A Long-term forecasting model for sea surface temperature based on wavelet decomposition and adaptive fusion

Sea surface temperature (SST) is a critical physical variable for characterizing ocean-atmosphere heat exchange and marine ecological processes. Its time series usually exhibit pronounced nonstationarity and multi-scale coupling characteristics. Existing methods face difficulties in long-term forecasting tasks in simultaneously balancing multi-scale feature modeling capability and computational efficiency, which limits their applicability in practical forecasting scenarios. To address these issues, this study proposes a multi-scale modeling framework for long-term SST forecasting, termed WFNet. This method enhances the modeling of relationships among SSTs in different regions by introducing an attention mechanism, and employs multi-level wavelet decomposition to decouple the original time series into components with different temporal-scale characteristics. On this basis, a multi-branch MLP architecture is adopted to model each scale component independently, and effective integration of multi-scale information is achieved through noise suppression and adaptive weighted fusion strategies, thereby improving prediction stability while reducing model complexity. To verify the effectiveness of the proposed model, forecasting experiments with lead times ranging from 30 to 360 days are conducted on daily mean SST datasets from four typical fishing grounds in the northern Yellow Sea of China. Systematic comparisons are performed with mainstream models including LSTM, DLinear, TCN, iTransformer, and xPatch. Experimental results demonstrate that WFNet achieves superior performance in terms of RMSE and MAE, while exhibiting significantly lower computational cost (GFLOPs) than self-attention-based models. This indicates that WFNet achieves a favorable balance between long-term forecasting accuracy and computational efficiency, suggesting its potential applicability in operational ocean forecasting and marine disaster early-warning support systems. • A lightweight wavelet-based neural network (WFNet) is proposed for long-term SST forecasting in the Northern Yellow Sea. • Multilevel wavelet decomposition explicitly separates low-frequency background variability and high-frequency fluctuations. • A magnitude-aware adaptive fusion strategy enhances the stability of multi-scale SST predictions. • WFNet achieves lower prediction errors than several baseline models across 30–360 day forecasting horizons, with up to 28.9% RMSE reduction. • The low computational cost (≈1.15 GFLOPs) makes WFNet suitable for efficient regional SST forecasting applications.