Ship Heave Motion Prediction Using LSTM Neural Networks Without Wave Excitation Information

Abstract

In off-shore multi-body operations, it is important to accurately determine the movements of the vessels to allow for safe operations. This is especially evident in multi-body operations that involve helicopters, due to the relative motion between the landing deck and the helicopter. To be able to augment human judgement in helicopter/off-shore vessel multi-body operations, it is crucial to have accurate predictions of the movements of the vessel. One of the key movements of off-shore vessels in these operations is the heave motion, which represents the up and down movements of the vessel. Modelling these movements mathematically is a difficult feat due to the strong non-linearity of these movements, and the complicated hydrodynamic forces and stochastic sea disturbances that are at the root of these movements. Therefore, Neural Networks bear a lot of opportunity in this regard, because of their strong ability of handling non-linearity. Especially, Long Short-Term Memory (LSTM) models are useful in this regard owing to their strong ability of handling sequences and their ability of learning both long- and short-term dependencies. However, most existing research uses wave excitation information to predict the heave motion even though ships are not often equipped with expensive wave detection systems. Therefore, this paper researches whether LSTM models are able to accurately predict heave motions 10 and 20 seconds ahead without the usage of wave excitation information. This paper shows that LSTM models are able to achieve accurate predictions without the usage of wave excitation information. Therefore, it is shown that LSTM models have the ability to augment human judgement in helicopter/off-shore vessel operations where this wave excitation information is not available. Especially in terms of 10 second ahead predictions LSTM is able to achieve promising prediction performance. However, the prediction quality of the 20 second ahead predictions is less satisfying. Finally, most existing papers used simple one-layer LSTM models, whereas this paper shows that using more complicated models lead to increased prediction performance. i