摘要
PM_(2.5)和O_(3)是影响我国城市和区域空气质量的主要因子,探究其污染特征并对其浓度进行预测是预防大气复合污染的基础工作。首先,利用2015—2023年长江中游宜昌和武汉两个主要城市国控站PM_(2.5)和O_(3)浓度监测数据,分析PM_(2.5)和O_(3)复合污染特征;然后,利用机器学习模型可解释工具(Shapely Additive Explanation,SHAP),揭示气温、相对湿度、降水量、日照、风速等气象因子对PM_(2.5)和O_(3)浓度的影响及贡献;再构建基于融合门控循环单元(GRU)等9种深度学习方法的PM_(2.5)和O_(3)浓度预测模型,并进行效果检验。结果表明:(1)2015—2023年长江中游宜昌和武汉O_(3)浓度每年平均依次升高3.89μg·m^(−3)和2.73μg·m^(−3),夏、秋季升高更明显;PM_(2.5)浓度则呈显著下降趋势,趋势率分别为−3.59μg·m^(−3)·a^(−1)和−3.36μg·m^(−3)·a^(−1),冬、春季下降更显著,表明近年来PM_(2.5)污染治理起到显著成效。(2)PM_(2.5)和O_(3)浓度月际间分别呈“U”和“M”型分布,两者呈弱的负相关关系。2015—2023年宜昌和武汉PM_(2.5)和O_(3)浓度“双高”天数分别为60 d和39 d,主要集中在2—5月和10—12月之间,且呈年下降趋势(2023年略有升高)。(3)对比分析9种深度学习预测模型表明,门控循环单元(GRU)、双向门控循环单元(BIGRU)、基于注意力机制门控循环单元(Attention-GRU)及基于注意力机制双向门控循环单元(Attention-BiGRU)共4种模型在宜昌和武汉PM_(2.5)和O_(3)浓度预测中效果较好,其中GRU运行时间最短,可有效提高PM_(2.5)和O_(3)浓度预测和服务的及时性。(4)构建的基于融合深度学习回归预测模型与GRU相比,宜昌和武汉O_(3)浓度预测均方根误差RMSE分别减小5%和8%,PM_(2.5)浓度预测RMSE分别减小20%和16%。该模型对PM_(2.5)和O_(3)复合污染日预测Ts评分宜昌为60.00%,武汉为69.23%,可为长江中游宜昌和武汉两城市受气象条件影响的大气PM_(2.5)和O_(3)浓度预测及复合污染防治提供科学依据。
PM_(2.5)and O_(3)were the main factors affecting urban and regional air quality in China.Exploring their pollution characteristics and predicting their concentrations are the basic work for the prevention and control of composite atmospheric pollution.The PM_(2.5)and O_(3)compound air pollution characteristics were analyzed using the concentration data of PM_(2.5)and O_(3)in Yichang and Wuhan from 2015 to 2023,which are two major cities in the middle reaches of the Yangtze River.The machine learning model explainable tool SHAP(shapely additive explanation)was used to reveal the contribution of meteorological factors such as temperature,humidity,precipitation,sunshine,wind speed,and other meteorological factors to the influence of PM_(2.5)and O_(3)concentrations.Based on nine deep learning methods including GRU(gated recurrent unit),a regression prediction model was constructed for PM_(2.5)and O_(3)concentrations,and their performance was evaluated.The results are as follows.(1)From 2015 to 2023,the O_(3)concentrations in Yichang and Wuhan increased annually by 3.89μg.m^(−3)and 2.73μg.m^(−3)on average,with a more evident increase in summer and autumn.However,the PM_(2.5)concentrations showed a significant decrease trend,with the trend rates of-3.59μg.m^(−3).a^(−1)and-3.36μg.m^(−3).a^(−1),respectively.The decrease was more notable in winter and spring,indicating the significant effectiveness of pollution control in recent years.(2)The monthly changes of PM_(2.5)and O_(3)concentrations showed“U”and“M”features,with a weak negative correlation between PM_(2.5)and O_(3).From 2015 to 2023,the compound air pollution days with high concentrations of both PM_(2.5)and O_(3)in Yichang and Wuhan were 60 and 39 days,respectively,which mainly concentrated in February to May and October to December.An annually decreasing trend was found,except for a slight increase in 2023.(3)Comparative analysis of a total of nine deep learning prediction models showed that four models,including GRU(gated recurrent unit),BIGRU(bidirectional gated recurrent units),Attention-GRU(attention gated recurrent unit),and Attention-BiGRU(attention bidirectional gated recurrent units),performed better in predicting PM_(2.5)and O_(3)concentrations in Yichang and Wuhan.Particularly,the GRU model had the shortest running time,which could effectively improve PM_(2.5)and O_(3)concentration prediction performance and services.(4)Compared with GRU,the root mean square error for O_(3)concentration prediction of the regression prediction model was reduced by 5%and 8%in Yichang and Wuhan,with PM_(2.5)concentration prediction reduced by 20%and 16%,respectively.Ts scores for daily predictions of PM_(2.5)and O_(3)were 60.00%in Yichang and 69.23%in Wuhan.The regression prediction model can provide a scientific basis for the prediction of PM_(2.5)and O_(3)concentration and the control of compound pollution in Yichang and Wuhan cities which were affected by meteorological conditions.
作者
马德栗
鞠英芹
史瑞琴
湛甜
MA Deli;JU Yingqin;SHI Ruiqin;ZHAN Tian(Hubei Province Climate Center,Wuhan 430074;Hubei Provincial Meteorological Engineering Center,Wuhan 430074)
出处
《暴雨灾害》
2025年第3期410-420,共11页
Torrential Rain and Disasters
基金
中华人民共和国水利部三峡气候局地监测项目(SK2021031)
中国长江三峡集团有限公司项目(NO0704182)
武汉市气象科技联合项目(2023020201010581)。
关键词
深度学习
气象特征
复合污染
预测模型
deep learning
meteorological characteristics
compound pollution
prediction model
