Preliminary study on wave height prediction with convolution neural network

Abstract

Nowadays, the Convolution Neural Network(CNN) has been widely utilized to classify the image categories in many fields. We initially conducted some numerical trainings to figure out the applicability of machine learning in wave height prediction. A series of wave fields were produced with Airy wave super-positions. We constructed an efficient CNN structure which is able to tune various training parameters of machine learning algorithm. Also image filtering techniques were adopted and evaluated in wave field prediction. A long-crested wave case and some short-crested wave ones were modeled and trained in wave height classification. In the former case, the overall accuracy was remarkably high, but the prediction with short-crested waves were not reliable. While the middle range of wave height categories were not predictable, the highest and lowest wave height categories showed the good prediction results. Some variations on training parameters and structures were lastly performed in order to increase the prediction accuracy in short-crested wave trainings while their effectiveness was not significant. ？ 2021 by the International Society of Offshore and Polar Engineers (ISOPE).