Short-Time Forecasting of Atmospheric NOx Concentration by Neural Networks

Abstract

The possibilities of modeling of NO and NO2 concentrations at air monitoring stations were analyzed in this paper. Artificial neural networks (ANN) were used in time series analysis to predict the levels of concentrations. Investigations were carried out based on two data sets collected at air monitoring stations in South Poland: Zabrze (1994–1997) and Ke˛dzierzyn-Koźle (1994–1999). The set of hourly values of NOx concentrations were used in the analysis. The main purpose of the studies was to estimate prediction accuracies for both chosen pollutant concentrations. The effect of changes of lookahead, the ANN time series parameter, was analyzed. The accuracies of different obtained models were compared. The main result was that the accuracy of prognosis quickly decreases when lookahead rises. Time series models of NOx concentration are not accurate, except for short lookaheads. The accuracy of NO2 concentration modeling is apparently better than NO concentration modeling. The nonlinear models are not distinctly more precise than their linear equivalents. The linear models seem to be more advantageous because of their simplicity and very short time of calculation. The precision of NOx modeling at the more polluted station (Zabrze) is higher than at the less polluted one ( Kędzierzyn-Koźle). Other methods of modeling could be more accurate in prediction of NOx levels, so possibilities of practical implementation of ANN time series analysis are limited.