MJOFormer: An adaptive land-ocean spatio-temporal transformer for Madden–Julian Oscillation forecasting

Abstract

The Madden–Julian Oscillation (MJO) represents the predominant driver of sub-seasonal variability within tropical regions. In deep learning of weather forecasting, achieving reliable accuracy in MJO prediction remains challenging, making sub-seasonal forecasts generally probabilistic. In this paper, we reveal the challenges that impede MJO forecasts, including distributional drift when the circulation passes from the Indian Ocean across Australia, hindered by obstacles like lands or shoals. In addition, we find it non-trivial to extract the sophisticated spatio-temporal relationships in climate data. To address these issues, we propose MJOFormer, an adaptive land-ocean spatio-temporal transformer, with (1) a land-ocean sampler to address distributional drift by adaptively partitioning across terrain; (2) a dynamic attention mechanism to compensate for the absence of spatial features by adaptively tackling the spatio-temporal correlation; (3) a cuboid method to improve efficiency by parallel training. Comprehensive experiments exhibit that MJOFormer possesses competitive, outperforming existing methods with better accuracy, stability, and efficiency.