Physics-Informed Residual Attention GRU for Short-Term Six-DOF Platform Motion Prediction of Floating Offshore Wind Turbines

Abstract

Accurate short-term prediction of six-DOF platform motion is important for the state monitoring and health management of floating offshore wind turbines. This paper proposes a physics-informed residual attention GRU model, PI-Res Attn-GRU, for short-term six-DOF platform motion prediction. The model uses the preceding 10 s of platform motion to predict the subsequent 5 s. First, a GRU temporal encoder transforms the historical six-DOF motion sequence into a sequence of hidden representations. An Attention-Last State Fusion Module then combines an attention-pooled feature capturing key historical dynamics with the final hidden state representing the latest motion state, thereby forming a context representation for future prediction. A Step-Delta incremental decoder starts from the last observed state and recursively predicts motion increments to generate the future motion sequence. In addition, physics-informed constraint terms are incorporated into the training loss to promote boundary position continuity, initial velocity consistency, and acceleration smoothness. Experiments are conducted on 22 floating offshore wind turbine simulation cases. An eight-fold leave-one-wind-speed-out cross-validation procedure is applied to the development set for model selection, while cases at wind speeds of 8, 16, and 24 m/s are held out exclusively for testing. The final model achieves an MAE of 0.0333, an MSE of 0.0046, an RMSE of 0.0681, and an R² of 0.9939 on the test set. Compared with LSTM, the best-performing baseline model, PI-Res Attn-GRU reduces the MAE, MSE, and RMSE by 54.1%, 72.0%, and 47.1%, respectively. The results demonstrate that the proposed model can accurately predict short-term six-DOF platform motion and generalize well across different operating conditions.