一种基于小波变换和LSTM的大气污染物浓度预测模型  

AN AIR POLLUTANT CONCENTRATION PREDICTION MODEL BASED ON WAVELET TRANSFORM AND LSTM

作　　者：何哲祥 李雷 HE Zhe-xiang;LI Lei(Department of Environment Engineering,School of Metallurgy and Environment,Central South University,Changsha 410083,China)

机构地区：[1]中南大学冶金与环境学院环境工程研究所,长沙410083

出　　处：《环境工程》2021年第3期111-119,共9页Environmental Engineering

基　　金：国家“十二五”科技支撑计划项目(2012BAC09B02);中南大学中央高校基本科研业务费专项资金(2019zzts513)。

摘　　要：针对现有大气污染物浓度预测模型存在预测精度不高、污染物种类单一等不足的问题,通过小波分解将高维大气污染物数据转换为低维数据,再对分解序列建立长短期记忆网络(LSTM)预测模型,最后通过小波重构将分解序列重构为污染物时间序列,建立了1种基于小波变换(WT)的LSTM大气污染物预测模型(WT-LSTM),用以预测目标区域内的次日平均ρ(PM_(2.5))、ρ(PM_(10))、ρ(SO_(2))、ρ(NO_(2))和ρ(O_(3))。采用长沙市2015—2018年10处国控站点的数据进行验证,结果表明:相对于LSTM、多元线性回归(MLR)和基于WT的WT-MLR模型,WT-LSTM的均方根误差和绝对平均误差均下降了50%,其对PM_(2.5)、PM_(10)、SO_(2)、NO_(2)和O_(3)的污染等级预测准确率均在80%以上。To solve the problem that current atmospheric pollutant prediction research has low accuracy of prediction and only pays attention to single pollutant type,a long short-term memory network atmospheric pollutant prediction model based on wavelet transform was proposed,to predict daily average PM_(2.5),PM_(10),SO_(2),NO_(2) and O_(3) concentration of the next day.First,the high-dimensional data was converted into low-dimensional data by wavelet decomposition,and subsequently,the long short-term memory network prediction model was established for low-dimensional data.Finally,the decomposition sequence was reconstructed into the pollutant time series by wavelet reconstruction.Based on the data collected from 10 national control stations in Changsha from 2015 to 2018,the model was verified.The results showed that for the prediction of atmospheric pollutants of the next day,compared with the LSTM,MLR,WT-MLR,the root mean square error and absolute mean error of WT-LSTM model decreased by 50%,and the accuracy of the pollution level predictions of the five air pollutants were all above 80%.

关 键 词：空气质量指数 机器学习 小波变换 大气污染物预测 

分 类 号：X831[环境科学与工程—环境工程] TP183[自动化与计算机技术—控制理论与控制工程]