Physics-informed LSTMNet for short-term marine buoy trajectory prediction under dynamic ocean environmental conditions

Abstract

Introduction Marine buoys are essential aids to navigation, yet their positions can vary continuously under dynamic environmental forcing. It is therefore important for navigational safety, maintenance planning, and reducing search-and-recovery costs to achieve accurate short-term prediction of buoy displacement. Physics-based models often depend on accurately specified hydrodynamic and meteorological parameters, whereas many existing data-driven approaches employ shallow, unidirectional recurrent architectures that may not fully capture the temporal context embedded in historical observations. Methods To address these limitations, this study proposes LSTMNet, a two-layer bidirectional long short-term memory network for next-step buoy trajectory prediction. The model learns temporal dependencies from multivariate buoy observations and predicts future buoy positions under dynamic marine conditions. Results The model was evaluated using hourly observations collected from a marine buoy during May 2026. On the held-out test set, LSTMNet achieved a mean geodesic distance error of 18.297 m, representing a 9.747% reduction relative to a single-layer unidirectional LSTM baseline. These results indicate that combining stacked recurrent learning with bidirectional temporal encoding improves the extraction of contextual dependencies from recent observations and enhances short-term buoy trajectory prediction. Discussion The proposed framework provides a practical data-driven approach for forecasting buoy displacement and may support safer and more efficient management of marine navigation infrastructure.