Adaptive-network-based fuzzy inference system (ANFIS) modelbased prediction of the surface ozone concentration

Abstract

This paper presents the results of modeling the tropospheric concentration of ozone in dependence on volatile organic compounds - VOCs (benzene, toluene, m- and p-xylene, o-xylene and ethylbenzene) and inorganic compounds - NOx (NO and NO2) CO, H2S, SO2, and particulate matter (PM10) in the ambient air, in parallel with meteorological parameters, i.e., temperature, solar radiation, relative humidity, and wind speed and direction. The modeling was based on measured results obtained during the year 2009. The measurements were performed at the measuring station located within an agricultural area, near the city of Zrenjanin (Serbian Banat, Serbia). Statistical analysis of obtained data, based on bivariate correlation analysis, indicated that accurate modeling could not be performed using the linear statistics approach. Moreover, considering that almost all the input variables have wide ranges of relative change (ratio of variance compared to range), the nonlinear statistic analysis method based on only one rule describing the behavior of the input variable most certainly would not present sufficiently accurate results. For these reason, the employed modeling approach was based on Adaptive-Network-Based Fuzzy Inference System (ANFIS). The model obtained using the ANFIS methodology resulted in high accuracy, with a prediction potential of above 80%, considering that obtained determination coefficient for the final model was R2 = = 0.802.