A Modeling Effort for Medium-Range Weather Prediction over India with a Modified Pangu-Weather Architecture

Abstract

Weather forecasting has become essential in today’s world due to its importance in various fields. The global Deep Learning (DL) weather prediction models are trained on long-term historical datasets provided by the numerical weather prediction (NWP) systems. A regional model can be advantageous for weather predictions over a specific region. The current study proposed a DL-based model for weather forecasting over India, with inference time less than a minute for a seven-day prediction. The model’s architecture is inspired by the Pangu-Weather (PW), a data-driven global weather forecasting model. The model was trained (for the period 1990–2017) with the Indian Monsoon Data Assimilation and Analysis (IMDAA), with a spatial resolution of 0.12 degrees, higher than any other reanalysis datasets available over this region. A sensitivity study on the latent space dimension (C) and global batch size (B) was conducted to further improve the model’s computational efficiency. The effect of the training data period on model performance was examined, suggesting the importance of both recent and long-term information for optimal performance. Two different prediction methodologies, viz., autoregressive and hierarchical, were adopted and compared against traditional baseline models with various statistical matrices. The hierarchical approach exhibited superior performance, but with periodic fluctuations in error, particularly for predictions over three days. The model also displayed encouraging performance in predicting extreme weather, viz., cyclones, heat waves, and cold waves.