Spatiotemporal Model with Attention Mechanism for ENSO Predictions

Climate disasters such as floods and droughts often cause significant losses to human life, national economy, and public safety. The El Nino Southern Oscillation (ENSO) is one of the most important interannual climate signals in tropical regions, and has a global impact on atmospheric circulation and precipitation. Accurate ENSO predictions can help prevent related climate disasters. Recently, convolutional neural networks (CNNs) have shown the best techniques for ENSO prediction. However, it is difficult for convolutional kernels to capture the long-distance features of ENSO due to the locality of convolution itself. We regard ENSO prediction as a spatiotemporal series prediction problem, and propose an ENSO non-stationary spatiotemporal prediction deep learning model based on a new attention mechanism and a recurrent neural network, called ENSOMIM. The model expands the Receptive field of the network to achieve the learning space characteristics of local and global interaction, and uses high-order nonlinear spatiotemporal neural networks to encode long-term time series features. In order to adequate training the model, we also add historical simulation data to the training set and conduct transfer learning. The experimental results indicate that ENSOMIM is more suitable for large-scale and long-term prediction. During the testing period from 2015 to 2023, ENSOMIM's Nino3.4 index's all-season correlation skill improved by 11% compared to classical CNNs, and the root mean square error decreased by 29%. It can provide effective predictions for a lead time of up to 20 months. Therefore, ENSOMIM can serve as a powerful tool for predicting ENSO events.