Abstract
The aim of this study was to establish a method for predicting heavy metal concentrations in PM(1) (aerosol particles with an aerodynamic diameter ≤ 1.0 μm) based on back propagation artificial neural network (BP-ANN) and support vector machine (SVM) methods. The annual average PM(1) concentration was 26.31 μg/m(3) (range: 7.00-73.40 μg/m(3)). The concentrations of most metals were higher in winter and lower in autumn and summer. Mn and Ni had the highest noncarcinogenic risk, and Cr the highest carcinogenic risk. The hazard index was below safe limit, and the integrated carcinogenic risk was less than precautionary value. There were no obvious differences in the simulation performances of BP-ANN and SVM models. However, in both models many elements had better simulation effects when input variables were atmospheric pollutants (SO(2), NO(2), CO, O(3) and PM(2.5)) rather than PM(1) and meteorological factors (temperature, relative humidity, atmospheric pressure and wind speed). Models performed better for Pb, Tl and Zn, as evidenced by training R and test R values consistently >0.85, whereas their performances for Ti and V were relatively poor. Predicted results by the fully trained models showed atmospheric heavy metal pollution was heavier in December and January and lighter in August and July of 2019. For the period covering the COVID-19 outbreak in China, from January to March 2020, most of the predicted element concentrations were lower than in 2018 and 2019, and the concentrations of nearly all metals were lowest during the nationwide implementation of countermeasures taken against the pandemic.