Importance sampling of floating offshore wind turbine fatigue environments using machine learning

Abstract

Floating offshore wind turbines (FOWTs) are subjected to complex environmental loading arising from the interaction between wind, waves, currents, structural dynamics, and control-system response. Fatigue assessment of such systems typically relies on large numbers of time-domain simulations performed for environmental conditions extracted from long-term hindcast databases. Due to the computational cost of fully coupled FOWT simulations, only a subset of the available environmental conditions can be simulated, making environmental-condition selection a critical aspect of fatigue assessment.

This thesis investigates the use of machine-learning-assisted importance sampling to improve the efficiency of fatigue damage estimation for the \gusto\ Tri-Floater concept. A benchmark database comprising 11,571 time-domain fatigue simulations was used to study fatigue behavior at 31 structural hot-spots and to evaluate alternative sampling strategies. The proposed methodology employs machine learning models to predict fatigue damage from environmental conditions and derives importance sampling density functions from the resulting damage contributions. Two machine learning approaches were investigated: Extreme Gradient Boosting (XGBoost) and symbolic regression through PySR.

The results show that fatigue accumulation behavior varies significantly between hot-spots. While most locations accumulate damage gradually over a broad range of environmental conditions, a smaller subset is dominated by highly localized and infrequent events. Wind speed and significant wave height were identified as the dominant fatigue drivers, while current-related parameters exhibited negligible influence on fatigue damage. XGBoost consistently outperformed PySR in both fatigue-damage prediction and reconstruction of damage-contribution distributions.

The XGBoost-derived importance sampling density functions significantly improved convergence toward the benchmark fatigue damage while simultaneously reducing statistical sampling uncertainty relative to equally distributed Monte Carlo sampling. For most hot-spots, reliable fatigue estimates were obtained using substantially fewer simulations than required by conventional sampling approaches. The results demonstrate that machine-learning-assisted importance sampling is a promising methodology for reducing the computational cost of FOWT fatigue assessment while maintaining accuracy and providing quantifiable uncertainty estimates.