A Deep Learning‐Based Long‐Term ENSO Forecasting Model: 3D‐STransformer

• Published in: Journal of geophysical research. Machine learning and computation Vol. 2; no. 3
• Main Authors: Lian, Jie, Wu, Xinjiao, Huang, Sirong, Chang, Zhanyuan
• Format: Journal Article
• Language: English
• Published: 01.09.2025
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000412

The El Niño‐Southern Oscillation (ENSO) significantly impacts global climate variability, causing extreme events like droughts, floods, and heatwaves. Accurate prediction of ENSO is critical for managing agriculture, water resources, disaster prevention, and economic planning. Despite advances in understanding ENSO's mechanisms and developing prediction models, forecasting its timing, intensity, and duration precisely continues to be a significant obstacle because of the nonlinear and complex characteristics of the phenomenon. In this study, we introduce a 3D‐STransformer model, with the aim of improving the accuracy and reliability of long‐term prediction by integrating multiple local and remote factors affecting ENSO dynamics, such as wind stress and upper‐ocean temperature at different depths. In addition, the model employs a multi‐head spatiotemporal attention mechanism to capture long‐range dependencies and complex interactions across time and space. We pre‐train the proposed model on the CMIP6 data set, then perform the transfer learning on the SODA data set, and finally validate the model on the GODA data set. The model employs an end‐to‐end, multistep rolling prediction strategy. It takes 12 months of input data and produces forecasts for the following 20 months. The experiment demonstrates that the model has superior performance, maintaining a high correlation in the Niño3.4 SST anomaly predictions up to 20 months ahead. The El Niño‐Southern Oscillation (ENSO) is a major climate phenomenon that significantly influences global weather, often triggering extreme events such as droughts and floods. Accurate long‐term prediction of ENSO is essential for effective preparedness. This study introduces a novel deep learning model, 3D‐STransformer, where “S” represents Spatial and “T” signifies both Temporal and the Transformer architecture's core capability. Designed to capture spatiotemporal dependencies via a multi‐head spatiotemporal attention mechanism, the model aims to enhance long‐term ENSO forecasts. It accepts inputs including wind stress fields and temperature anomalies at various depths of the upper ocean. The model's multi‐head spatiotemporal attention mechanism extracts critical details across both time and space, while a patch partitioning technique directs focus toward key ocean regions. These features enable 3D‐STransformer to maintain high accuracy even when predicting ENSO events up to 20 months in advance. Overall, experimental results indicate that 3D‐STransformer outperforms traditional forecasting methods, offering a more reliable tool for scientists and decision‐makers. A 3D‐STransformer model based on the multi‐head spatio‐temporal attention mechanism was proposed for El Niño‐Southern Oscillation (ENSO) prediction Both local and remote factors that influence the prediction of ENSO of target areas are considered in the method Improved the accuracy of ENSO predictions over a 20‐month forecast period