农业大数据学报 ›› 2019, Vol. 1 ›› Issue (1): 78-87.doi: 10.19788/j.issn.2096-6369.190108

• 应用研究 • 上一篇    下一篇

气象大数据超短临精准降水机器学习与典型应用

张晨阳1,2,3(),杨雪冰1,3,张文生1,2,3,*()   

  1. 1. 中国科学院自动化研究所,北京 100190
    2. 中国科学院大学,北京 100049
    3. 气象大数据与机器学习联合实验室,北京 100190
  • 收稿日期:2018-08-28 出版日期:2019-03-26 发布日期:2019-04-04
  • 通讯作者: 张文生 E-mail:zhangchenyang2016@ia.ac.cn;zhang@ia.ac.cn
  • 作者简介:张晨阳,男,博士。研究方向:机器学习,深度学习,农业气象数据挖掘|杨雪冰,男,博士、助理研究员。研究方向:机器学习,农业气象数据挖掘;Email: <email>yangxuebing2013@ia.ac.cn</email>
  • 基金资助:
    国家自然科学基金(61602482、61472423)

Accurate Precipitation Nowcasting with Meteorological Big Data: Machine Learning Method and Application

Zhang Chenyang1,2,3(),Yang Xuebing1,3,Zhang Wensheng1,2,3,*()   

  1. 1. Institute of Automation, Chinese Academy of Sciences, Beijing 100190
    2. University of Chinese Academy of Sciences, Beijing 100049
    3. Joint Laboratory of Meteorological Big Data and Machine Learning, Beijing 100190
  • Received:2018-08-28 Online:2019-03-26 Published:2019-04-04
  • Contact: Zhang Wensheng E-mail:zhangchenyang2016@ia.ac.cn;zhang@ia.ac.cn

摘要:

超短临精准的降水估计对农业生产、水文监测、洪涝减灾、大型活动、电力系统等方面具有重要意义。由于天气系统具有高度不确定性,传统基于物理模型和统计分析的气象方法在降水估计中难以满足高分辨率条件下的精度要求,如何提升超短临降水估计的精准性在研究和应用领域是极具挑战性的问题。本文提出了基于地形的加权随机森林(terrain-based weighted random forests, TWRF)方法用于雷达定量降水估计(quantitative precipitation estimation, QPE)。该方法可视为随机森林方法的推广,并在此基础上考虑了反射率垂直廓线(vertical profile of reflectivity, VPR)的特征重要性变化以及复杂地形区域降水的山岳抬升效应。对TWRF在中国杭州湾地区Z9571雷达45~100km覆盖范围内2014年6月份和7月份的降水过程上进行了实验验证,结果表明TWRF方法优于传统气象方法及随机森林方法,并表明利用整个VPR与基于地形的建模可以有效提升雷达QPE效果。

关键词: 气象大数据, 定量降水估计, 随机森林, 机器学习, 天气雷达

Abstract:

Accurate precipitation nowcasting is essential for agricultural production, hydrological monitoring, flood mitigation, large organizations, and electrical systems. Because of the high uncertainty of weather systems, the performance of precipitation estimation by conventional meteorological methods based on physical models and statistical analysis is unsatisfactory. Determining how to improve the accuracy of precipitation estimation and forecasting in high resolution is challenging. This study proposes the method of terrain-based weighted random forests (TWRF) for radar-based quantitative precipitation estimation (QPE). This method can be regarded as a generalization of random forests via consideration of variations in the vertical profile of reflectivity (VPR) and orographic enhancement of precipitation for complex terrains. The performance was tested within the 45~100 km range of the Z9571 radar in Hangzhou, China during rainfall events in June and July, 2014. The experimental results showed that TWRF is better than conventional methods and random forests, and further indicate that utilization of the entire VPR and terrain-based modeling are effective for radar QPE.

Key words: meteorological big data, quantitative precipitation estimation, random forests, machine learning, weather radar

中图分类号: 

  • S-1