基于三种机器学习方法的降水相态高分辨率格点预报模型的构建及对比分析

冬季降水相态及其转变时间的精细化客观预报对提高气象预报和服务质量具有重要的现实意义。利用京津冀地区国家级自动气象站观测资料及网格化快速更新精细集成产品，统计分析了京津冀地区复杂地形下各类降水相态温度和湿球温度平均气候概率的分布差异及不同降水相态时网格化快速更新精细集成产品中可能影响降水相态判断的特征信息。然后将地面观测天气现象资料、复杂地形下降水相态气候特征及高分辨率模式输出产品作为特征向量，分别基于梯度提升（XGBoost）、支持向量机（SVM）、深度神经网络（DNN）3种机器学习方法建立了降水相态的高分辨率客观分类模型，并对同样条件下3种机器学习方法对雨、雨夹雪和雪3种京津冀主要降水相态的预报效果进行了对比检验，进一步提升了雨夹雪复杂降水相态的客观分类预报技巧。 

The construction and comparison of high resolution precipitation type prediction models based on three machine learning methods

Refined and objective prediction of precipitation type and its transition time in winter is of great practical significance for improving the quality of forecast service. This paper establishes a high-resolution precipitation type prediction model based on temperature and weather phenomena data collected at 174 national automatic weather stations for the period 1955—2019 in Beijing-Tianjin-Hebei and the high-resolution forecast products of rapid update multi-scale analysis and forecast system-integrated subsystem (RMAPS-IN) using three machine learning methods, i.e., the XGBboost, the support vector machine (SVM) and the depth neural network (DNN) prediction models. Firstly, differences in spatial distribution between various precipitation types and corresponding climatologically mean probabilities of air temperature and wet bulb temperature at 174 national stations in Beijing-Tianjin-Hebei region are statistically analyzed. The fine integrated products provided by RMAPS-IN, i.e., 2 m air temperature, dew point temperature, relative humidity, snowline height, the ratio of frozen precipitation to total precipitation in the near surface atmosphere for different precipitation types, and the analysis fields of three-dimensional meteorological elements such as temperature and wet bulb temperature are analyzed. The observational weather phenomena, climatological characteristics of precipitation type over complex terrain and high-resolution model output products are taken as feature vectors. The classification model of precipitation type is then established based on the XGBboost, SVM and DNN, and the prediction effects of three machine learning algorithms on rain, sleet and snow are compared and evaluated. The results show that: (1) the accuracy of the three machine learning methods for rain, sleet and snow prediction can be significantly improved by adding climatological features of precipitation type over complex terrain to the feature parameters; (2) the prediction ability of the three machine learning methods for rain and snow is better than that for sleet; (3) the XGBboost and DNN have the same prediction ability, which are obviously better than SVM.