide range of simulation efforts to train the surrogate models. Our results show that Kriging yields a higher accuracy per invested simulation compared to PCE. This improved understanding of utilizing PCE and Kriging in fatigue reliability assessment may significantly benefit decision support in probabilistic design of wind turbines.