Abstract
The South China Sea (SCS) holds global geopolitical, economic, and environmental importance, making the accurate forecasting of its marine variables essential for effective ocean management and hazard mitigation. While traditional numerical models are fundamental to ocean forecasting, they are often constrained by high computational complexity and low efficiency. Recently, deep learning-based data-driven GOFS, such as XiHe, have shown great promise by achieving forecasting accuracy competitive with traditional numerical models at a fraction of the computational cost. However, although deep learning-based GOFS can efficiently capture large-scale ocean variability, their direct application to dynamically complex regional seas remains challenging because regional coastal gradients, mesoscale structures, and local error patterns are often insufficiently resolved. To address this issue, we propose a Swin-Transformer Corrector (STC), a dedicated regional post-processing model for multivariate correction of frozen deep learning-based global ocean forecasts over the SCS region. Rather than replacing the original forecasting system, STC is designed as a lightweight plug-in corrector that preserves the largescale predictive prior of the global model while improving regional forecast fidelity. Specifically, it employs a hierarchical Swin Transformer backbone to capture multiscale spatial error structures, explicitly uses high-resolution features to retain coastal and mesoscale information, and applies residual correction to refine the baseline forecasts efficiently. Experiments show that STC significantly improves prediction accuracy, achieving an average reduction of 20.35% in RMSE, while also demonstrating strong adaptability under extreme conditions such as Tropical Cyclone Haikui.