Abstract In a climate change scenario, India is experiencing an escalation in the frequency of localized heavy rainfall and flood events in recent years. Consequently, it is imperative to evaluate and improve the capacity for high‐resolution spatiotemporal rainfall prediction. For dynamical models, accurately forecasting rainfall characteristics at the district level with a finer temporal resolution (1 h) and a lead time of 72 h presents a substantial challenge. This research endeavour seeks to address this knowledge deficiency by evaluating the efficacy of an ensemble methodology employing spatial attention‐based deep learning (DL) techniques, juxtaposed with deterministic multiphysics models trained at both state and district scales, to forecast rainfall up to 3 days ahead for the state of Odisha, India. The models have been trained (and tested) on 7 (1) monsoon depressions (MDs), 2 (1) post‐monsoon depressions (PMDs), and 6 (1) monsoon deep depressions (DDs), and tested on 1 MD, 1 PMD, and 1DD. Both DL models exhibited superior performance compared to the deterministic ensemble weather research and forecasting (WRF) model, achieving an average mean absolute error of (>70 mm), while the DL model trained at the district scale (0.4) of the equitable threat score (ETS) during afternoon and evening hours indicate the superiority of the ensemble‐based DL models in predicting various rainfall categories. The approach is tailored to depression‐driven heavy rainfall, and the results demonstrate substantial improvements over deterministic multi‐physics WRF outputs for these synoptic systems. These findings underscore the capability of an ensemble utilizing DL models to predict rainfall with precision at both spatial and finer temporal scales.
Trivedi et al. (Wed,) studied this question.