Investigation of machine learning methods for prediction of ozone concentration to determine outdoor air quality level

Due to an increase in the field of industrialization and urbanization that the world witnessed in recent years, the amount of air pollution is increased in various countries. This situation constitutes a risk for the humans especially those with respiratory diseases. Therefore, prediction of the air pollution level is an important task that may be useful for developing early warning systems. In this thesis study, performances of six different machine learning methods have been investigated for predicting outdoor ozone concentration in Istanbul, Turkey. These methods are support vector machines, multilayer perception, random forests, gradient boosted decision trees, k-nearest neighbor, and elastic net. The predictions are made according to one-step-ahead time-series forecasting scheme with a rolling walk forward validation. Six major atmosphere components that are based on air quality index readings, namely, SO2, PM10, O3, PM2.5, NO2, and CO that are utilized to train the models. Firstly, prediction performance of these methods are calculated by training regression models. Next, the predicted values are converted into one of the five air quality index levels. This conversion makes it possible to calculate performance metrics regarding a multi-class classification problem. The lowest RMSE of 2246 was observed by using MLP and the highest accuracy score of 0.790 was observed by SVM.