A new objective post-processing system with cutting-edge machine learning algorithm for numerical weather prediction is presented. The core of the system, a matrix of artificial neural network trained by using historical in-situ observations and model results, can be applied on the latest numerical weather prediction results and produce real-time forecast for specific stations. The authors investigate the performance of post-processing system for two stations, Zhangjiakou and Beijing, for the period 2005?2020. It is shown that the forecasts produced by the new system are significantly more accurate than those produced by raw model forecasts, single-variable linear regression, and multi-variable linear regression, especially in terms of 3-day forecast. The authors developed all the core and auxiliary code by serving as Zhangjiakou and Beijing post-processing systems, which are routinely deployed since Nov. 1, 2020, to facilitate the weather service for the Olympic Winter Games at Beijing in 2022.

提出一种基于先进机器学习算法的纯客观、实时天气预报后处理方法。该方法使用历史数值天气预报结果和实况观测值, 训练出一个人工神经网络模型, 再将该模型应用于每日实时发布的数值天气预报结果中, 得到台站级别的天气要素预报结果。于张家口市和北京市分别建立该模型, 并使用2005—2020年的观测数据和模式数据进行训练、验证和调优。结果表明, 该模型预报的5 个常规气象要素的预报误差普遍优于一元线性回归、多元线性回归以及数值天气预报模式的原始输出值, 尤其对3天以内的天气预报具有明显优势。基于该模型发展的全自动实时后处理系统已于2020年11月1日开始每日自动化地输出预报结果, 并服务于 2022 年北京冬奥会的气象保障工作。