Abstract: As one of the important variables in the ocean, SST (sea surface temperature) and its changes significantly impact global climate and marine ecology; therefore, it is necessary to forecast SST. Deep learning is highly efficient for data processing; however, it is rarely used in the short-term forecasting of the equatorial Pacific SST pattern. Based on the LSTM (long short-term memory) network, this paper constructs a daily forecast model of SST for the tropical Pacific Ocean (10°S–10°N, 120°E–80°W) for the next 10 days. The proposed model forecasts SST using observations from 1982 to 2010 as a training set and data from 2011 to 2020 as a test set. The results show that the forecast RMSE (root mean square error) for the eastern equatorial Pacific region is larger than those for the central and western areas. The RMSE for the east basin is approximately 0.6°C on the first day of the forecast, while those for the western and central regions are less than 0.3°C. The forecast skill of the model for different phases of the El Niño–Southern Oscillation is examined. RMSE is the largest in the La Niña period, followed by normal years, and the smallest in the El Niño period. RMSE in the La Niña period is more than 20% in some regions than in the El Niño period. The forecast error is positive in the east and negative in the west. The number of predictable days is more than 10 days. Specifically, the number of predictable days near the equatorial cold tongue is 4–7 days, while that in the western equatorial regions is 3 days. The model shows lower forecast skills in the eastern equatorial Pacific regions than in the western regions. In terms of skills for various months, the skill level is lower in October and November than in other months. In general, the SST forecast model based on LSTM can satisfactorily capture the evolutionary characteristics of the SST time series, and the forecast performance is high for the pan-tropical Pacific regions. Furthermore, the running time of the proposed data-driven forecast model for predicting the daily average SSTs for the next 10 days is very fast and more efficient than that of traditional dynamical models.