PREDICTION OFAIR OUALITY INDEX BASED ONMETEOROLOGICAL VARLABLES USING MACHINELEARNING TECHNIQUES

摘要

It can be observed that the distribution of pop- ulation worldwide is becoming increasingly concen- trated in certain regions. The crealion of new metropoles through a decrease in rural population leads to many disadvantageous situations. Industri- alization caused by increasing population will bring about health problems in the short term and also problems resulting from global warming in the long term. The deterioration in air quality is among the most important such problem. In order to measure a complex phenomenon such as air quality and to make it easier for people to understand. each country applies certain air quality indexes in line with do- mestic legislation. For this study, with regard to esti- mations of Turkey's national air quality index (AOD based on Environmental Protection Agency (EPA) standards, data have been collected from a station in Ankara city for the 2010-2014 period. In addition to PMio, 03, SO2, NOz, CO pollutant concentra- tions, a dataset consisting of six different sets of me- teorological data has also been created. In the study, XGBoost (XGB), random forest regression (RFR) and extra trees involving tree-based ensemble learn- ing (TBEL) algorithms have been used to estimate the air quality index. In addition to these algo- rithms, analyses have been carried out with the sup-port vector regression (SVR), artificial neural net- works (ANN) and k-nearest neighbors regression (k- NNR). According to the test data set results, an aver- age R-Squared score of 0.97 was obtained with re- gard to the TBEL algorithms, while a score of 0.76 was obtained for the other algorithms. In addition, the TBEL algorithms showed a close prediction per-formance among themselves in terms of test scores, and a best R-Squared score of0.99 has been obtained with regard to the RFR algorithm.