A machine learning model for 2D significant wave height forecasts around Scotland

Abstract

With the expansion of the marine renewable energy industry comes the need for quick and reliable marine forecasts. Significant wave height is one of the most important parameters for making decisions in the operation and maintenance phase of projects. Traditional numerical wave models (e.g. SWAN) require large computational resources, having an impact on costs and time. Data-driven models, on the other hand, are faster and lighter. Machine learning tools have been employed for forecasting wave heights for the past two decades, with a broad range of models and locations, although most studies focus on a single spatial point. With deep learning came the move to 2D models. However, most research has been carried out in wind-wave dominated environments in the western Pacific. This study presents a novel spatio-temporal model covering Scottish waters, where a range of conditions are present, from mixed sea conditions to swell-dominated seas. The proposed model, a ConvLSTM model, takes previous hours of wind data from the ERA5 reanalysis dataset as an input to forecast significant wave height in the future. Since the model is in two dimensions, it can capture spatial as well as temporal dependencies. The non-linearity of waves makes this a complex problem. However, the results demonstrate that it is possible to forecast wave heights in different conditions with a high degree of accuracy, particularly in the short term (1 to 6 hours), with root mean square errors below 50 cm, proving this to be a useful tool for wave forecasting at marine energy sites.