Machine Learning Approaches for the Estimation of Particulate Matter (PM2.5) Concentration Levels: A Case Study in the Hyderabad City, India

摘要

Particulate matter concentration is one among several variables monitored at regular intervals to calculate air quality indices (AQI) which are intended to help understand the acute and chronic effects of air quality on human health. The tine particulate (PM2.5) samplers installed at pollution monitoring stations continuously monitor the concentration of pollutant in air over time. The specific time-averaged concentration is then estimated from the continuous records. Missing data records in the PM2.5 time series is quite normal, which is attributed by faulty equipment, routine maintenance schedules, or replacement of equipment. When one or more point observations in a time series are missing, it is very essential to estimate or predict the missing values. This study presents the application of machine learning techniques such as support vector regression (SVR), group method of data handling (GMDH) network, and evolutionary adaptive neuro fuzzy inference system to estimate the 24-h average PM2.5 concentration levels at a particular station using PM2.5 concentration levels observed at neighborhood stations as inputs. The performance of these models are evaluated in terms of widely used statistical metrics such as centered root mean square difference (CRMSD), normalized Nash-Sutcliffe efficiency (NNSE), and correlation coefficient (R). The findings of the study reveal that the GMDH model provided reasonably accurate estimates of daily PM2.5 levels.