Abstract: Sea surface temperature (SST) is one of the important contents of marine scientific research. Abnormal fluctuations of SST lead to marine and meteorological disasters. Accurate prediction is of great significance for the study of the marine environmental protection and the development of marine economy. For the seasonality and non-stationarity of the SST sequence, firstly, the data are preprocessed through Seasonal and Trend decomposition using loess (STL), and the decomposition obtains the seasonal component, trend component and residual component. Next, select the corresponding prediction method to build a combining model. Seasonal components are predicted using the Transformer with time embedding encoding module, which fully exploits the global information of the sequence, and solve the problem of capturing long-range dependency; trend components are predicted using linear regression; residual components are predicted using auto regression model. Select the single-point SST data of the South China Sea, and apply the combination prediction model based on STL for modeling to predict the SST value of 5 days. The experimental results show that the model of this paper can effectively capture the variation pattern of SST and improve the prediction accuracy in a single-point SST prediction task.