Application of intelligent correction in ocean numerical forecasting based on spatiotemporal convolutional network (STCN)

Abstract Marine numerical forecasting is susceptible to complex spatiotemporal systematic errors, particularly during extreme weather events, such as typhoons, which majorly impede its efficacy in disaster prevention and mitigation. To address this challenge, the present study proposes an intelligent correction model based on a spatiotemporal convolutional network. A decoupled convolutional design that separates temporal and spatial dimensions is employed, enabling hierarchical parallel extraction and deep synergistic fusion of spatiotemporal error features from multivariable forecast fields. The experimental validation was conducted using multisource data acquired from the northern South China Sea from November 2023 to September 2025. The findings indicate that the proposed model has a substantial impact on enhancing the precision of critical variables, including sea surface temperature, sea surface elevation, and wave height. Specifically, the root mean square error ( RMSE ) of sea surface temperature decreased by an average of 23.8%, and the RMSE of wave height decreased by up to 24.0% under typhoon conditions. The spatiotemporal convolutional network (STCN) model has been demonstrated to enhance the spatial structure of typhoon-induced ocean responses and improve extreme value predictions in case studies of Typhoons Mitag and Ragasa in 2025. These enhancements contribute to an increased physical plausibility of the forecast fields. The present study corroborates the STCN model’s superiority in addressing systematic errors in numerical oceanic forecasts, particularly in the context of extreme weather events. It offers a reliable data-driven approach to enhance the overall accuracy of operational ocean forecasting.