Prediction of motions and mooring tensions for the OC3 spar in short-crested seas using a LSTM NN model, with application to fatigue damage assessment

This article presents a Long Short-Term Memory (LSTM) neural network model for predicting one-hour motions and mooring tensions of the OC3 5-MW spar under short-crested seas. Spectral spreading is synthesized via a cosine-power function, and the total wave elevation is provided to the network as a set of concurrent wave-elevation time series (early fusion), associated with the selected discrete wave propagation directions. The LSTM network was trained on an extensive dataset comprising over 300 one-hour fully coupled simulation scenarios, generated using OpenFAST. The neural network was then used to predict motions and mooring tensions for metocean conditions different from the ones used for training purposes. Good agreement with the directly simulated values was achieved, as evidenced by the high consistency of the time-series responses and the close alignment between the statistical distributions of responses from the surrogate model and direct simulation. Finally, fatigue assessment was performed at each fairlead with both the predicted tensions and the simulated reference data, finding similar computed annual damage. Once trained, the surrogate model delivers a substantial speed-up in computational time compared to direct simulation, enabling large-scale fatigue assessments at minimal computational cost. Thus, the presented workflow is reusable and scalable for a specific case study, supporting thousands of predictions and real-time monitoring use. Finally, aspects related to the overall computational effort involved in training and deploying the surrogate model are also critically addressed. Results provide insights that could facilitate the design process of offshore wind turbines, indicating potential for substantially accelerated analysis without sacrificing accuracy.