Performance Evaluation And Model Calibration Of Co Mmercially Available Light-scattering PM_2.5Monitors

Objective:This study was aimed to evaluate the performance of light-scattering PM2.5 monitors,to explore the influencing factors of light-scattering method in PM2.5 measurement,and to develop calibration models using the machine learning algorithms.Methods:From December 2020 to March 2021,the performance of two professional-grades(MicroPEM and LD-6S)and three low-cost co mmercially available PM2.5 monitors(A,B,and C)was evaluated through side-by-side comparison with the tapered element oscillating microbalance(TEOM)as a reference instrument.The precision of the monitor is evaluated by calculating the correlation coefficient and deviation value between the results from two identical devices.The accuracy of the monitor is evaluated by calculating the correlation coefficient and deviation value between the results of the light scattering monitors and the TEOM.The influencing factors of the light scattering method are explored by comparing the ratios of the concentrations determined by the light scattering method and TEOM among different groups divided by quartiles of the relative humidity,temperature and PM2.5 concentrations.The multiple linear regression,random forest,support vector machine and artificial neural network algorithms were used to develop calibration models for light scattering method,when the PM2.5 concentrations determined by TEOM method was used as the dependent variable,and the relative humidity,temperature and PM2.5 concentrations were used as prediction variables.The performance of the models was evaluated by 10-fold cross validation method(10-fold CV).The coefficient of determination(R2),root mean square error(RMSE)and average absolute error(MAE)were used as model performance parameters.Results:(1)Precision:The two professional-level light-scattering monitors both demonstrated a high degree of precision.The correlation coefficient values(r)were higher than 0.98(rMicroPEM=0.982;rLD-6S=0.995),and the average deviations were(0.2±16.4)μg/m3 and(2.0±7.8)μg/m3 for MicroPEM and LD-6S,respectively.The precision varied largely among the three low-cost monitors with r values ranged from 0.827 to 0.990,and the average deviation ranged from 2.3 μg/m3 to 28.6 μg/m3.(2)Accuracy:The results of the two professional-level light-scattering monitors were both highly related to the results of TEOM,and the r values were 0.943 and 0.972 for MicroPEM and LD-6S,respectively.However,there was systematic deviation between the results of these two monitors and TEOM,with the median of ratios were 1.42 and 0.97 for MicroPEM and LD-6S,respectively.The correlation coefficients between the results from the three low-cost monitors and TEOM varied from 0.813 to 0.942.All of the three low-cost monitors overestimated the PM2.5 concentrations,with the median of ratios ranged from 1.32 to 1.43.(3)Influencing factors of light-scattering method:The environmental temperature,relative humidity and PM2.5 pollution levels were found to be the important influencing factors of light-scattering method.The ratio values of the results from the light-scattering and the TEOM method increased with the increasement of the environmental relative humidity as well as the PM2.5 concentrations for all tested monitors.However,the effect of environmental temperature was inconsistently among different light-scatting monitors.(4)Calibration model based on machine learning algorithms:With the ambient temperature,relative humidity and PM2.5 measured by the light-scattering method as predictive variables,all the calibration models based on traditional multiple regression and more sophisticated machine learning algorithms(random forest model,support vector machine model and artificial neural network model)could improve the accuracy of light-scattering method,and the models based on the later algorithms performed better than that based on the former algorithm,which was indicated by higher R2 values and lower RMSE and MAE value in the 10-fold cross validation.Support vector machine model and artificial neural network model performed best for 4 and 1 light-scattering PM2.5 monitors,respectively.Conclusions:In general,the co mmercial professional light-scattering PM2.5 performed good in precision,and their results were highly related to the reference method.However,the performance of the co mmercial low-cost light-scattering PM2.5 monitors varied among monitors,but there was no clear correlation between their performance and price.There was systematic deviation between the results of the light-scattering method and the reference method for all tested monitors,which needed to be calibrated.The relative humidity,temperature and PM2.5 pollution level were important influencing factors for light-scatting method,and the accuracy of light-scattering method can be improved by calibrations model including these influencing factors.The models based on more sophisticated machine learning algorithm performed better than the traditional multiple linear regression model.