Phase-resolved real-time forecasting of three-dimensional ocean waves via machine learning and wave tank experiments

Accurate prediction of ocean waves plays an essential role in many ocean engineering applications, such as the control of wave energy converters and floating wind turbines. However, existing studies on phase-resolved wave prediction using machine learning mainly focus on two-dimensional wave data, while ocean waves are usually three-dimensional. In this work, we investigate, for the first time, the phase-resolved real-time prediction of three-dimensional waves using machine learning methods. Specifically, the wave prediction is modeled as a supervised learning task aiming at learning mapping relationships between the input historical wave data and the output future wave elevations. Four frequently-used machine learning methods are employed to tackle this task and a novel Dual-Branch Network (DBNet) is proposed for performance improvement. A group of wave basin experiments with nine directional wave spectra under three sea states are first conducted to collect the data of 3D waves. Then the wave data are used for verifying the effectiveness of the machine learning methods. The results demonstrate that the upstream wave data measured by the gauge array can be used for control-oriented wave forecasting with a forecasting horizon of more than 20 s, where the directional information provided by the upstream gauge array is vital for accurately predicting the downstream wave elevations. In addition, further investigations show that by using only local wave data (which can be easily obtained), the very short-term phase-resolved prediction (less than 5 s) can be achieved.