Predicting ocean-wave conditions using buoy data supplied to a hybrid RNN-LSTM neural network and machine learning models
Gülüstan Doğan, Meghan Ford, Scott James
Abstract
The ability to accurately predict ocean-wave conditions is paramount for many maritime activities. A framework comprising a bi-directional recurrent neural network (RNN) with long-short term memory (LSTM) cells (RNN-LSTM) was developed for timely and accurate predictions of ocean-wave conditions. For comparison, we employed a suite of non-deep learning data-driven techniques to predict significant wave height. To obtain precise predictions, models need accurate and consistent data. The dataset was an aggregation of condition data taken from wave-observing floats (buoys) that recorded readings in increments of 20 and 30 minutes. The buoy measured wave height, wave period, sea surface temperature, and other attributes off the coast of Wilmington, NC. These features were used in various approaches to predict future conditions. Using an RNN-LSTM deep learning methodology, wave conditions were predicted with an average validation accuracy of 81–83%. The Gradient Boosting machine learning model predicted significant wave height with an RMSE of 0.237. This approach has global application as it can be deployed using any sufficient data set collected from wave-monitoring buoys.