Comparative Analysis of different Statistical Methods for Prediction of PM2.5 and PM10 Concentrations in Advance for Several Hours

摘要

Atmospheric particulate matter (APM) is harmful for living being due to their small size which is ranging from ultra-fine particles up to particles with aerodynamic diameter up to 10 micrometers and hence because of their ability to penetrate deeper into human respiratory system. Particulates less than 2.5 micrometer (PM2.5) are more hazardous as compared to coarse particles of size 10 micrometer (PM10). The damage due to APM can be minimized through appropriate preventive measures. In order to gage the sway of air on the health and welfare of every living being it is necessary to perform an analysis of air quality for accurate decisions about preventive measures. Different machine learning methods including Support Vector Machines, Decision Trees, Neural Networks and Linear Discriminant analysis have been proposed for robust forecasting and prediction. This work is aimed at analyzing and benchmarking different methods for the prediction of average PM2.5 concentrations. For this purpose, data was acquired during the hours of the day from the ambient air and indoor environment in the suburb of Muzaffarabad (Azad Jammu and Kashmir, Pakistan). Linear and Radial Support Vector Regressors and RF algorithm were used for model generation and prediction. Results from these methods are then compared using root mean squared error (RMSE) for accurate predictions. The finding indicated that RF method and Radial Support Vector Regressor provided better prediction with RMSE as compared to Linear Support Vector Regressor.