Prediction of Global Mesoscale Eddy Trajectories Using a Temporal Convolutional Network

Abstract Mesoscale eddies are ubiquitous throughout the global ocean, and their propagation and evolution exert important influences on oceanic energy transport and salt exchange. Accurate prediction of short-term mesoscale eddy propagation is therefore crucial for advancing understanding of mesoscale dynamics and improving marine forecasting skill. To address the limited ability of traditional statistical approaches to represent complex nonlinear propagation, a Temporal Convolutional Network (TCN) is developed for future eddy position prediction. The results show that the TCN, when constrained by multi-source background information including surface winds, sea surface currents and bathymetry, can effectively capture the nonlinear short-term propagation behavior of mesoscale eddies. For 7-day lead time prediction on a global scale, the model yields a median distance error of 19.85 km, outperforming widely used baselines including LSTM, MLP and maximum likelihood estimation. However, discrepancies remain under low-latitude, strong background-flow, and high-speed propagation conditions, suggesting that stronger physical constraints should be incorporated in future work to improve generalization and physical consistency.