干旱气象 ›› 2026, Vol. 44 ›› Issue (2): 338-347.DOI: 10.11755/j.issn.1006-7639-2026-02-0338

• 技术报告 • 上一篇    

北京地区能见度变化特征及其临近外推预报研究

姜江1(), 夏江江2, 刘祺2, 乔媛1   

  1. 1 北京市气象服务中心北京 100089
    2 中国科学院大气物理研究所北京 100010
  • 收稿日期:2025-05-07 修回日期:2025-09-19 出版日期:2026-05-20 发布日期:2026-05-18
  • 通讯作者: 乔媛(1987—),女,高级工程师,主要从事气象服务与业务应用工作。E-mail: qiaoyuan@@bj.cma.gov.cn。
  • 作者简介:姜江(1987—),女,高级工程师,主要从事气象服务与业务应用工作。E-mail: 625162362@qq.com
  • 基金资助:
    北京市自然科学基金青年项目(8214066);“基于机器学习方法的北京道面能见度临近预报应用研究”资助

Visibility variation characteristics and its nowcasting extrapolation in Beijing

JIANG Jiang1(), XIA Jiangjiang2, LIU Qi2, QIAO Yuan1   

  1. 1 Beijing Meteorological Service CenterBeijing 100089, China
    2 Institute of Atmospheric PhysicsChinese Academy of SciencesBeijing 100010, China
  • Received:2025-05-07 Revised:2025-09-19 Online:2026-05-20 Published:2026-05-18

摘要:

能见度作为表征大气透明度的重要气象要素,对交通安全、航空运输及公众健康具有重要影响。本文基于2013—2022年自动气象站数据,分析北京地区能见度的区域特征和变化趋势,同时开展深度学习能见度临近外推预报研究。结果表明:1)不同等级能见度出现频率依次为“好”(49.70%~75.12%)、“一般”(14.39%~40.82%)、“较差”(2.75%~10.95%)、“差”(0.10%~6.28%)、“极差”(0~3.01%);2)基于前72、60、48、36、24、12、6 h开展的循环神经网络(Recurrent Neural Network,RNN)、长短期记忆网络(Long Short-Term Memory,LSTM)方法能见度临近外推预报试验中,基于前48 h RNN外推未来1 h能见度效果最优,基于前36 h RNN外推未来2 h能见度效果最优,而基于前6 h LSTM外推未来1 h和2 h能见度效果均最优;3)RNN和LSTM均可以把握未来0~2 h不同地理位置能见度的变化,也具备对于受天气形势影响的能见度“转折点”的判断能力。整体上,RNN外推预报效果优于LSTM,但两种模型的预报结果均存在一定的滞后性。

关键词: 北京, 能见度, 深度学习, 外推预报

Abstract:

As an essential meteorological indicator of atmospheric transparency, visibility holds significant implications for traffic safety, aviation operations, and public health. In this study, the characteristics and changes of visibility in Beijing were analyzed based on the data of automatic weather stations from 2013 to 2022, and the research on the nowcasting extrapolation of visibility based on deep learning was carried out. The main conclusions are as follows: 1) The frequency of different visibility levels ranges from 49.70% to 75.12% for “good”, 14.39% to 40.82% for “fair”, 2.75% to 10.95% for “bad”, 0.10% to 6.28% for “poor”, and 0 to 3.01% for “very poor”; 2) In the visibility nowcasting extrapolation experiments conducted using Recurrent Neural Network (RNN) and Long Short-term Memory Network (LSTM) based on the previous 72, 60, 48, 36, 24, 12, and 6 hours, the RNN based on the previous 48 hours had the best effect in extrapolating future one hour visibility, the RNN based on the previous 36 hours had the best effect in extrapolating future two hours visibility, and the LSTM based on the previous 6 hours had the best effect in extrapolating both future one hour and two hours visibility; 3) Both RNN and LSTM can predict the changes in visibility at different locations in the next 0 to 2 hours, and also have the ability to identify the “turning points” of visibility affected by weather conditions. Overall, the extrapolation forecasting effect of RNN is better than that of LSTM, but the forecast results of both models still have a certain lag.

Key words: Beijing, visibility, deep learning, extrapolation

中图分类号: