To address the challenges of refined numerical weather prediction in complex terrain regions, we developed a hybrid regional forecast framework (NG-WRF) driven by NeuralGCM (NG), a physically constrained artificial intelligence (AI)-based global model. This framework leverages NG’s superior dynamic and thermodynamic consistency for upper-air initial and boundary conditions, supplemented by GFS surface variables to drive the WRF model, thereby achieving a high-resolution downscaling process. In evaluation experiments based on a precipitation event in the Shigatse region of the Tibetan Plateau, this study found that the NG-WRF forecast framework performs better than the traditional GFS-driven scheme. Notably, the spatial correlation coefficient between the NG-WRF model’s 36-hour cumulative precipitation forecast and observations nearly doubled compared to the GFS-driven forecast, yielding higher threat scores at the 10-mm cumulative precipitation threshold. Furthermore, the NG-WRF scheme maintains high stability and coordination in large-scale fields as lead times increase, with spatial correlations for 2-m temperature and humidity exceeding 0.94. These results demonstrate that integrating physically constrained AI models with regional numerical models is a robust strategy for enhancing high-precision forecasting over complex topography. 为应对复杂地形区域精细化数值天气预报的挑战, 本文开发了基于物理约束人工智能全球模式NeuralGCM (NG) 驱动的区域中尺度数值预报模式WRF的混合区域预报框架 (NG-WRF) .该框架利用NG模型在高空初始条件和边界条件方面卓越的动力学与热力学一致性, 辅以GFS地面变量驱动WRF模式, 从而实现高分辨率降尺度过程.在基于青藏高原日喀则地区降水事件的评估试验中, 本研究发现NG-WRF预报框架相比传统GFS驱动方案表现出更好的效果.值得注意的是, NG-WRF框架的36小时累计降水预报与观测值的空间相关系数较GFS驱动预报提升近一倍, 在10毫米累计降水阈值下TS评分更高.此外, 随着预报提前期延长, NG-WRF方案在大尺度场中保持着高稳定性和协调性, 2米高度温湿度空间相关系数均超过0.94.这些结果表明, 将物理约束型人工智能模式与区域数值模式集成, 是提升复杂地形高精度预报的有效路径.
何秉亮 et al. (Wed,) studied this question.