设为首页 加入收藏 登录旧版
基于机器学习的短临预报方法及其在空气质量保障中的应用
Short-Term Forecast Method Based on Machine Learning and Its Application in Air Quality Assurance
投稿时间:2020-05-16  修订日期:2020-09-05
DOI:10.19316/j.issn.1002-6002.2021.03.09
中文关键词:  数值预报  机器学习  短临预报  空气质量
英文关键词:numerical forecast  machine learning  short-term forecast  air quality
基金项目:国家重点研发计划(2018YFC0213206)
作者单位
肖林鸿 中国科学院大气物理研究所, 大气边界层物理和大气化学国家重点实验室, 北京 100029
中科三清科技有限公司, 北京 100029 
陈焕盛* 中国科学院大气物理研究所, 大气边界层物理和大气化学国家重点实验室, 北京 100029 
陈婷婷 中科三清科技有限公司, 北京 100029 
吴剑斌 中国科学院大气物理研究所, 大气边界层物理和大气化学国家重点实验室, 北京 100029
中科三清科技有限公司, 北京 100029 
杨文夷 中国科学院大气物理研究所, 大气边界层物理和大气化学国家重点实验室, 北京 100029 
王文丁 中国科学院大气物理研究所, 大气边界层物理和大气化学国家重点实验室, 北京 100029 
田敬敬 中科三清科技有限公司, 北京 100029 
张稳定 中科三清科技有限公司, 北京 100029 
通讯作者:陈焕盛*  中国科学院大气物理研究所, 大气边界层物理和大气化学国家重点实验室, 北京 100029  
摘要点击次数: 175
全文下载次数: 12
中文摘要:
      基于多个数值模式(NAQPMS、CMAQ)预报结果和站点观测资料,采用岭回归机器学习方法构建了一种多污染物短临预报方法,并应用于2017年厦门金砖国家峰会和2019年武汉军运会空气质量保障工作,评估总结了短临预报方法的可行性和预报效果。结果表明:短临预报表现出很好的预报能力,可以有效改进数值模式的预报趋势和量值。从日均模拟效果来看,短临预报的颗粒物相关系数达到0.9以上,均方根误差小于5 μg/m3,而且预报的臭氧相关系数和均方根误差分别为0.99和2 μg/m3。相较于日均数值模式预报结果,短临预报可使颗粒物和臭氧的预报相关系数至少提升16%和9%,预报均方根误差至少下降50%和86%。对于小时模拟效果,短临预报能够更好地把握颗粒物的小时变化特征,其相关系数较模式提升了0.2以上;同时短临预报能够更好地再现观测到的臭氧日循环特征,对白天的峰值和夜晚的低值预报更好,其相关系数高达0.9。总体来说,短临预报方法可为重大活动空气质量保障提供较为精准的临近预报结果,具有较好的应用价值。
英文摘要:
      Based on the simulation results of several numerical models (NAQPMS,CMAQ) and the observation dataset,a multi-pollutant short-term forecast system was constructed using ridge regression machine learning algorithm. The forecast system was successfully applied to air quality assurance for the 2017 Xiamen BRICS Summit and the 7th CISM Military World Games. The feasibility and performance analysis of such system are summarized in this work. The results showed that the method has good performance in the short-time forecasting, which can improve the simulation results of the numerical models,in both forecasting trend and magnitude. For daily average forecasting results,the correlation coefficient of particulates is higher than 0.9,the root mean square error is less than 5 μg/m3.And for daily ozone,the correlation coefficient and root mean square error is 0.99 and 2 μg/m3,respectively. Compared with the numerical model results,the short-term forecast method can improve the forecast results by increasing the correlation coefficient of predicted particulates and ozone by at least 16% and 9%,and reducing the root mean square error by at least 50% and 86%. For hourly simulation results,the short-term forecast method shows better performance in capturing the hourly variability of particulate matter with correlation coefficients more than 0.2 higher than results of numerical models. And the short-term forecast method can better reproduce the observed ozone cycle characteristics and provide better predictions of daytime peaks and nighttime lows,with correlation coefficient of 0.9. In general,the short-term forecast method can provide more accurate nowcast results,which can be used in the air quality assurance in certain important public events.
查看全文  查看/发表评论  下载PDF阅读器