Motion prediction of semi-submersibles using time-frequency deep-learning model with input of incident waves

Abstract

Machine learning techniques have inspired reduced-order solutions in hydrodynamic response prediction and hold the potential to provide valuable insights for structural design and real-time monitoring. In this study, to improve the feature extraction capability and interpretability, the time-frequency analysis methods were applied to the architectural design of deep learning model, and the incident waves were fed into the time-frequency (TF) model to predict the motions of semi-submersibles subjected to head waves. The training and testing datasets were derived from a scaled model test. Upon validation, the internal visualization demonstrated that the TF model can provide an accurate minute-level prediction with great interpretability. An in-depth analysis of time window selection was conducted to improve the model's performance and stability. Furthermore, multi-step predictions were systematically explored, successfully reconstructing the full 3-h motion time series with an average accuracy exceeding 90% for heave, pitch, and surge. The model's generalization ability was also evaluated under unfamiliar wave conditions and with limited training datasets. It is envisioned that the TF model could be a promising alternative to provide precise and rapid predictions of the hydrodynamic performance of offshore platforms.