Explainable physics-guided attention network for long-lead ENSO forecasts

The El Ni & ntilde;o-Southern Oscillation (ENSO) is the primary interannual variations of the climate system, significantly impacts global climate patterns, ecosystems, and economies. Most cutting- edge ENSO prediction methods rely on traditional numerical models and novel data driven technologies. The numerical ways are based on dynamic equations and contribute to the physical representation of ENSO. However, the numerical model is relatively complex, leading to resource consumption, and it fails to address the inherent uncertainty like spring predictability barrier (SPB) and signal-to-noise ratio problem for long-lead forecasts particularly in long-term forecasts exceeding one year. Data-driven methods can effectively alleviate the SPB and improve the effective hindcasting time. However, they lack guidance from physical mechanisms, which results in a lack of physical interpretability in their outcomes. This can even lead to physically inconsistent results. In this study, we introduce an explainable physics-guided intelligent spatio-temporal forecasting model for ENSO (PGtransNet_ENSO). The model incorporates key characteristics and factors of ENSO events, including internal variability, external forcing, Bjerknes positive feedback mechanism, delayed attention mechanism to account for temporal lag effects, and El Ni & ntilde;o/La Ni & ntilde;a event types and intensities encoding. PGtransNet_ENSO maintains high accuracy even with limited data availability and enhances the model's convergence speed. Extensive experimental confirm its capability to deliver dependable ENSO predictions up to 12 months in advance. Moreover, the model outputs demonstrate robust physical consistency with established dynamical principles, thereby enhancing the interpretability of its underlying mechanisms.