Marine meteorological prediction is of great significance in ocean engineering, climate change research, ocean transportation and disaster warning. However, existing models cannot fully capture the complex nonlinear relationships of marine meteorological systems, resulting in biased forecast results. To this end, this paper proposes a correction method for marine meteorological prediction data based on deep learning (DL-CMMP), which aims to reduce the deviation between the predicted value and the true value, thereby correcting the accuracy of ocean meteorological predictions. Firstly, a marine meteorological accuracy optimization model based on the Transformer architecture was designed. By optimizing the data structure and attention mechanism, it can effectively capture the dependencies between meteorological variables and reduce the computational complexity. Secondly, a spatiotemporal loss function combining the interaction of physical factors is proposed. By introducing temporal, Spatial and physical constraints, it ensures that the model can better reflect the physical interactions between marine meteorological variables during the optimization process, thereby enhancing the generalization ability and correction effect of the model. Experimental results show that the proposed method performs better than other methods in the marine meteorological correction task, and provides a practical solution for the integration of physical models and data-driven models in the field of marine meteorology.