Multi-Year ENSO Forecasts Using Parallel Convolutional Neural Networks With Heterogeneous Architecture

The El Nino-Southern Oscillation (ENSO) is one of the main drivers of the interannual climate variability of Earth and can cause a wide range of climate anomalies, so multi year ENSO forecasts are a paramount scientific issue. However, most existing works rely on the conventional iterative mechanism and, thus, fail to provide reliable long-term predictions due to error accumulation. Although methods based on deep learning (DL) apply the parallel modeling scheme for different lead times instead of a single iteration model, they leverage the same DL model for prediction, which can not fully mine the variability of different lead times, resulting in a decrease of prediction accuracy. To solve this problem, we propose a novel parallel deep convolutional neural network (CNN) with a heterogeneous architecture. In this study, by adaptively selecting network architectures for different lead times, we realize variability modeling of different tasks (lead times) and thereby improve the reliability of long-term predictions. Furthermore, we propose a relationship between different prediction lead times and neural network architecture from a unique perspective, namely, the receptive field originally proposed in computer vision. According to the spatio-temporal correlated area and sampling scale of lead times, the size of the convolution kernel and the mesh size of sampling are adjusted as the lead time increases. The Coupled Model Intercomparison Project phase 5 (CMIP5) from 1861 to 2004 and the Simple Ocean Data Assimilation (SODA) from 1871 to 1973 were used for model training, and the GODAS from 1982 to 2017 were used for testing the forecast skill of the model. Experimental results demonstrate that the proposed method outperforms the other well-known methods, especially for long-term predictions.

Automated summary

The study introduces a novel parallel deep convolutional neural network model, which improves the reliability of long-term predictions by adaptively selecting network structures for different lead times, further enhancing prediction accuracy by adjusting the size of the convolution kernel and sampling grid to accommodate different lead times.