The significance of periodic parameters for ANN modeling of daily SO2 and NOx concentrations: A case study of Belgrade, Serbia

Abstract

In recent decades, artificial neural networks (ANNs) have been used for the prediction of concentration of air pollutants in urban areas. Beside meteorological variables, periodic parameters, such as hour of the day or month of the year, have been frequently used to improve the performance of ANN models by representing variations of emission sources. In this paper, different forms of periodic parameters, i.e. smoothed cosines based approximation and normalized historical mean values, were combined with meteorological variables in order to analyze the sensitivity of the ANN model to them. Ward neural network and general regression neural network were used and compared for the prediction of daily average concentrations of SO2 and NOx in Belgrade, Serbia. Multiple performance metrics have demonstrated that models based on periodic parameters outperform the corresponding models that used only meteorological variables as inputs. Also, a newly proposed normalized historical mean MOYnmv (month of the year) proved to be more appropriate in majority of cases than the traditional cosines based approximation (MOYcos). A simple rule for the selection of the most efficient MOY form was defined depending on their mutual correlation (r). Results have shown that if MOYnmv is correlated with MOYcos with r > 0.8, then ANN models what uses MOYnmv provide more accurate predictions.