Forecast Of Air Pollutant Concentrations On BP Neural Network

In recent years,air pollution has become a serious problem.Air quality deterioration has a huge potential hazards on personal health and the environment.Atmospheric pollutants concentration forecast is very important,therefore,it will not only help in People's Daily lives,and s has a guiding significance for the government to make relevant policy.In 2013,the state council promulgated.The action plan for the control of air pollution.The action order local government to set up monitoring and warning system.Beijing-tianjin-hebei,Yangtze river delta,the pearl river delta and other provinces,deputy provincial city,the provincial capital cities are included in the city or region were ordered to carry out air quality forecast warning work.Study of pollutant concentration model in Kunming can make contributes to air quality forecast warning work of Kunming.Represented by statistical model and machine learning model of the mechanism model are widely used in pollutant concentration forecast.The BP neural network,with its strong ability of nonlinear fitting is widely used in the concentration prediction of pollutant.The models of the prediction of SO2,NO2,03,CO,PM10 and PM2.5 concentrations were established by back propagation(BP)neural network and variable selection in this paper,which were monitored the concentrations of above pollutants and founded daily concentrations forecasting models from five monitoring points of Kunming from 2014-1-1 to 2015-11-28.Furthermore,the mean impact value(MIV)method was used to detect the major factors of daily concentrations of pollutants above and the results were regarded as input variables of BP to forecast daily concentration changes of different pollutants respectively.There are some conclusion blow.Through variable screening results as,concentration of other pollutants the day before forecasting day have a great influence on the forecast object.BP neural network model of prediction results are good,and prediction results coincided well with the observations.Standardization of standardized mean deviation(NMB)are less than 18 and the average error(NMB)are less than 40.Residual standard deviation RMSE are less than 30 and most of correlation coefficient R are much greater than 0.6.Using the MIV method for input variables selection can help to improve prediction accuracy of BP neural network models.Some individual prediction model such as NO2,CO of guansahng position,SO2 of bijiguangchang position,S02 of longquanzhen,S02 of chenggongxinqu position,S02,03 of dongfengdonglu position can not improve the prediction accuracy.Various pollutants IAQI accuracy is higher which are bigger than 70%,and the primary pollutants accuracy can reach 50%.The accuracy of AQI are bigger than 65%.