Abstract: To address the limitations of existing deep learning methods in effectively capturing multi-scale evolution patterns of sea surface temperature (SST) and their insufficient capability to model nonlinear variations, this study first integrates Variational Mode Decomposition (VMD) with a 3D SwinLSTM network to develop a hybrid forecasting model. VMD is employed to decompose SST's spatiotemporal signals across multiple scales, revealing its complex multi-scale variability characteristics and enhancing the model's ability to capture multi-scale information, thereby improving forecast accuracy. Furthermore, by incorporating exogenous drivers such as heat flux and momentum flux, the model's prediction precision is further enhanced, with experimental validation conducted at the Western Pole of the Indian Ocean Dipole (WIO). Results demonstrate that compared to various baseline models, the proposed hybrid model integrating multi-scale variability features and exogenous drivers exhibits superior performance, achieving a root mean square error (RMSE) of 0.7 ℃—verifying the effectiveness of multi-scale periodicity information in SST forecasting. The introduction of exogenous drivers further reduces RMSE to 0.62 ℃, confirming their significant contribution to improved prediction accuracy..