Abstract: Based on the sea surface temperature (SST) from satellite remote sensing and prediction of global 0.1° wave-tide-circulation coupled ocean model developed by First Institute of Oceanography, MNR, China (FIO-COM), linear regression model and Long Short-term Memory (LSTM) neural network were used to correct the errors in the SST prediction. A series of comparative tests were carried out with the data from 2016 to 2021. The linear regression model, which was based on assumption of local linearity, predicted the error at the next moment. The LSTM was built and trained with historical deviation data in the 56 months from 2016 to 2020, and was tested with the data in 2021. The results show that both the linear regression model and the LSTM neural network can improve the prediction, but the latter performs better than the former, and the SST error is reduced by about 70%. Compared with the linear regression model, the probability density distribution of the deviation at each point after LSTM correction is concentrated around 0. The LSTM method is statistically better than the linear fitting and its results are more stable, and can be further applied to multi-factor deviation correction of ocean numerical prediction.