Advancing Ocean Forecasting in the Russian Arctic: A Performance
Analysis of MariNet Model in Comparision to FourCastNet and PhyDNet
Abstract
Marine forecasts are essential for safe navigation, efficient offshore
operations, coastal management, and research, especially in areas with a
such harsh conditions as the Arctic Ocean. They require accurate
predictions of ocean currents, wind-driven waves, and other oceanic
parameters. However, physics-based numerical models, while precise, are
computationally demanding. Consequently, data-driven methods, which are
less resource-intensive, may offer a more efficient solution for sea
state forecasting. This paper presents an analysis and comparison of
three data-driven models: our newly developed convLSTM-based MariNet,
FourCastNet and the PhydNet, a physics-informed model for video
prediction. Using metrics such as RMSE, Bias and Correlation, we
demonstrate the areas where our model surpasses the performance of the
prominent prediction models. Our model achieves improved accuracy in
forecasting ocean dynamics compared to FourCastNet and PhyDNet. We also
find that our model requires significantly less training data, computing
power, and consequently provides less carbon emmisions. The results
suggest that data-driven models should be further explored as a
complement to physics-based models for operational marine forecasting.
They have the potential to enhance prediction accuracy and efficiency,
enabling more responsive and cost-effective forecasting systems.