PCFNet: A Sea Surface Temperature Prediction Network Fusing Physical Constraints With Frequency Domain Feature Extraction

Sea surface temperature (SST) predictions are critical for marine climate research and ecosystem management. To address the deficiencies of the existing data-driven models in regard to physical consistency constraints and multiscale feature modeling, this paper proposes a deep learning model that fuses physical constraints and frequency domain features for the SST prediction. This model employs the fast fourier transform (FFT) to process spatio-temporal features. At the same time, it incorporates a constraint integration module (CIM), which embeds the physical conservation properties of the advection-diffusion equation (ADE) into the prediction process, thereby enhancing the modeling capability for the dynamic evolution patterns of SST. The experimental results on both daily and weekly scales demonstrate that the proposed PCFNet achieves consistently higher prediction accuracy compared with existing SST forecasting models. In particular, it provides more stable short-term predictions and effectively mitigates multi-step error accumulation. These results further confirm the effectiveness of integrating frequency-domain structure modeling and cross-scale interaction mechanisms. Overall, this study offers a promising and interpretable framework for SST prediction by combining physical priors with data-driven learning.