Comparison of physics-based and machine learning methods for phase-resolved prediction of waves measured in the field

Abstract

Phase-resolved predictions of surface waves can be used to optimize a wide variety of marine applications. In this paper, we compare predictions obtained using two independent methods for field data, with horizons sufficient to control wave energy converters.

The first method is physics-based prediction. In this method, a set of optimal representative angles, obtained using an optimization algorithm given time histories of a wave buoy motion in 3D, are used for forward propagation based on linear wave theory. The second method is a machine learning method using an Artificial Neural Network (ANN) which requires longer records for training.

Field measurements were obtained from the Southern Ocean of Albany, WA. The field data were collected by an upwave ‘detection’ array of 3 Sofar Spotter wave buoys and a downwave ‘prediction’ point coincident with a Datawell Waverider-4. All buoys were soft-moored, and data were collected over 3 months in 2022. Selected intervals during this period are presented in the paper to compare and contrast the predictions made by the two different methods. We find that some wave fields can be predicted well over more than a period in advance, all that is required for active control of a wave power take-off in a renewable energy application. In contrast, highly spread sea states remain a challenge. The methods are also compared in terms of the complexity and time required for making predictions. Further discussions are made on the applicability of the results to other locations.