Comparative analysis of maximum daily ozone levels in urban areas predicted by different statistical models

Many large urban areas experience elevated concentrations of ground-level ozone pollution, which is reported to cause adverse effects on human health and the environment. The prediction of ground-level ozone is an important topic, which attracts attention from research communities and policy makers. This study investigates the potential Of using the multi-layer perceptron (MLP) neural network technique to predict daily maximum ozone levels in the Bangkok urban area. The MLP was trained and validated using ambient air quality monitoring data and observed meteorological data for the high ozone months (January to April) in the area during a four year period, 2000-2003. The inputs to the MLP included the average concentration of air pollutants (nitrogen oxide, nitrogen dioxide, and non-methane hydrocarbon) and meteorological variables (wind speed and direction, relative humidity, temperature. and solar radiation) during the morning rush hours. The MLP network, which contained 8 input layer neurons, two hidden layers (10 hidden neurons for the first hidden layer, 14 hidden neurons for the second hidden layer) and 1 output layer neuron, was found to give satisfactory predictions for both the training and validated data sets. The performance of the MLP was better than the multivariable linear regression model developed based on the same dataset. For the validated dataset, the MLP predicted the daily maximum 1-h ozone concentration in the study area with a mean absolute error of 10.3 ppb, a root mean square error of 13.5 ppb, a coefficient of determination (R-2) of 0.85, and an index of agreement of 0.89.