一种基于卷积神经网络的源解析因子识别方法  

An identification method of source apportionment factor based on convolutional neural network

作　　者：孟祥来 孙扬[1] 廖婷婷 张琛 张成影 MENG Xianglai;SUN Yang;LIAO Tingting;ZHANG Chen;ZHANG Chengying(Innovation Transformation Base,Institute of Atmospheric Physics,Chinese Academy of Sciences,Huainan 232000;University of Chinese Academy of Sciences,Beijing 100049;Plateau Atmospheric and Environment Key Laboratory of Sichuan Province,College of Atmospheric Science,Chengdu University of Information Technology,Chengdu 610225)

机构地区：[1]中国科学院大气物理研究所,创新转化基地,淮南232000 [2]中国科学院大学,北京100049 [3]成都信息工程大学大气科学学院,高原大气与环境四川省重点实验室,成都610225

出　　处：《环境科学学报》2022年第8期117-126,共10页Acta Scientiae Circumstantiae

基　　金：国家重点研发计划(No.2018YFC0214003,2016YFA0602004)。

摘　　要：为解决源解析因子识别过程中人为参与及判定过程复杂的问题,提高源解析因子识别工作的效率,提出基于卷积神经网络(CNN)的源解析因子识别方法.通过文献调研,构造可供CNN模型训练的因子识别数据集,对因子识别模型进行训练与调试,并以北京市南部采样点的PM_(2.5)组分观测数据对模型进行验证.同时,利用正定矩阵分解模型(PMF)解析得到不同因子数时的源谱矩阵,输入因子识别模型并与人工分析比对.结果表明,9个因子时模型的识别效果最佳,可以实现既无重复识别又无“无法判定”的情况,与源解析因子人工识别结果吻合,证明了所提出方法的合理性与可行性.该方法不仅对源解析中因子识别问题具有一定的实用价值,同时对减排策略的制定与动态调整也具有积极意义.To solve the problem of human involvement and the complexity of determination process in the identification of source apportionment factor,and to improve the efficiency of the source apportionment factor recognition work,a method of source apportionment factor identification based on convolutional neural network(CNN)was proposed. A factor recognition dataset for CNN model was constructed through literature research and used to train and debug the factor identification model. The model was validated with PM_(2.5) component observations from sampling sites in southern Beijing. The positive matrix factorization model(PMF)was used to analyze the source spectrum matrix for different number of factors,which was input into the factor identification model and compared with the manual analysis. The comparison results showed that the model achieved the best identification results at 9 factors,and could achieve neither duplicate identification nor "undetermined" cases,which matched with the manual identification results of source apportionment factors,and proved the rationality and feasibility of the proposed method. The method is of practical value to the problem of factor identification in source apportionment,and also has positive significance to the formulation and dynamic adjustment of emission reduction strategies.

关 键 词：源解析 正定矩阵分解(PMF) 卷积神经网络(CNN) PM_(2.5) 

分 类 号：X16[环境科学与工程—环境科学] X32