A measured-data-driven surrogate framework with hyperparameter optimization for fatigue damage prediction of FOWTs

Accurate fatigue damage prediction of floating offshore wind turbines (FOWTs) under complex wind-wave environments is essential for ensuring long-term structural safety and optimizing operation and maintenance strategies. To address the high computational cost of conventional time-domain fatigue analysis, this study develops a measured-data-driven artificial neural network surrogate framework with sequential hyperparameter optimization for fatigue damage prediction of FOWTs, in which grid search is used for network architecture selection and Bayesian optimization is employed for learning-rate optimization. Long-term measured wind-wave data from the Beishuang marine observation station in Fujian, China, were used to construct site-specific environmental inputs, which were then introduced into OpenFAST simulations to generate a physically consistent fatigue damage dataset through rainflow counting. Comparative results demonstrate that the BO-GS-ANN model outperforms conventional ANN, Kriging, and GS-ANN models. While the conventional ANN, Kriging, and GS-ANN models achieved R 2 values of approximately 0.50–0.53, 0.74, and 0.51–0.70, respectively, the BO-GS-ANN model further increased R 2 to 0.72–0.94, with the best configuration achieving an RMSE of 3.05 × 10 −4 and an MAE of 1.45 × 10 −4 .