干旱气象 ›› 2020, Vol. 38 ›› Issue (4): 655-664.

• 业务技术应用 • 上一篇    下一篇

深度神经网络方法在山东降水相态判别中的应用

朱文刚1,李昌义1,曲美慧2,温晓培1
Application of Deep Neural Networks Method in Precipitation Phase Identification in Shandong Province   

  1. (1.山东省气象科学研究所,山东济南250031;2.吉林省气象科学研究所,吉林长春130062)
  • 出版日期:2020-08-31 发布日期:2020-09-04
  • 作者简介:朱文刚(1985— ),男,硕士,工程师,主要从事数值天气预报技术和资料同化应用研究. E-mail: zhu122812@163.com。
  • 基金资助:
    山东省重点研发计划项目(2016GSF120017)和山东省气象局青年科研基金项目(2016SDQN08)共同资助

Application of Deep Neural Networks Method in Precipitation Phase Identification in Shandong Province

ZHU Wengang1, LI Changyi1, QU Meihui2, WEN Xiaopei1   

  1. (1. Shandong Institute of Meteorological Sciences, Jinan 250031, China;
    2. Jilin Institute of Meteorological Sciences, Changchun 130062, China)

  • Online:2020-08-31 Published:2020-09-04

摘要: 利用欧洲中心(ECMWF)ERA-Interim再分析资料,通过分析2008—2017年山东冬半年不同降水相态(雨、雪和雨夹雪)下温度和位势厚度特征,统计得到8个降水相态判别因子(T2 m、T1000、T975、T950、T925、T850、H850-700、H1000-850),并给出每个判别因子降水相态阈值指标。然后利用8个判别因子和阈值建立降水相态判别方程和训练DNN模型,通过随机检验发现DNN法对雨、雪和雨夹雪的预报准确率分别提高1.9%、0.2%和21.6%;利用ECMWF细网格预报资料进行个例检验,雨、雪和雨夹雪共106站中判别方程法判别错误29站,DNN法判别错误14站,即DNN法的降水相态判别能力优于判别方程法,且明显提高了对雨夹雪的判别能力。


关键词: 关键词:降水相态, 判别方程, DNN法, 温度阈值, 厚度阈值, 预报准确率

Abstract:  Based on ECMWF ERA-Interim reanalysis data, 8 factors (T2 m, T1000, T975, T950, T925, T850, H700-850 and H850-1000) for identifying precipitation phases were obtained through analyzing the temperature and geopotential thickness of precipitation phases (rain, snow, sleet) in winter half year from 2008 to 2017 in Shandong Province, and the threshold indicators of the 8 factors were provided. The discriminant equation for precipitation phase identification was established and the deep learning DNN model was trained using the 8 factors and their threshold values, and the forecast accuracy of rain, snow and sleet increased by 1.9%, 0.2% and 21.6% using DNN method through randomization test, respectively. The inspection using ECMWF fine grid model products indicated that among a total of 106 stations of rain, snow and sleet, the discriminant equation and DNN method carried out wrong identifications for 29 and 14 stations, respectively. The results show that the DNN method performed better than the discriminant equation, and in particular, it significantly improved the identification ability of sleet.


Key words: Key words: precipitation phase, discriminant equation, DNN method, temperature threshold, thickness threshold, forecast accuracy