Regional Ensemble ENSO Prediction Based on Graph Neural Networks with Self-Attention

ENSO exerts profound impacts on global climate change through ocean-atmosphere interactions and serves as a critical factor in global climate prediction. However, its prediction remains challenging due to the complex spatiotemporal interactions and evolution processes, as well as the varying degrees of correlation and teleconnection across different geographical regions. To address this issue, this study proposes an advanced ENSO forecasting framework based on regional predictions and model ensemble. The framework leverages a graph self-attention mechanism (GAT) to learn and capture the spatiotemporal dependency signals of ENSO, which are then incorporated as physical constraints into a spatiotemporal graph convolutional neural network (STGCN) for regional predictions. Furthermore, machine learning algorithms, including XGBoost and SVR are employed to integrate the predictions from different regions. Experimental results based on reanalysis data demonstrate the effectiveness and robustness of the proposed framework, achieving a correlation skill exceeding 0.8 within a 12-month lead prediction period, and significantly improving the computational efficiency by filtering key signals.