Tez özetleri Astronomi ve Uzay Bilimleri Anabilim Dalı

Feature Selectıon by Data Mınıng Methods for Predıctıon of Abrupt

Changes ın Tıme Serıes
In time series analysis, generally, a variable is modelled by taking into consideration its previous values and previous/current values of other variables. In this study, we consider not only modelling of a variable as a function of time but also classification of abrupt changes in the values of a variable into two categories: abrupt rise and abrupt fall. As an application of the study, we have picked an air pollution dataset due to its gradually increasing importance. As a result, in this study, besides forecasting of the level of ozone as one of the air pollutants, we have also worked on determining factors that may cause abrupt changes in ozone level.Contrary to the positive effects of high level ozone concentration in stratosphere for protecting the Earth against ultraviolet radiation, in lower troposphere it has negative effects on human health and environment. Exposure to high level ozone concentration even for two-three hours can cause serious damage in respiratory systems of children and asthma patients. The goal of this study is to determine the feature groups that are related to abrupt changes in the level of ozone. Canonical Correlation Analysis (CCA), Support Vector Machines (SVM), Linear Discriminant Analysis (LDA) and k-Nearest Neighbour (KNN) techniques are used to explore which combination of features are predictive of abrupt changes in ozone level. The simulation dataset used in this study is collected in Ankara, Turkey, by an automatic air quality monitoring station operated by the ministry of environment and urban planning. The dataset consisted of one year of measurements of air pollutants and the meteorological factors. The analysis of ozone time series has shown that NO, NO₂, NO_X and wind speed are effective variables for forecasting the future ozone levels. On the other hand, temperature and relative humidity are more effective variables for classification of whether an abrupt rise or fall will occur in the level of ozone. Furthermore, particulate matters and SO₂ are found to be the most effective for rise/fall classification when considering even more abrupt changes in ozone levels.