Objective To establish a PM2.5 component concentration estimation model based on random forest algorithm with NO3- as an example, and to investigate the large influencing factors for NO3- concentration and the continuous time series characteristics of NO3- concentration.
Methods The study used the meteorological, land use, emission inventory, and air quality monitoring data of PM2.5, NO2, PM10, SO2 and CO between 2013 and 2017 as the independent variables and NO3- concentration data as the dependent variable, and various method such as value extraction to points, inverse distance weight interpolation, and setting of 1 km buffer area were used to standardize various data sets. A random forest model was established, and the fitting effect of the model was validated by the ten-fold crossover method.
Results The result of model verification showed that there was a high degree of fitting between the simulated value and the monitoring value, and daily, monthly, and annual mean concentrations had R2 of 0.61, 0.77, and 0.83, respectively. According to the importance ranking of the feature parameters of NO3- concentration in the model, the mass concentration of PM2.5 had the highest importance score of 0.387, and meteorological factors such as albedo lagging for 2 days, albedo lagging for 1 day, 10 m longitudinal wind speed and height of the boundary layer were closely associated with the change in NO3- concentration. In addition, the primary PM2.5 sources emitted by transportation, residential, industry, and power sectors all ranked among the 20 most important sources.
Conclusion The multi-parameter random forest model has certain advantages in PM2.5 composition simulation. Factors such as PM2.5 mass concentration, NO2, 10 m longitudinal wind speed, and primary PM2.5 sources emitted by residential and traffic sectors have a great influence on the simulation of NO3- concentration. NO3- concentration has the characteristics of seasonal distribution.
DownLoad: