ENSONet: a data-driven ENSO forecasting model with concise spatial location learning parameters and temporal embedding

The El Niño-Southern Oscillation (ENSO) is a highly notable climate phenomenon with significant implications for global weather patterns and climate change. Accurately predicting ENSO holds substantial scientific and economic value. However, due to the intricate relationship between the evolution of the oceans and the atmosphere across spatial and temporal scales, currently, the most advanced physically based dynamical models struggle to deliver skillful predictions beyond a 1-year. Deep learning models often prioritize complex module stacking, neglecting the incorporation of crucial spatial and temporal information and providing inaccurate predictions over long distances. To overcome these challenges, ENSONet is proposed in this study. It identifies the Niño high correlation region and temporal relationships by designing concise spatial location learning parameters and temporal embedding. The progressive prediction architecture employs multiple learning to enhance long-term prediction accuracy and effective distance. Additionally, novel prediction-relevant regions are discovered from ocean features using spatial and temporal attention modules, and intricate prediction patterns are learned by finely modeling spatio-temporal relationships. Extensive experiments on real-world datasets demonstrate that ENSONet identifies regions directly associated with the Niño index and uncovers new regions of interest through continuous learning. By successfully predicting changes in ENSO from 1984 to 2023, it showcases its proficiency in learning complex predictive patterns. In conclusion, ENSONet not only expands the prediction horizon to the 18th month but also demonstrates a remarkable enhancement in prediction accuracy, with an average improvement of 28.99%, thus achieving state-of-the-art performance.