A Generalized Machine Learning Model to Predict Significant Wave Height from Wind Parameters

Abstract

Due to recent energy crises in the post-pandemic world, exploration of marine renewable energy sources is more crucial than ever. Significant wave height is a key parameter for wave energy extraction, it also has a wide range of applications, including ship navigation, oil and gas extraction, and the construction of coastal structures. Among the existing methods of measuring significant wave height, direct measurements using buoys are very expensive and limited in number, moreover these provide data with low time and spatial resolutions whereas numerical models are based on mathematical equations, assumptions and becomes complex when they are applied for generalization purpose. With a view to facilitate the utilization of wave energy and foster research activities by providing cheap dataset of wave properties with high spatial and time resolution, this work focuses on developing a generalized machine learning model that is able to predict significant wave height from wind parameters on a huge area around the coastlines of USA and Canada. Four machine learning models have been used in this work; 2 deep learning models (Artificial Neural Network (ANN) and Self Normalizing Neural Network (SNN)) and 2 gradient boosting tree-based models (XGBoost and LightGBM) and performance of these models have been evaluated on test data, distinct from the one used for training. The deep learning models have showed greater fitting capacity compared to tree based model on the training data, achieving the lowest Mean Squared Error(MSE) (0.047 for ANN, 0.063 for SNN, 0.226 for XGBoost, 0.108 for LightGBM) and highest R2 score (0.953 for ANN & 0.937 for SNN, 0.894 for XGBoost, 0.892 for LightGBM) whereas the gradient boosting models demonstrate better generalizing capacity compared to the deep learning models on both the known (data from these buoys are included in the training set) and unknown (data from these buoys are not included in the training set) buoys. Furthermore, impact of outlier detection and removal using Tukey’s Fence method on the performance of ANN & SNN has been evaluated and found to be insignificant.