Investigation of recurrent neural network architectures for prediction of vessel trajectory

Abstract

Modern deep learning algorithms are able to handle large amounts of data and therefore are particularly important in automating vessel movement prediction in intensive shipping. This could be one of the support tools for monitoring, managing the increasing maritime traffic and its participants. Applying deep learning algorithm, a recurrent networks is created that is able to predict the further vessel movement. The developed architectural model is based on sequences when data change over time, therefore the article investigates the most optimal recurrent network structure and network hyper-parameters, which aim to obtain the most accurate prediction results. Different recurrent network architectures were used to compare the results those are: fully-connected (simple) recurrent neural network, basic (vanilla), bidirectional, stacked Long Short-Term Memory network, autoencoder, and gated recurrent unit. The accuracy of the predictions for each architecture is monitored by varying the number of cells size in the hidden layer. The research was performed on a specific sample of data from the Netherlands (North Sea) coastal region and the proposed algorithm can be applied as one of the ways to improve maritime safety. The research showed that the most accurate prediction of the vessel trajectory prediction is achieved with the bidirectional Long Short-Term Memory network architecture in which the variance is less shifting even with the smallest cell selection, and autoenoder network architecture which depends on the choice of the appropriate cell size, because distribution range increasing in 100 and 150 cells.