Real-time Operational System of AI-Based Medium-Range Weather Forecasting Models by KMA

Weather forecasting has evolved through various Numerical Weather Prediction (NWP) models over the years, becoming essential tools for understanding and predicting complex weather systems. Recent advancements in Artificial Intelligence (AI) technologies have significantly impacted the weather forecasting. AI-based weather prediction models capable of forecasting beyond 10 days (i.e., Medium-Range Weather Forecasting) have demonstrated performance comparable to or superior to traditional NWP models like the Integrated Forecasting System (IFS) under specific conditions. FourCastNet2 showed accuracy similar to IFS in short-term predictions of large-scale variables and performed better with variables having complex microstructures. Pangu-Weather surpassed traditional NWP models in accuracy across all time ranges and variables, and GraphCast presented new possibilities for medium- range predictions using the latest graph neural network (GNN) architecture. In this study, we conducted real-time forecasts using three AI prediction models (i.e., Pangu- Weather, GraphCast, FourCastNet2) introduced by the Korea Meteorological Administration (KMA), leveraging the initial fields of KIM (Korea Integrated Model), UM (Unified Model), and IFS, at 6-hour intervals from +006h to +288h. We explain the preprocessing, optimization, and display system development processes for real- time operations and present the results of performance and reliability comparisons between AI prediction models and operational NWP models. This study aims to evaluate the feasibility of real-time operation of AI-based weather forecasting models and contribute to enhancing the reliability of future weather forecasts.