Research Of Prediction Model On Atmospheric PM_2.5 Concentration Using Support Vector Regression

Recently,many cities in China are suffering severe air quality problems.Haze weather occurs frequently,which affects people's lives,physical health,and ecological environment seriously.And PM_2.5-based atmospheric pollutant is one of the key factors that aggravate urban air pollution.Therefore,mastering the change regulations,pollution levels and related influencing factors for PM_2.5.5 concentration,and exploring suitable atmospheric PM_2.5 concentration prediction model have important scientific significance for controlling urban air pollution.Moreover,it can provide some basis and help for public travel arrangements and related decisions of government departments.Therefore,the Wanliu monitoring station in Haidian District of Beijing is selected as the research object,and some mainly atmospheric pollutants concentration and related meteorological factors are collected from March 1,2014 to April 30,2015.Then,the change regulations,related influencing element of atmospheric PM_2.5concentration are studied in different seasons and different hours by using mathematical statistics methods.At the same time,Machine learning methods are applied to predict the hourly and daily mass concentration of atmospheric PM_2.5.The main research contents of this article include:1?Mathematical statistics analysis method is used to analyze the change regulations and relevant influencing factors of atmospheric PM_2.5 concentration at the current site.The results of the study indicates that there is a certain seasonal change and daily change rule for the atmospheric PM_2.5 concentration,and the atmospheric PM_2.5exhibits more serious pollution in autumn and winter.At the same time,PM_2.5.5 is generally characterized by nighttime pollution over daytime.On the other hand,the correlation analysis results show that the atmospheric PM_2.5 concentration present a weak negative correlation with temperature and wind speed,while there is a weak positive correlation with the atmospheric pressure and relative humidity.At the same time,PM_2.5 presents significant positive correlations with CO,NO₂ and SO₂ in the four different seasons,while the correlation with O₃ in the four seasons is not obvious,but only showed significant negative correlation in winter.2?Through the support vector regression?SVR?prediction model for PM_2.5concentration based on grid search,the direct multi-step prediction method and the iterative multi-step prediction method are compared.The results show that with the increase of the prediction step,the prediction error of the two prediction methods will increase,and the iterative multi-step prediction method will increase the prediction error sharply due to the error accumulation effect.Therefore,the direct multi-step prediction method was selected as the main method for PM_2.5 concentration prediction.3?Because the three parameters?penalty factor,insensitivity coefficient and kernel parameter?in the SVR model have great influence on the performance of the model,the Quantum Particle Swarm Optimization?QPSO?algorithm is introduced to select the three parameters in order to improve the prediction accuracy.Therefore,the QPSO-SVR model is built to realize the prediction of atmospheric PM_2.5 concentration in the next 4 hours.Compared with Particle Swarm Optimization?PSO?-SVR model,Genetic Algorithm?GA?-SVR model and Grid Search?GS?-SVR model,the results showed that the proposed QPSO-SVR model outperforms the other three models in terms of accuracy,speed and robustness.4?Taking into account the non-linear and non-stationary characteristics of atmospheric PM_2.5 concentration time series,the“decomposition and integration”prediction method was introduced and the mixed CEEMD-QPSO-SVR forecasting model was proposed to realize the daily average concentration of atmospheric PM_2.5 on the second day.The results show that compared to Empirical Mode Decomposition?EMD?and Ensemble Empirical Mode Decomposition?EEMD?,Complementary Ensemble Empirical Mode Decomposition?CEEMD?possesses better effect on the time series decomposition of atmospheric PM_2.5 concentration,and the“decomposition and integration”prediction method is more accurate than the single prediction method.