Numerical weather prediction (NWP) centers around the world operate a variety of NWP models, and recent advances in AI-driven NWP models have increased the availability of diverse NWP outputs. While this expansion holds the potential to improve forecast accuracy, it also raises a critical challenge of identifying the most reliable predictions for specific forecast scenarios. Traditional approaches, such as ensemble or weighted averaging, combine multiple NWP outputs but often generate unrealistic atmospheric fields, complicating the production of reliable and consistent forecasts in operational settings. In this study, we introduce DeepMedcast, a deep learning method that generates intermediate forecast, or "medcast", between two or more NWP outputs. Unlike ensemble averaging, DeepMedcast can provide consistent and explainable medcast without distorting meteorological fields. This paper details the methodology and case studies of DeepMedcast, discussing its advantages and potential contributions to operational forecasting.

翻译：世界各地的数值天气预报（NWP）中心运行着多种NWP模型，而近期人工智能驱动的NWP模型进展进一步增加了多样化NWP输出的可获取性。尽管这种扩展有望提升预报精度，但也带来了一个关键挑战：如何针对特定预报场景识别最可靠的预测结果。传统方法（如集合平均或加权平均）虽能融合多个NWP输出，但常产生不真实的大气场，使得业务环境中难以生成可靠且一致的预报。本研究提出DeepMedcast——一种能够在两个或多个NWP输出间生成中间预报（或称"中值预报"）的深度学习方法。与集合平均不同，DeepMedcast能在不扭曲气象场的前提下提供具有一致性和可解释性的中值预报。本文详述了DeepMedcast的方法论与案例研究，并探讨了其在业务化预报中的优势与潜在贡献。