In 2025, the Hong Kong Observatory introduced, for the first time in operation, artificial intelligence (AI) global weather prediction models for operational forecasting of tropical cyclones (TCs) over its area of responsibility, including the South China Sea and part of the western north Pacific. This paper documents the first-year operational experience of using such AI models and provides verification results for their track forecasting, temporal consistency and genesis. The AI models are found to have reduced forecast track errors, converge towards the actual tracks earlier, and have higher temporal consistency among successive forecast runs, relative to the selected traditional global numerical weather prediction (NWP) models. When combined the traditional global NWP models with AI models, a grand ensemble is possible to enhance the robustness of tropical cyclone warning service for Hong Kong, supporting earlier and more consistent forecast to facilitate early preparation work of the public and the emergency preparedness parties. While AI models are expected to be an indispensable tool for operational TC warning, their limitations on TC intensity and wind-structure prediction should be acknowledged. It should also be noted that present evaluation covers a single TC season, over a regional subset of TCs and a selected set of AI models; multi-year and multi-basin verification is needed to assess generality.
Choy et al. (Fri,) studied this question.