Predicting tropospheric ozone concentrations in different temporal scales by using multilayer perceptron models

This study encompasses ozone modeling in the lower atmosphere. It was aimed to develop an appropriate neural network model in order to predict ozone concentrations in various temporal scales as a function of meteorological variables and air quality parameters. All data were collected from Dilovasi, Turkey as this site represents typical industrial regions with major air pollution problems. In the study performance of the multilayer perceptron models were tested for both annual and seasonal periods as meteorological conditions highly influence the ozone levels. Among the various architectures, a network of two hidden layers with fifteen neurons was found to give successful predictions. Modeling efficiency of the developed network was also evaluated for day light and night time data of warming season exhibiting highest ozone levels. Furthermore, principle component analysis was performed by using annual data in order to reduce the number of input variables describing ozone formation. Model run with principle components has also provided satisfying performance.