MODELING OCEANIC VARIABLES WITH GRAPH-GUIDED NETWORKS FOR IRREGULARLY SAMPLED MULTIVARIATE TIME SERIES

Forecasts of ocean dynamic variables are essential to ensure safe operations at sea and in coastal regions. However, one difficulty with such forecasts is the need to handle multiple scales and repetitions in data, as well as noise caused by sensors malfunction. We describe a data-driven approach to predict oceanic variables under those circumstances; we take as a case study the prediction of water current velocity and sea surface height in an estuarine system in the southeastern coast of Brazil. We propose a generic method that can be applied to a variety of practical cases with little to no adaption, using a Graph Neural Network to model the system dynamics. We provide evidence that our method produces robust forecasts. It does so by employing forecast data from the state-of-the-art physics-based model "Santos Operational Forecasting System" (SOFS). The approach has lower computational costs and requires almost no domain-specific knowledge. We compare our model with SOFS and ARIMA-like forecast models in experiments.