Abstract El Niño and the Southern Oscillation (ENSO) is the strongest inter‐annual signal in the global climate system with worldwide climatic, ecological, and societal impacts. Over the past decades, the research on ENSO prediction and predictability has attracted broad attention. Typical prediction efforts based on physically coupled models (e.g., SINTEX‐F, CanCM4) demonstrate skill at short lead times (approximately 6–12 months) but tend to lose predictability rapidly over longer horizons. Benefiting from their ability to capture nonlinear dependencies and improve long‐term accuracy, deep learning methods such as Convolutional Neural Network (CNN) and Long Short‐Term Memory (LSTM) networks have been widely applied in the prediction of ENSO‐related indices, such as the Niño 3.4 index. In this study, we propose a newly designed neural network, named ACTNet, by incorporating a self‐attention mechanism into a CNN + LSTM architecture. ACTNet is designed to process the past 12 months of global sea surface temperature (SST), heat content, zonal wind (UA), and meridional wind (VA) as inputs, where CNN layers extract spatial patterns, LSTM layers capture temporal dependencies, and a self‐attention mechanism highlights critical spatiotemporal relationships for accurate ENSO prediction. It can predict the Niño 3.4 index at a monthly resolution up to 24 months in advance reasonably well, achieving correlation coefficients exceeding 0.5. Compared to conventional CNN and CNN + LSTM models, ACTNet demonstrates improved spatiotemporal feature extraction and long‐lead prediction skill. Another, independent aspect is ENSO‐type prediction based on historical observed SST anomalies. Since ENSO events manifest in different types—such as Eastern Pacific and Central Pacific El Niño, as well as their La Niña counterparts—distinguishing these types is crucial for understanding regional climate impacts. To this end, we further employed an LSTM model to classify events into six defined ENSO types based on Niño 3 and Niño 4 indices, achieving a classification accuracy of 70.5% at a 12‐month lead time.
Yu et al. (Thu,) studied this question.