ENSO is one of the most influential coupled air-sea phenomena in earth system.

The accurate prediction for ENSO is of significant importance and practical value. In this study, we explore a brand novel method for ENSO prediction based on deep learning technology. The network is deep enough with 179 layers in total and can be fed with multiple precursors such as spatio-temporal physical fields and index series related to ENSO at the same time. Based on CNN-ResNet algorithm, the model is promising to catch spatial dependence of physical fields and dig the relationship between ENSO and various parameters well.

The DL models outperform the general performance of conventional prediction models, with higher accuracy, better predictability at longer lead times. At the lead time of 6 months, correlation coefficient between predictions and observations can reach 0.8,while RMSE be no more than 0.62°C. At the lead time of 12 months, correlation coefficient can reach 0.72, while RMSE can be no more than 0.7°C. Such deep learning model also has potential for detecting effective precursors. Ocean temperature over the upper 300 meter, the warm water volume, and the surface ocean currents over tropical Pacific are proved to be key factors in ENSO prediction, which is consistent with previous studies for ENSO physics. Predictions from DL model shows that ENSO event tend to develop into weak El Niño at the end of the year 2018, while to be neutral in 2019.

We are still working on improving the training speed by simplfying the model and overcoming convergence problems.