Research Of PM_2.5 Concentration Variation Rule And Prediction Model

In recent years,China's air pollution problem is very serious,especially in the Beijing-Tianjin-Hebei region,where long-term,continuous haze weather often occurs.PM_2.5,as one of the major factor aggravating haze,not only affects people's daily life,but also poses a serious threat to people's health.Therefore,mastering the changing rules and influencing factors of PM_2.5.5 concentration and exploring an effective PM_2.5concentration prediction model have important scientific significance.In this paper,the changing rules and related influencing factors of PM_2.5.5 concentration are studied firstly.Then two different prediction models are established according to different sample sets,and the hourly prediction of PM_2.5.5 concentration is conducted.The main research contents of this paper are as follows:1?Based on the data of PM_2.5.5 concentration at 12 monitoring stations in six districts of Beijing,the annual variety regulation of PM_2.5.5 concentration from December 2013 to November 2018 is studied.The results show that the PM_2.5concentration has similar changes in the past five years,with obvious seasonal variation characteristics,and presents a trend of decreasing year by year.Meanwhile,PM_2.5.5 concentration data from December 2014 to November 2015 in six urban areas are selected to study its seasonal and diurnal variation characteristics.The results show that the pollution level of PM_2.5.5 is more serious in autumn and winter;At the same time,in spring and summer,the pollution level in daytime is higher than that in nighttime;In autumn,the pollution level is similar at different times of the day;In winter,the pollution level of PM_2.5.5 in nighttime is higher than that in daytime.2?Other pollutant concentration data(PM₁₀,SO₂,NO₂,CO and O₃concentrations)and meteorological data?temperature,air pressure,relative humidity and wind speed?from the monitoring station of Beijing Chaoyang Agricultural Exhibition Hall from December 2014 to November 2015 are selected to study the correlation between PM_2.5.5 concentration and them.The results show that PM_2.5.5 has obvious positive correlations with PM₁₀,SO₂,NO₂ and CO in the four seasons,while there is a weak positive correlation with O₃ in summer,but a negative correlation in the other three seasons.At the same time,PM_2.5.5 shows a significant positive correlation with relative humidity,while the correlations with temperature,pressure and wind speed are not obvious.3?Based on the Least Square Support Vector Machine?LSSVM?model,this paper establishes the NDFA-LSSVM prediction model by optimizing its super parameters through the improved New Dynamic Firefly Algorithm?NDFA?.The monitoring data of Beijing Chaoyang Agricultural Exhibition Hall in December 2014and March 2015 are selected as the research objects to achieve the prediction of PM_2.5concentration in the next two hours.Compared with Grid Search?GS?-LSSVM model,Genetic Algorithm?GA?-LSSVM model and Firefly Algorithm?FA?-LSSVM model,the comparison results show that the prediction accuracy of NDFA-LSSVM model are better than other three models.In view of the non-linear and non-stationary nature of PM_2.5.5 concentration time series,this paper introduces the Complementary Ensemble Empirical Mode Decomposition?CEEMD?method to decompose the original PM_2.5concentration time series,and then establishes the CEEMD-NDFA-LSSVM prediction model.Compared with the prediction results of NDFA-LSSVM model,CEEMD-NDFA-LSSVM model shows better prediction effect.4?Considering the limitations of the least squares support vector machine in processing large data sample sets,this paper establishes the CEEMD-Long Short-Term Memory?LSTM?model based on the data from 0 o'clock on April 3,2014 to 23 o'clock on November 30,2018 at the monitoring site of Beijing Chaoyang Agricultural Exhibition Hall.Meanwhile,the prediction results of CEEMD-LSTM model are compared with the prediction results of LSTM,Empirical Mode Decomposition?EMD?-LSTM model and Ensemble Empirical Mode Decomposition?EEMD?-LSTM model.The results show that the decomposition of PM_2.5concentration time series can effectively improve the prediction effect of LSTM model.Meanwhile,among the three decomposition methods,CEEMD method has better decomposition effect on PM_2.5.5 concentration time series.Finally,CEEMD-LSTM model and CEEMD-NDFA-LSSVM model are compared in three sample sets.The results show that CEEMD-LSTM model achieves relatively good prediction effect on the large data sample set,while CEEMD-NDFA-LSSVM model has better prediction effect on the small data sample set.