Deep learning for short-range monsoon rainfall forecast using ground truth rainfall data (Papers Track)
Aastha Jain (Columbia University); Jatin Batra (Tata Institute of Fundamental Research); Apoorva Narula (Georgia Institute of Technology); Rajeevan Madhavan Nair (Atria University); Sandeep Juneja (Ashoka University)
Abstract
The Indian summer monsoon is a highly complex and critical weather system that directly affects the livelihoods of about a billion and a half people across the Indian subcontinent. Accurate short-term forecasting remains a major scientific challenge due to the monsoon's sensitivity to multi-scale drivers, including local land-atmosphere interactions and large- scale ocean-atmosphere phenomena. In this study, we address the problem of forecasting daily rainfall across India during the summer months, focusing on both one-day and three-day lead times. We use Autoformers - deep learning transformer-based architectures designed for time series forecasting. These are trained on historical gridded precipitation data from the Indian Meteorological Department (1901--2023) at spatial resolutions of $0.25^\circ \times 0.25^\circ$. The models also incorporate auxiliary meteorological variables from ECMWF’s reanalysis datasets. Forecasts are benchmarked against ECMWF’s High-Resolution Ensemble System (HRES), widely regarded as the most accurate numerical weather predictor. \textbf{Our results provide the first evidence that a machine learning model can outperform HRES for short-term monsoon forecasts.} Specifically, compared to our model, forecasts from HRES model have about 22\% higher error, for a single day prediction, and over 27\% higher error, for a three day prediction. Such enhanced forecast accuracy translates into tangible climate adaptation benefits, enabling earlier flood warnings and helping farmers protect crops from unexpected downpours. We also find that incorporating historical data up to 20 days prior reduces forecast error, particularly in landlocked regions.