Development and assessment of multiple regression and neural network models for prediction of respirable PM in the vicinity of a surface coal mine in India

摘要

The findings presented in this paper are based on the study conducted in a large opencast coal project in India. As a part of this study, particulate matter (PM) levels at different distances from the pit boundary were measured to understand dispersion of particles of different sizes emitted from the mine. Portable spectrometers and weather stations were deployed for measurement of PM concentration and meteorological parameters, respectively. The objectives were to quantify PM (PM10, PM2.5, and PM1) escaping from an active mine, estimate spatial variations of PM concentration in mine surrounding, evaluate role of meteorology, distance from the mine on PM concentration in the mining locality, and develop a model for prediction of PM10, PM2.5, and PM1 levels in the surroundings of a mine. A test of equality of means, correlation analysis, bivariate regression analysis, stepwise regression analysis, and the general linear model univariate procedure (analysis of covariance (ANCOVA)) were used for data analysis. At pit boundary, the average PM10, PM2.5, and PM1 concentrations were 2.75, 2.44, and 2.58 times the corresponding background level, and it reduced to 1.51, 1.79, and 1.90 times at a distance 500 m from the mine. Stepwise regression analysis revealed that between temperature and RH, which are highly correlated, RH is a better predictor of PM concentration. Although the proposed models had moderate R-2 values (AdjR(2): PM10 =0.35; PM2.5 =0.43; PM1 =0.42), the models were reasonably good in predicting PM concentration (index of agreement =0.80-0.89), with better prediction for fine particles (R-2 =0.48 for PM10, 0.62 for PM2.5, and 0.66 for PM1). The general linear model results revealed that distance is the largest predictor (16%) for PM10, but RH explains the highest variability for PM2.5 (28.2%) and PM1 (31.4%) concentrations. Wind speed was the least powerful determinant (4.5-9.9%). Using meteorological parameters (RH, temperature, and wind speed) and distance as input neurons, a feed-forward back-propagation artificial neural network model has been developed for prediction of particulate matter concentration.