Spatial-temporal transformer network for multi-year ENSO prediction

The El Niño-Southern Oscillation (ENSO) is a quasi-periodic climate type that occurs near the equatorial Pacific Ocean. Extreme periods of this can cause terrible weather and anomalies on global scale. Therefore, it critical to accurately, quickly, effectively predict occurrence ENSO events. Most existing research methods rely powerful data-fitting capability deep learning which does not fully consider spatio-temporal evolution its character, resulting in neural networks with complex structures but poor prediction. Moreover, due large magnitude ocean variability over long intervals, they also ignored nearby prediction results when predicting Niño 3.4 index for next month, led errors. To solve these problem, we propose transformer network model inherent characteristics sea surface temperature anomaly map heat content along changes space time by designing an effective attention mechanism, innovatively incorporate temporal into feature procedure influence seasonal variation phenomenon. More importantly, better conduct long-term prediction, recurrent strategy using previous as prior knowledge enhance reliability Extensive experimental show our provide 18-month valid validates effectiveness method.