A Forecasting Paradigm Shift: From Traditional Methods to Physics-AI Combination in China’s National-Level Severe Convective Weather Operations

Severe convective weather (SCW) is one of the primary meteorological phenomena causing significant disasters and casualties in China. Since 2009, the National Meteorological Center (NMC) of the China Meteorological Administration has made remarkable progress in SCW monitoring, forecasting, and warning technologies and operations. This paper summarizes the operational developments and scientific achievements of NMC in SCW from four aspects including operations, mechanism studies, monitoring and forecasting technologies, and operational platforms. The NMC has established a comprehensive operational system for SCW, achieving continuous improvement in forecast accuracy and refinement. Key advances are as follows. The climatological characteristics and environmental conditions of SCW in China have been revealed, along with the impacts of the Asian monsoon and cold vortices on convective storms. Identification technologies have been developed for tornadoes, thunderstorm winds, and downbursts using multi-source observations, physical structure characteristics, and deep learning (DL) algorithms. By integrating multi-source observational data with numerical weather prediction and applying DL methods, DL models for nowcasting (0‒2 h), short-term (0‒12 h) and short-range (0‒72 h) forecasting have been constructed, with Fenglei nowcasting model achieving a reliable 3-h forecast. SWAN (Severe Weather Analysis and Nowcasting) 3.0 platform has been implemented, enabling real-time coordinated nowcasting across all-level meteorological departments for the first time in China. Future efforts will focus on enhancing high-precision observation networks, advancing ultra-high-resolution numerical models, developing large-scale and specialized models by integrating physical mechanisms and DL methods, addressing key technical challenges including high false alarm rates, limited lead time, and intensity prediction difficulties.