Accurate sea surface temperature (SST) prediction in the East China Sea remains challenging because of its highly dynamic oceanic and atmospheric conditions, yet it is essential for regional fisheries management and marine hazard early warning. Here, we propose SwinTrans-ConvLSTM, a spatiotemporal deep-learning framework tailored for SST forecasting in the East China Sea. The model couples the global representation capability of the Swin Transformer with the local temporal-evolution modeling strength of ConvLSTM, incorporates air–sea temperature contrast and vector wind-field features, and is optimized using a curriculum-learning strategy constrained by a physics-informed gradient loss. Experiments show that SwinTrans-ConvLSTM achieves a mean absolute error of only 0.071 °C and a root mean square error of 0.156 °C on the test set, reducing prediction errors by approximately 30 % relative to state-of-the-art baselines. Crucially, hindcasts of the extreme 2022 marine heatwave event further demonstrate that the model can reproduce heatwave occurrence frequency and duration while substantially mitigating the systematic underestimation of extreme SST peaks inherent in purely data-driven models. These results highlight the critical role of thermodynamic-variable reconstruction and training-strategy optimization in improving the robustness of marine extreme-event prediction, and provide a promising technical pathway for high-resolution operational forecasting under complex ocean conditions.