Minimization of Forecast Error Using Deep Learning for Real-Time Heavy Rainfall Events Over Assam
Dhananjay Trivedi, Omveer Sharma, Sandeep Pattnaik
Abstract
Predicting Heavy Rainfall Events (HREs) with lead time poses a significant challenge for meteorological agencies, especially in mountainous regions like Assam. In this study, we simulated a real-time HRE that occurred between June 13 and 17, 2023, resulting in severe flooding in Assam. To enhance rainfall prediction, we integrated output from the Weather Research and Forecasting (WRF) model into a Deep Learning (DL) model. When comparing the district-level performance of WRF and DL models, it becomes evident that the DL model excels in capturing HREs with a significant accuracy of 54.4%, outperforming WRF’s accuracy of only 22.8%. The proposed model demonstrates a mean absolute error (MAE) of under 30 mm, outperforming WRF’s more than 50 mm MAE for Days 2-4, as compared with the India Meteorological Department (IMD). Remarkably, the DL model accurately represents rainfall intensity and magnitude in the western and southern parts of Assam. This study is the first of its kind to focus on a district-scale analysis in Assam.