A Skillful Prediction of Monsoon Intraseasonal Oscillation Using Deep Learning

• Published in: Journal of geophysical research. Machine learning and computation Vol. 2; no. 2
• Main Authors: Anirudh, K. M., Raj, Prasang, Sandeep, S., Kodamana, Hariprasad, Sabeerali, C. T.
• Format: Journal Article
• Language: English
• Published: Wiley 01.06.2025
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000504

The northward‐propagating 30–60 days mode of monsoon rainfall anomalies over India, commonly referred to as the monsoon intraseasonal oscillation (MISO), plays a critical role in driving the active and break spells over the monsoon zone of the country. These oscillations are essential to understanding and predicting the variability of the Indian summer monsoon, which has significant implications for agriculture and water management. This study uses daily precipitation data from the TRMM/GPM satellite to derive MISO indices (MISO1 and MISO2). These indices were obtained through an extended empirical orthogonal function analysis conducted on 25 years of daily rainfall anomalies over the Indian region. The long time series of MISO1 and MISO2 indices generated from this analysis were then used to forecast future values using a transformer‐based deep learning model. The deep learning model demonstrated skilful predictions of the MISO indices for 2018–2022, with forecast lead times extending to 18 days. Notably, the model outperformed conventional operational numerical weather prediction models in predicting the MISO indices. These results indicate the potential for more reliable sub‐seasonal to seasonal (S2S) predictions of the Indian monsoon. The findings from this work highlight the effectiveness of using advanced deep learning techniques, such as Transformer architectures, in enhancing the predictability of complex atmospheric phenomena like MISO, thereby improving the outlook for monsoon forecasting. Plain Language Summary While the chaotic nature of convective processes limits short‐term predictability of tropical weather, we believe longer‐range forecasts are possible due to the influence of large‐scale oscillations and sea‐surface temperature variations. One key oscillation is the intraseasonal oscillation, with a periodicity of 30–60 days, which suggests predictability on the sub‐seasonal to seasonal (S2S) scale. A major manifestation of this is the monsoon intraseasonal oscillation (MISO), the dominant mode of sub‐seasonal variability in the Indian summer monsoon. Current operational numerical weather prediction (NWP) models struggle to forecast monsoon rainfall on the S2S scale. Here, we use a deep learning model called Transformer, the foundation of large language models like ChatGPT, to predict MISO. Our results show that the Transformer can effectively predict MISO with a lead time of about 2 weeks. If implemented operationally, this method could enable reliable predictions of monsoon active and break phases 2 weeks in advance, using far fewer computational resources than traditional NWP models. Key Points A transformer model reliably predicts monsoon intraseasonal oscillation indices 18 days ahead Slow error growth in transformer as compared to conventional NWP methods Transformer model shows better skill than two operational dynamical models