Prediction of FOWT mooring tension during non-stationary typhoon transit: a Transformer based framework with Bayesian optimization

Abstract

The dynamic evolution of typhoons is marked by pronounced non-stationarity and extreme nonlinearity, featuring rapid wind speed fluctuations that critically threaten the structural integrity of Floating Offshore Wind Turbines (FOWT). Conventional steady-state simulations frequently inadequately represent the transient coupling effects induced by a rapidly transiting typhoon. Therefore, a 12-h non-stationary typhoon model is developed to explicitly capture the continuous “approach-peak-decay” lifecycle, thereby ensuring physical consistency in the simulated aerodynamic load response. Furthermore, a deep learning framework is developed to collaborate Transformer architecture with Gaussian process-based Bayesian optimization (GP-BO) strategy. The framework leverages mooring tension response data generated throughout the entire non-stationary typhoon transit to train and construct a high-precision surrogate model. Comparative analysis with the traditional Long Short-Term Memory with Self-Attention mechanism (LSTM-SA) model demonstrates that the established predictive model exhibits exceptional generalization capability, enabling accurate prediction of dynamic responses under previously unencountered sea conditions, with a correlation coefficient exceeding 0.90 against true values. This research provides a highly efficient and reliable surrogate modeling approach for the real-time safety monitoring and design optimization of FOWT mooring systems under extreme meteorological conditions.