Modeling of effects of meteorological parameters on atmospheric concentrations of volatile organic compounds using artificial neural networks

Atmospheric volatile organic compounds (VOCs) has been attracting gradually increasing interest of researchers due to the increase in their ambient concentrations. VOC concentrations has been gradually increasing as a result of urbanization and industrialization. It has been shown by previous research studies that these VOCs pose adverse effects on human health and welfare. They have carcinogenic and mutagenic effects on living beings for long durations of exposure. Besides, they participate in a number of atmospheric photochemical reactions resulting in the formation of secondary pollutants, the most important of which is tropospheric ozone. A great number of research papers have been dedicated to ambient VOCs in recent years. These studies focused on assessing ambient levels and determining sources of VOC species. Unfortunately, current literature lacks research studies related with neural modeling of ambient VOC concentrations. This dissertation presents the results of a modeling study in which effects of meteorological factors on ambient levels of VOCs were investigated through artificial neural network (ANN) approach. Thirty-three ANN topologies were constructed for thrity-three VOC species. The network topologies consisted of an input layer, a hidden layer, and an output layer. The input parameters were ambient temperature, relative humidity, wind speed, wind direction, and day-hour of sampling. A multilayer perceptron, feed-forward artificial neural network (ANN-MLP) model was applied to all 207 measurement results for thrity-three VOC species. Trial-and-error schemes showed that 10 to 20 neurons in the hidden layer produces best results. Sigmoid function was used for all neurons in the hidden layer, while the activation function was one of hyperbolic tangent or purelin functions. Levenberg-Marquart algorithm was applied as the learning algorithm. The best network topologies were determined through a great number of trials. The correlation coefficients for thirty-three species ranged from 0.547 to 0.818, which indicates ANN successfully predicts ambient concentrations of a number of VOCs. The best correlation coefficients were obtained for 2,2,4-trimethylpentane (R:0.81), benzene (R: 0.74), 2-methylpentane (R: 0.78), and hexane (R: 0.70), while the correlation coefficients for undecane (R: 0.58), m&p-xylene (R: 0.595), pentane (R: 0.47), o-xylene (R: 0.56), and 3-methylpentane (R: 0.54) were the lowest ones. Sensitivity analyses were also performed for each VOC species. The results showed that ambient temperature, relative humidity, and wind direction are the most effective input parameters. Ambient VOC concentrations were inversely proportional to the relative humidity, that?s, the concentration decreases with increasing humidity. In general, an increase in ambient temperature lead to a decrease in VOC concentration. Since the measurements were taken within Davutpaşa Campus of Yildiz Technical University and the campus is surrounded by a great number of industrial facilities, highways and connection roads with high traffic loads, and the Central Bus Station of Istanbul, it is somewhat difficult to efficiently predict the VOC concentrations in such a complex airshed, and the results are considered as a success.