A Forward Surrogate-Assisted Physics-Informed Framework for OpenFAST-Based Wind Turbine Model Updating

For wind turbine condition monitoring and digital twin applications, OpenFAST-based parameter updating provides high-fidelity dynamic constraints for model calibration and response consistency evaluation. However, efficiently incorporating physical information into data-driven model training remains a key challenge. In a supervised learning framework, introducing physics-consistency constraints requires the responses associated with the network-predicted parameters. These responses are used to compute frequency- and waveform-related loss terms. Evaluating these responses directly using OpenFAST would require repeated high-fidelity time-domain simulations during training, resulting in a substantial computational burden. To address this issue, a forward surrogate model is proposed for OpenFAST model updating. It replaces high-fidelity response evaluation during training and enables efficient computation of physics-informed loss terms. The proposed surrogate model takes the parameters to be updated as inputs and predicts the corresponding system responses. These outputs include frequency-domain features and time-series waveforms, thereby approximating the forward mapping of OpenFAST. Based on simulation datasets generated from parameter variations, the surrogate model is trained to capture the relationship between structural parameters and dynamic responses. It is then embedded into the loss calculation process of the parameter prediction network, providing frequency and waveform losses without directly calling OpenFAST. The results show that the proposed method can accurately predict both frequency-domain and time-domain responses of the high-fidelity model while substantially reducing the computational burden of physics-informed loss evaluation. After applying the surrogate model, the responses obtained by rerunning OpenFAST with the predicted parameters show improved agreement with the target responses, with the response prediction MAE reduced by 9.6% and the R² increased by 10.1%.