Predicting Future Wave Heights by Using Long Short-Term Memory

Abstract

In this paper, several deep learning models are trained using Long Short-Term Memory (LSTM), which is a special type of a recurrent neural network that can handle time series data due to its memory. These models are also compared to a Temporal Convolutional Network (TCN) model, which is comparable to LSTM models in terms of prediction. These models use the wind speed, wind direction, significant wave height, and mean wave direction as input features to forecast the significant wave height. Furthermore, the wind and mean wave direction are subtracted from the first value of the time series to forecast wave heights for multiple locations. To build a model with a prediction reliability, a lower quantile of 2.5 percent and an upper quantile of 97.5 percent are first predicted for a range that is too high. Then, the highest class with a prediction probability of greater than 50 percent is used to improve the wave height forecast. It has been observed that a class of a maximum wave height over a longer period of time can lead to better results than the wave height for a single time point. However, all models cannot sufficiently forecast the wave height over several days, which is needed to determine weather windows, i.e. periods of time when a maritime infrastructure is accessible, e.g. for maintenance. Weather data from the North Sea are used for the training, validation, and test data.