Development of a Data-driven weather forecasting system over India with Pangu-Weather architecture and IMDAA reanalysis Data

Abstract: Numerical Weather Prediction (NWP) has advanced significantly in recent decades but still faces challenges in accuracy, computational efficiency, and scalability. Data-driven weather models have shown great promise, sometimes surpassing operational NWP systems. However, training these models on massive datasets incurs high computational costs. A regional data-driven approach offers a cost-effective alternative for localized forecasts. This study develops a regional weather forecasting model for India by efficiently modifying the Pangu-Weather (PW) architecture. The model is trained using the Indian Monsoon Data Assimilation and Analysis (IMDAA) reanalysis dataset with limited computational resources. Prediction results are evaluated using Root Mean Square Error (RMSE), Anomaly Correlation Coefficient (ACC), Mean Absolute Percentage Error (MAPE), and Fractional Skill Score (FSS). At a 6-hour lead time, MAPE remains below 5%, FSS exceeds 0.86, and ACC stays above 0.94, demonstrating robustness. Three forecasting approaches, static, autoregressive, and hierarchical, are compared. Errors increase with lead time in all cases. The static approach exhibits periodic fluctuations in error metrics, which are absent in the autoregressive method. The hierarchical approach also shows fluctuations, though with reduced intensity after three days. Among these, the hierarchical approach performs best while maintaining computational efficiency. Furthermore, the model effectively predicts cyclone tracks using the hierarchical approach, achieving results comparable to observational and reanalysis datasets.