Floating offshore wind turbine model and validation and parameter estimation

Abstract

Recently, the forecast of the dynamic response and mooring tension of floating offshore wind turbines (FOWTs) has received a lot of attention from researchers. This focus is significant for maintaining seakeeping function and ensuring safety for FOWTs installed in deep water. Nevertheless, the conventional numerical and experimental approaches are not economical. This constraint makes them unreliable for generating the extensive datasets needed. This study presents an accurate deep learning-based approach, which refers to a Convolutional Neural Network (CNN) and directional Long Short Term Memory (Bi-LSTM), to predict the tensions in three mooring lines by mapping the time series data (wave elevation and 6 DOF platform response) to time series representing mooring tension. The model is applied to the UMaine VolturnUS-S Reference Platform developed for the IEA Wind 15 Megawatt Offshore Reference Wind Turbine. The proposed model merges CNN and BiLSTM to extract the coupling relationships and capture the temporal dependence among different inputs, respectively. A hybrid CNN+BiLSTM model was trained, validated, and tested on {1188} time series simulations from OrcaFlex with 70\%, 15\%, and 15\%, respectively, considering multiple sea state conditions. The model achieved acceptable accuracy with RMSE of \num{0.0288} and a predictive accuracy $R^2$ of \num{0.9631}, with a total training time of \SI{12263.1} $s$. To verify the feasibility of the CNN+BiLSTM network, the model was tested on unseen data. However, it experiences catastrophic spikes under extreme loading conditions and struggles to capture the low-frequency drift components. These occur due to strong nonlinearity and non-stationarity in the time series data. To simplify these challenges, the variational mode decomposition (VMD) method has been integrated into the network as a preprocessing step, separating drift-dominated trends from faster wave-induced oscillations. Integrating VMD ($K$ = \num{3}, $\alpha$ = \num{2000}) doubles the average error, with an RMSE of \num{0.0614} (normalized unit), but preserves the $R^2$ value of \num{0.9627}. Interestingly, this approach eliminates all extreme failures and effectively captures the low-frequency mooring tension forces, which are predominantly driven by second-order wave drift. However, it struggles to capture nonlinear wind-induced oscillation loads. Also exhibits a significant computational cost, with a training time of \num{375673.86} $s$. Hence, it is recommended to add more features to the model (such as wind data, tower data, and turbine data), customize VMD parameters (K and $\alpha$), integrate Physics-Informed Neural Networks (PINN), and transfer learning for better generalizability across platforms for future work.