Ground-level ozone predictions using outlier identification leveraged sample weighted regressors

Abstract

ABSTRACTGround-level ozone is a pollutant, greenhouse gas, and respiratory irritant which may facilitate skin cancer development and be involved in cardiovascular, respiratory and a range of other diseases. A re-distribution in the hourly ozone concentrations has occurred in the past decades while the interest in obtaining precise methods for the prediction of ozone measures has risen. Weather conditions influence ozone levels, specifically we used maximum temperature per hour, solar radiation per hour, date and hour of the measurement in order to fit prediction models. Weather stations may provide defective data with missing values or incorrect measures which may lead to a decrease in the performance of data driven predictors. This paper proposes a new method that deals with raw data without preprocessing by weighting the effect of automatically detected outliers. The method is evaluated against other traditional outlier removal techniques for a case study in Ponferrada, Spain. Our method yielded great performance for ground-level ozone prediction in simpler and more sophisticated regression techniques, such as linear regression and multi-layer perceptron algorithms.