Comparison of statistical extrapolation methods for the evaluation of long-term extreme response of wind turbine

Prediction of extreme responses with given mean recurrence intervals (MRIs) is an essential task for reliability-based design of wind turbines. It involves statistical extrapolation of extreme response distribution at very upper tail region for the estimation of long-term extreme responses from short-term field or simulation data. In this study, the accuracy and effectiveness of the global maxima method, translation process method and average conditional exceedance rate (ACER) method in evaluating extreme response distribution are examined through comparisons with the predictions from controlled Monte Carlo simulation (MCS). The controlled MCS combines the importance splitting (ISp) method with multivariate autoregressive (MAR) modeling of stochastic wind excitations, enables a direct simulation of long-term extreme responses with much reduced computational efforts. The extreme value distribution determined provides a valuable opportunity to verify the adequacy of different statistical extrapolation approaches. This study reconfirms the limitation of global maxima method that the predictions are greatly relying on the prescribed probability distribution model. This study also highlights the challenges in translation process method for strongly non-Gaussian response process when moment-based Hermite model is used for representing the translation function between non-Gaussian response process and underlying Gaussian process. An improved Hermite model that better represents the translation function at upper tail is proposed, and its improved performance in the estimated extreme value distribution of strongly non-Gaussian response process is achieved. The ACER method, which is essentially a method of extrapolating the process crossing rate, is also found to be able to well represent the tail behavior of the extreme value distribution.