LAUSR

Peroxyacyl nitrates (PANs) are important secondary pollutants in ground-level atmosphere. Accurate prediction of atmospheric pollutant concentrations is crucial to guide effective precautions for before and during specific pollution events. In this study, four models based on the back-propagation (BP) artificial neural network (ANN) and multiple linear regression (MLR) methods were used to predict the hourly average PAN concentrations at Peking University, Beijing, in 2014. The model inputs were atmospheric pollutant data and meteorological parameters. Model 3 using a BP-ANN based on the original variables achieved the best prediction results among the four models, with a correlation coefficient (R) of 0.7089, mean bias error of -0.0043 ppb, mean absolute error of 0.4836 ppb, root mean squared error of 0.5320 ppb, and Willmott's index of agreement of 0.8214. Based on a comparison of the performance indices of the MLR and BP-ANN models, we concluded that the BP-ANN model was able to capture the highly non-linear relationships between PAN concentration and the conventional atmospheric pollutant and meteorological parameters, providing more accurate results than the traditional MLR models did, with a markedly higher goodness of R. The selected meteorological and atmospheric pollutant parameters described a sufficient amount of PAN variation, and thus provided satisfactory prediction results. More specifically, the BP-ANN model performed very well for capturing the variation pattern when PAN concentrations were low. The findings of this study address some of the existing knowledge gaps in this research field and provide a theoretical basis for future regional air pollution control.