Using random forest and gradient boosting trees to improve wave forecast at a specific location

The main objective is to present alternative algorithms to neural networks when improving sea state forecast by numerical models considering main spectral bulk parameters at a specific location, namely significant wave height, peak wave period and peak wave direction. The two alternatives are random forest and gradient boosting trees. To our knowledge, they have never been used for error prediction method. Therefore, their performances are compared with the performances of the usual choice in the literature: neural networks. We showed that the RMSE of the variables updated with gradient boosting trees and random forest are respectively 20 and 10% lower than the RMSE obtained with neural networks. A secondary objective is to show how to tune the hyperparameter values of machine learning algorithms with Bayesian Optimization. This step is essential when using machine learning algorithms and can improve the results significantly. Indeed, after a fine hyperparameter tuning with Bayesian optimization, gradient boosting trees yielded RMSE values in average 8% to 11% lower for the correction of significant wave height and peak wave period. Lastly, the potential benefits of such corrections in real life application are investigated by computing the extreme wave run-up (R_2%) at the study site (Biarritz, France) using the data corrected by the different algorithms. Here again, the corrections made by random forest and gradient boosting trees provide better results than the corrections made by neural networks.