The Kuroshio, a powerful western boundary current in the North Pacific, exhibits multi-scale variability that profoundly affects regional weather, climate, marine ecosystems, and fisheries, rendering its accurate prediction indispensable. However, this variability is driven by complex multi-scale physical processes, necessitating high-resolution numerical models that are computationally expensive and often constrained by limited timeliness. In recent years, the emergence of data-driven models has opened new avenues for ocean forecasting, and the global ocean intelligent prediction systems are now approaching or even surpassing traditional numerical models across various metrics. Despite these advances, their performance in the Kuroshio region remains limited. To address this challenge, this study develops an eddy-resolving (1/12°) Kuroshio Intelligent Prediction System (KIPS) based on the Swin Transformer architecture. Specifically designed to capture Kuroshio dynamics, KIPS uses an autoregressive strategy to generate daily forecasts of three-dimensional temperature, salinity, current, and sea surface height, with a lead time of up to 10 days. KIPS achieves higher accuracy compared to existing numerical and AI-based prediction systems, while significantly reducing computational costs. In operational forecasts, KIPS successfully captures several recent eddy shedding and merging events in the southern Kuroshio region of Japan, demonstrating agreement with near-real-time satellite observations. These results underscore the value of integrating prior physical knowledge into region-specific forecast systems to improve fine-scale ocean prediction.

Ocean forecasting,Kuroshio