测绘学报 ›› 2022, Vol. 51 ›› Issue (3): 361-372.doi: 10.11947/j.AGCS.2022.20200385

• 地图学与地理信息 • 上一篇    下一篇

基于深度学习的中国连续空间覆盖PM2.5浓度预报

毛文婧1, 王卫林1, 焦利民1, 刘安宝2   

  1. 1. 武汉大学资源与环境科学学院, 湖北 武汉 430079;
    2. 福建经纬测绘信息有限公司, 福建 福州 350001
  • 收稿日期:2020-08-13 修回日期:2021-05-08 发布日期:2022-03-30
  • 通讯作者: 焦利民 E-mail:lmjiao@whu.edu.cn
  • 作者简介:毛文婧(1997-),女,博士,研究方向为GeoAI在城市的应用。E-mail:wenjingmao@whu.edu.cn
  • 基金资助:
    国家自然科学基金(41971368);国家重点研发计划(2017YFA0604404)

Continuous spatial coverage PM2.5 concentration forecast in China based on deep learning

MAO Wenjing1, WANG Weilin1, JIAO Limin1, LIU Anbao2   

  1. 1. School of Resources and Environmental Sciences, Wuhan University, Wuhan 430079, China;
    2. Fujian Jingwei Surveying Information Co., Ltd., Fuzhou 350001, China
  • Received:2020-08-13 Revised:2021-05-08 Published:2022-03-30
  • Supported by:
    The National Natural Science Foundation of China (No. 41971368); The National Key Research and Development Program of China (No. 2017YFA0604404)

摘要: 实现全国连续空间覆盖未来小时尺度的PM2.5浓度实时、高精度预报是一个难题。本文建立基于深度学习的多层长短期记忆迭代模型和改进的空间反向传播神经网络S-BPNN模型来实现全国小时尺度PM2.5浓度的空间预报。首先,研究基于空间相关性将全国1286个空气质量监测站点在空间上进行自适应分区,并对各个分区分别构建多层LSTM迭代预报模型实现未来24 h各个监测站点的PM2.5浓度的实时预报。其次,应用改进的S-BPNN空间化模型实现未来24 h全国连续空间覆盖的PM2.5浓度精细化制图。然后,利用2016—2019年中国PM2.5监测站的历史数据进行训练和验证,结果显示预报模型和空间化模型的相关系数R2分别为0.88和0.87,表明模型都能实现较高的精度。最后,基于提出的预报模型和空间化模型,辅助从监测站实时获取的大气污染数据和气象数据,通过搭建的大气污染物浓度预报智能化在线信息原型系统可实时发布预报结果并可进行空间化展示。研究实现了全国连续空间覆盖的PM2.5浓度高时空精度的实时预测,以支持大气污染联防联控和公众环境空间质量信息服务。

关键词: 大气污染, 时空预报, PM2.5, 深度学习

Abstract: To achieve real-time and high-precision spatiotemporal forecast of PM2.5 concentration in China with spatial coverage is still a difficult problem. In this paper, two models based on deep learning were established to realize the spatiotemporal forecast in the hourly scale of PM2.5 concentration in China:a multi-layer long and short-term memory iterative model and an improved spatial back-propagation neural network (S-BPNN) model. First of all, we based on spatial correlation to divide 1286 air quality monitoring stations across the country in space adaptively and built a multi-layer LSTM iterative model for each region to achieve the PM2.5 concentration forecast of each monitoring site in the next 24 hours. Secondly, the improved S-BPNN spatialization model was applied to realize the refined mapping of PM2.5 concentration in a large continuous spatial coverage across the country in the next 24 hours. We integrated historical data of PM2.5 monitoring stations in China from 2016 to 2019 for training and verification. The results show that the correlation coefficients R2 of the forecast model and the spatialized model are respectively 0.88 and 0.89, indicating that the two models could achieve high accuracy. Finally, based on the proposed two models and related data crawled from monitoring stations in real-time, the intelligent online information prototype system for air pollutant concentration forecast can be built to release the forecast results of stations in real-time and display spatial results. The study has realized the real-time prediction of PM2.5 concentration with high spatio-temporal accuracy across the country and has strongly supported joint prevention and control of air pollution and public environmental spatial quality information services.

Key words: air pollution, spatio-temporal forecast, PM2.5, deep learning

中图分类号: