A Machine Learning Approach to Water Velocity Estimation for Better Navigation of Marine Gliders

Marine gliders are autonomous underwater vehicles used for data collection by ocean and climate scientists. To conserve energy on months-long trips these gliders use a propulsion system driven mostly by gravity. This variable buoyancy propulsion system, however, is particularly susceptible to error introduced by strong or chaotic ocean currents. For this reason, most navigation systems rely on prediction of local water velocity to plan their paths. Current physics-based simulation models are limited by on-board computation resources. Consequently, we introduce several variations of recurrent neural network based prediction algorithms and compare their performance to a commonly used statistical forecasting method.