Capturing Long-Range Dependencies for Improved MISO Prediction via Deep Learning 

The subseasonal-to-seasonal (S2S) prediction gap is a major challenge in operational forecasting, especially for the Indian Summer Monsoon. Prediction skill of the dynamical models for its dominant mode of variability, the Monsoon Intraseasonal Oscillation (MISO), drops sharply beyond one week. This 30–60-day northward-propagating mode governs active and break spells, with major implications for agriculture, water resources, and disaster preparedness across South Asia. In this study, we developed a deep learning forecasting framework built on a Transformer architecture to capture the long-range dependencies inherent to intraseasonal variability. We used 25 years of high-resolution (0.25° × 0.25°) daily TRMM/GPM precipitation data to derive MISO indices (MISO1 and MISO2) via extended empirical orthogonal function (EEOF) analysis for the boreal summer months (June–September). These indices formed the basis for training and evaluating the Transformer model. When evaluated for the 2018–2022 period, the Transformer substantially outperformed traditional numerical weather prediction models, accurately forecasting the phase and amplitude of MISO with lead times of up to 18 days. It also produced better phase alignment and reduced phase-lag errors compared to NWP systems at extended leads.  The approach was further extended to predict NLSA-based MISO indices. In addition, a Vision Transformer (ViT) was used to make preliminary forecasts of spatial rainfall patterns associated with MISO propagation. These results highlight the potential of advanced deep learning architectures to enhance S2S prediction of monsoon intraseasonal variability, supporting improved early warning systems and decision-making in monsoon-affected regions.