Hindcast of Significant Wave Heights in Sheltered Basins Using Machine Learning and the Copernicus Database

Abstract

Long-term time series of wave parameters play a critical role in coastal structure design and maritime activities. At sites with limited buoy measurements, methods are used to extend the available time series data. To date, wave hindcasting research using machine learning methods has mainly focused on filling in missing buoy measurements or finding a mapping function between two nearshore buoy locations. This work aims to implement machine learning methods for hindcasting wave parameters using only publicly available Copernicus data. Ensemble regression and artificial neural networks were used as machine learning methods and the optimal hyperparameters were determined by the Bayesian optimization algorithm. As inputs, data from the MEDSEA reanalysis wave model were used for the wave parameters and data from the ERA5 atmospheric reanalysis model were used for the wind parameters. The results of this study show that the normalized RMSE of the test data improved by 29% for Rijeka and 12% for Split compared to the original MEDSEA wave hindcast at buoy locations. The proposed method was extremely efficient in removing bias in the original MEDSEA hindcasts (e.g., NBIAS = -0.35 for Rijeka) to negligible values for both Split and Rijeka (NBIAS < 0.03).