Real-time in situ prediction of ocean currents

Abstract

The prediction of ocean currents is essential for the path planning and control of Autonomous Underwater Vehicles. Regional physics-based forecast models provide valid predictions but are too computationally expensive for real-time prediction necessary for AUV navigation. While vehicle sensors can measure the spatial evolution of currents, temporal prediction remains an open problem as existing data-driven models with real-time capabilities have only been shown to work at locations where data have been used to develop the model. We propose in this paper two predictive tools using deep learning techniques, a Long Short-Term Memory (LSTM) Recurrent Neural Network and a Transformer, to perform real-time in-situ prediction of ocean currents at any location. A data set from the National Oceanic and Atmospheric Administration is split in two distinct sets to train and test the models. We show that the LSTM and the Transformer have an averaged Normalized Root Mean Squared Error respectively of 0.10 and 0.11 over all test sites with a standard deviation respectively of 0.024 and 0.031. Comparisons with Harmonic Method predictions at various locations in the territorial sea of the United States show that both models provide state-of-the-art accuracy without having been trained with data from these sites.