Abstract: ENSO (El Niño/Southern Oscillation) is an important air-sea coupling phenomenon that occurs in the equatorial central-eastern Pacific Ocean, and plays an important role in the global climate system. El Niño/La Niña occurs when the SST (Sea Surface Temperature) in the Eastern Tropical Pacific is warm/cool abnormally for more than 5 consecutive months. Therefore, it is of great scientific value to study and predict the SST dynamics in this region. In this paper, we use the Attention-LSTM model that introduces the Attention mechanism into the input layer of LSTM (Long Short-Term Memory) neural network to make one-year predictions of SST data obtained from multi-time and multiple Tropical Pacific observation buoy stations in El Niño and La Niña years based on exploring the effect of training sample length on the prediction results. In the SST prediction of the experimental site, the mean square error of the LSTM algorithm is about 0.5 ℃, while the mean square error of the Attention-LSTM algorithm is less than 0.31 ℃, which proves that the prediction accuracy of the Attention-LSTM algorithm is higher than that of the traditional LSTM model; At different stations in the Eastern Pacific Ocean in the year of ENSO, the Attention-LSTM algorithm also has a certain accuracy improvement on the Spring Predictability Barrier (SPB) phenomenon of SST.