El Niño/Southern Oscillation (ENSO) is a complex coupled ocean-atmosphere event, usually manifested as an abnormal increase or decrease in Sea Surface Temperature (SST) in the equatorial Pacific Ocean. ENSO is one of the critical factors causing global climate extremes and ecosystem turbulence. Therefore, long-term robust ENSO predictions are essential for people’s productive life and social development. The statistical and dynamic models still face challenges in making long-lead-time ENSO forecasts, and the accuracy can be improved. Here, we have created three innovative efforts for ENSO forecasting: (1) To compensate for the lack of Remote Sensing (RSS) dataset, we apply transfer learning with the large SST dataset simulated from the LASG/IAP Climate Ocean Model Version 3 (LICOM3). (2) We propose a novel encoder-decoder structure with multiple Spatiotemporal prediction units (ST-Pred) named ED-PredRNN. The ST-Pred changes the original transfer direction by adding spatiotemporal memory cells, allowing the model to extract more information. (3) To reduce the memory bandwidth load, we use FP32 and FP16 mixed-precision computing methods to improve model training and inference speed.The result shows that during 2011–2020 verification period, the correlation skills of the Niño 3.4 index predicted by ED-PredRNN are over 0.77/0.64/0.53 within the lead time of 6/12/18 months respectively, which is superior to existing statistical models. Also, the precise predictions of the El Niño event (such as 2015/16) and the La Niña event (such as 2010/11), as well as a reasonable forecast for neutral ENSO conditions (2014), demonstrate the effectiveness and superiority of our proposed model.