Study Of Statistical Methods Based On Machine Learning Algorithms For Air Quality Forecasting In Lanzhou

Air pollutants(especially PM2.5)seriously endanger human health.Based on the analysis of the pollution characteristics of three major air pollutants SO2,NO2 and PM10 during 2001~2011,six major air pollutants PM10,PM2.5,NO2,SO2,CO and O3 during 2013~2015,we built several forecasting models for predicting daily average concentrations of 6 major pollutants for next 24 h and 48 h by using BP neural networks,least–square support vector machine(LS-SVM)and Elman neural networks with European medium-term weather forecasting center(ECMWF)and T639 operational forecasting product as predictive factors,with wavelet decomposition as a progressively pollution data processing procedure.Then we compared the predicted and measured concentrations to analyze the advantages and disadvantages of each modeling method.Finally,we combined different predictions from several models of good accuracy into each predicted daily concentration of 6 major air pollutants by using support vector regression method(SVR),we tested its performance by simulating real operational forecasting procedure.The results are as follow:(1)during 2013 ~ 2015 PM10 was still the major pollutant in Lanzhou and the main cause of heavy pollution in spring;during 2013-2015 annual average concentration of SO2 decreased significantly compared with that during 2001~2011,and from 2013 onwards it was lower than the annual average concentration of NO2 in the same period;the annual average concentration of O3 increased year by year,in 2015 the days for O3 to be the major pollutant increased significantly,becoming one of the most important pollutants in the summer.(2)The evaluation indexes of 24 h and 48 h forecast models of LS-SVM are better than those of BP neural network and Elman neural network.The stability and performance of the models established by BP neural network are relatively poorer,and the prediction accuracy of BP 48 h model showed the highest attenuation.(3)The 24 h and 48 h models built with ECMWF showed better performance in predicting daily concentrations of PM10,PM2.5,NO2,SO2 and CO;Models built with T639 showed some advantages in predicting O3 concentration against ECMWF.(4)The prediction accuracy of both 24 h and 48 h prediction models of LS-SVM is improved by pretreating the pollutant data by wavelet decomposition method.(5)The forecasting accuracy of the combined forecasting model for the daily average mass concentration of the six major pollutants is higher than that of the unaudited model.Compared the predicted 24 h and 48 h AQI of the combined forecasting model with the actual AQI,the mean absolute error for 24h(48h)is 9.874(12.315),the mean absolute percent error is 12.4%(15.1%),the root-mean-square error is 14.033(17.095);the detected rate of air quality index level is 76.7%(71.5%),the unpredicted rate is 9.1%(11.0%),the false alarm rate is 14.2%(17.5%);The detected rate of the type of major pollutant is76.3%(70.0%).The results of this study have a certain reference value to improve the accuracy of operational air quality forecasting for Lanzhou.