Applying Convolutional LSTM Network to Predict El Niño Events: Transfer Learning from The Data of Dynamical Model and Observation

Abstract

Neural network as a statistical method is widely used for weather forecasting. But for the prediction of El Niño grid data, the data record is short. In this paper, we use deep learning to handle spatiotemporal information and transfer learning to transfer knowledge from dynamical model (Zebiak–Cane model) data to the prediction of realistic El Niño. A ConvLSTM (Convolutional Long Short-Term Memory Network) architecture is constructed to predict the grid data of sea surface temperature and thermocline depth at lead times from 3 to 12 months. Cross-validation is used to evaluate the predictions. The entire data record from 1980 to 2018 is divided into 10 groups and used for training and validation. The experiment results show that transfer learning have a positive impact on the El Niño prediction, especially for the strong Eastern-Pacific type. Compared with the predictions of the Zebiak-Cane model, it can be inferred that the role of the model data in transfer learning is greater than the observation data.