Abstract: Since ozone near the earth’s surface has adverse health effects and is bad for vegetation, it can be said that it is very important to issue an ozone day warning in advance as early as possible so that people can take precautionary measures against the ozone day. In this paper, we define the problem of predicting ozone days up to 14 days in advance as a classification problem into ozone days and non-ozone days, whereas the ozone day prediction problem has been mainly treated as a regression problem in the past. To make a time series data the class values of the original data, which had information on ozone days or not by date, were moved to the number of predicted days from 1 day to 14 days to determine whether or not it might be an ozone day or not. They were moved as many days as the number of predicted days. We experimentally proved that such a problem can be solved using random forests and a novel massive oversampling technique for the time series data. We applied principal component analysis to eliminate possible multicollinearity between attributes because our target dataset has 72 attributes as conditional attributes, which can be considered quite a lot, and statistical analysis of a t-test was performed to confirm that the oversampled data belonged to the same population. The results of the experiment showed that by using statistically reliable oversampled data to train the random forests, it could be possible to generate random forests of ozone day classification for the time series data with much higher accuracy than using the original data alone so that very accurate ozone day prediction models of random forests could be generated for upto 14 days in advance. That is, an ozone day prediction accuracy of up to 99.84% was achieved when the model predicted ozone day or not seven or eight days in advance whereas the worst accuracy is 99.64% which is the accuracy of three days in advance.

Keywords: Ozone day prediction, time series analysis, classification, principal component analysis, oversampling, random forests

Pages: 474-490