A Step Towards Building Medium-Range Data-Driven Weather Forecasting System Over India

Abstract

Traditional numerical weather prediction (NWP) models, while effective, are computationally expensive and often struggle with fine-scale phenomena due to their inherent limitations. This study explores the potential of deep learning (DL) models, specifically convolutional neural networks (CNNs), for predicting the geopotential height at 500 hPa (Z500), a critical meteorological variable, three days ahead over India. The study uses the Indian Monsoon Data Assimilation and Analysis (IMDAA) dataset for the period 1990 to 2020 at six-hour time intervals for 13 pressure levels. The model uses an encoder-decoder architecture and is improved in a step-by-step approach by altering the filter configuration. The performance of the CNN models is compared against several benchmarks, which include persistence, climatology, weekly climatology, and linear regression. Using all atmospheric levels as input, the bestperforming CNN model achieved the root mean square error (RMSE) of 25.44 m, better than all the benchmarks. The study highlights the capability of CNN models to capture complex, non-linear atmospheric trends, providing a costeffective alternative to traditional NWP models.