Session: 08-04-01 Deep Learning based ROM for Wave Dynamics and Prediction in Marine Engineering

Paper Number: 127930

127930 - Application of Machine Learning and Deep Learning for Enhanced Spatiotemporal Wave Parameter Prediction 

Traditional methods of wave prediction, which are mainly reliant on extensive CFD simulations, such as the utilization of spectral wave models like SWAN, WAVEWATCH III, or TOMAWAC, have prompted the question: Can faster wave prediction be achieved? The answer, as demonstrated by this study, lies in the advancements of machine learning and deep neural networks.

In this research, the spatiotemporal relationship between wind and wave conditions is established using the XGBoost machine learning method and deep neural networks. This approach enables effective predictions of wave height, wave period and other parameters within the waters of the North Atlantic and North Sea.

Six years of hourly wind data from ECMWF ERA5 (2016-2021) is used as input features, while wave parameter observations from Cefas WaveNet buoys and the previously validated TOMAWAC wave data, developed by the authors, are employed for model training and verification. The final output features enable a comparison that ultimately leads to wave predictions for the year 2022. Building upon this foundation, a versatile model for typical weather conditions and a specialized model for extreme weather scenarios are devised, facilitating more precise predictions. Furthermore, the consideration of tidal currents as input features are introduced, particularly in areas where waves and tidal currents coexist, revealing the modulation effect caused by tidal currents within the wave predictions.

The data-driven model, rooted in wind data, proves adept at predicting wave characteristics across different times and locations and investigates wave-current interactions. Notably, the trained machine learning and deep learning model delivers significant efficiency gains compared to traditional CFD models. One year's worth of data can be predicted in just a few seconds, whereas over 24 hours are required (on 16 logical CPUs) by the TOMAWAC wave model. This leap in training efficiency is a crucial development in the realm of wave prediction. The full version of the paper will provide all the details behind the techniques used and a discussion of the results.

Presenting Author: Tian Tan University of Edinburgh

Presenting Author Biography: Tian is presently a PhD candidate at the University of Edinburgh while concurrently a part-time research assistant on the EPSRC Supergen-funded WavE-Suite project. Her research covers the ocean-scale numerical modelling of waves and tidal currents, alongside tank experiments of wave energy converters. Her current passion lies in the application of machine/deep learning methods for the spatiotemporal prediction of wave parameters.

Authors:

Tian Tan University of Edinburgh
Vengatesan Venugopal Institute for Energy Systems, School of Engineering, The University of Edinburgh