主办单位:中国气象局沈阳大气环境研究所
国际刊号:ISSN 1673-503X
国内刊号:CN 21-1531/P

气象与环境学报 ›› 2022, Vol. 38 ›› Issue (5): 34-41.doi: 10.3969/j.issn.1673-503X.2022.05.004

• 论文 • 上一篇    下一篇

长短期记忆神经网络(LSTM)模型在低能见度预报中的应用

方楠1(),谢国权1,*(),阮小建2,任晨平1,姜舒婕3,张玮玮2   

  1. 1. 浙江省预警信息发布中心, 浙江杭州 310052
    2. 浙江省气象服务中心, 浙江杭州 310052
    3. 浙江省人工影响天气中心, 浙江杭州 310052
  • 收稿日期:2021-03-23 出版日期:2022-10-28 发布日期:2022-11-04
  • 通讯作者: 谢国权 E-mail:fjt901020@126.com;xiegq1@139.com
  • 作者简介:方楠, 男, 1990年生, 工程师, 主要从事气象灾害预警模型和机器学习算法应用研究, E-mail: fjt901020@126.com
  • 基金资助:
    浙江省气象科技计划项目(2019YB04);浙江省气象服务中心气象科技服务开发项目(2020YB010)

Application of long short-term memory neural network (LSTM) model in low visibility forecast

Nan FANG1(),Guo-quan XIE1,*(),Xiao-jian RUAN2,Chen-ping REN1,Shu-jie JIANG3,Wei-wei ZHANG2   

  1. 1. Zhejiang Early Warning Center, Hangzhou 310052, China
    2. Zhejiang Meteorological Service Center, Hangzhou 310052, China
    3. Zhejiang Weather Modification Center, Hangzhou 310052, China
  • Received:2021-03-23 Online:2022-10-28 Published:2022-11-04
  • Contact: Guo-quan XIE E-mail:fjt901020@126.com;xiegq1@139.com

摘要:

利用浙江省义乌市2015—2019年逐小时气象观测数据(相对湿度、风速、地气温差、能见度)和空气质量指数(Air Quality Index, AQI)数据, 分析了义乌地区低能见度天气(观测能见度lt; 10 km)的分布特征和气象要素条件。利用长短期记忆神经网络(Long Short Term Memory Neural Network, LSTM)模型对逐小时能见度进行模拟, 分别对比了观测能见度作为输入变量与否的模拟效果; 根据义乌地区低能见度天气条件的特征, 将模拟时段分为三个时期(11月至翌年2月, 3—6月, 7—10月), 对比了分时期模拟的效果; 以及评估了模型的预报步长。结果表明: 高湿、高污染、气温高于地温和低风速是义乌地区低能见度天气的主要特征。LSTM模型对单站能见度有较好的模拟效果, 当输入参数中加入历史观测能见度时, 能大幅提高模拟准确度, 日均能见度模拟结果均方根误差RMSE=0.63 km, 平均绝对误差MAE=0.51 km, 拟合优度R2=0.99;分时期进行模拟能得到更精准的模拟结果。本研究中选用的输入要素在冬季(11月至翌年2月)模拟效果最好, RMSE=2.35 km, MAE=1.46 km, 低能见度均方根误差RMSE_10 km=1.81 km, 低能见度平均绝对误差MAE_10 km=1.13 km, R2=0.83; 3—6月的模拟中, 输入变量中不加AQI模拟效果更好, 这意味着3—6月义乌地区的低能见度天气以雾天气为主导, 加入过多变量并不一定能提高模型准确度; 随着预报步长增大, 模型预报效果变差, 预测步长等于3 h, R2=0.71, 预测结果已不具备实际应用意义。

关键词: 大气能见度, 长短记忆神经网络, 预报模型, 空气质量

Abstract:

Based on the hourly meteorological observation data (relative humidity, wind speed, ground-air temperature difference, and visibility) and air quality index (AQI) data from 2015 to 2019 in Yiwu of Zhejiang province, we analyzed the distribution and meteorological conditions of low visibility (observed visibility lt; 10 km) in Yiwu.The Long Short-term Memory Neural Network (LSTM) was used to simulate the hourly visibility, and the simulation results with and without observed visibility as an input member were compared.According to the meteorological conditions of low visibility, the simulation period was divided into three periods (from November to February, from March to June, and from July to October).We compared the simulation performance in different periods and evaluated the model prediction steps.The results showed that the high humidity, high pollution, air temperature higher than the surface temperature, and low wind speed are the main meteorological characteristics influencing the low visibility weather in Yiwu.The LSTM has a good performance in simulating visibility at a single station.When the historical observed visibility has been used as an input parameter, the simulation accuracy can be greatly improved, with the root mean square error (RMSE) of 0.63 km, the mean absolute error (MAE) of 0.51 km, and the fitting goodness R2 of 0.99.Better simulation can be reached after dividing different periods, and the best simulation occurred in winter (from November to February) using the input elements selected in this study, with RMSE=2.35 km and MAE=1.46 km, and for low visibility weather, the RMSE_10km=1.81 km, MAE_10km=1.13 km, and R2=0.83.In the simulation from March to June, the simulation of low visibility without AQI as one input member performs better, which means that low visibility in Yiwu is dominated by foggy weather in this period.Adding too many variables does not necessarily improve the accuracy of the model.With the increase in the forecast step size, the simulation performs worse.When the forecast step size equals 3 h, R2 reaches 0.71 and the simulation result has no practical application significance.

Key words: Atmospheric Visibility, Long Short-term Memory Neural Network (LSTM), Forecast model, Air quality

中图分类号: