Sea surface temperature (SST) is an important marine hydrologic parameter, and its accurate prediction is critical in marine-related fields. Deep learning has been widely and increasingly used for SST prediction in recent years because of its strong analytical ability. However, the volatility and randomness of SST time series still constitute a challenge for accurate prediction. In this study, variational mode decomposition (VMD) is first introduced as a denoising module to reduce the influence of SST sequence noise on the prediction results. Then, to solve the lag phenomenon of depth models in SST prediction, the method of transfer learning is adopted to combine the concepts of long-short term memory (LSTM) and broad learning system (BLS). LSTM is used as the feature mapping node of BLS to improve the prediction accuracy. As a result, an SST prediction model based on VMD, LSTM, and BLS is proposed. The SST of the East China Sea is selected as an example for verification. By comparing with benchmark models, support vector regression (SVR), LSTM, gate recurrent unit (GRU), and existing deep models, it is shown that the proposed model is relatively stable and efficient in SST prediction, which provides a new idea for the development of SST prediction.