A data-efficient transfer learning framework with multi-fidelity simulation for predicting wind-induced responses of offshore wind turbines

Accurately predicting wind-induced responses is critical for the design and operation of offshore wind turbines (OWTs), which is often restricted by the scarcity of field data and the computational cost of high-quality simulations. To overcome this challenge, this study proposes a novel transfer learning framework that enables accurate predictions by efficiently leveraging abundant, lower-cost simulated data and a small amount of high-quality targeted data. The framework employs a two-stage strategy: first, a deep temporal model (LSTM) is pre-trained on a large dataset generated through efficient multi-physics simulations (OpenFAST) to learn generalized dynamic representations. Second, the model is adapted using a very limited set of high-quality data derived from detailed fluid-structure interaction (FSI) simulations to refine its predictions toward targeted accuracy. The results demonstrate that the proposed framework can achieve high-precision predictions of tower-top displacement responses with a minimum mean absolute percentage error of 10.46% under unseen wind speed condition, without the need for massive high-quality data. Meanwhile, it significantly reduced computational and data acquisition costs. The proposed transfer learning framework provides a versatile and data-efficient paradigm not only for OWT response prediction but also for broader structural health monitoring and digital twin applications where high-quality data are scarce.