Air quality warning system based on a localized PM2.5 soft sensor using a novel approach of Bayesian regularized neural network via forward feature selection

Abstract

It is highly significant to develop efficient soft sensors to estimate the concentration of hazardous pollutants in a region to maintain environmental safety. In this paper, an air quality warning system based on a robust PM2.5 soft sensor and support vector machine (SVM) classifier is reported. The soft sensor for the estimation of PM2.5 concentration is proposed using a novel approach of Bayesian regularized neural network (BRNN) via forward feature selection (FFS). Zuoying district of Taiwan is selected as the region of study for implementation of the estimation system because of the high pollution in the region. Descriptive statistics of various pollutants in Zuoying district is computed as part of the study. Moreover, seasonal variation of particulate matter (PM) concentration is analyzed to evaluate the impact of various seasons on the increased levels of PM in the region. To investigate the linear dependence of concentration of different pollutants to the concentration of PM2.5, Pearson correlation coefficient, Kendall's tau coefficient, and Spearman coefficient are computed. To achieve high performance for the PM2.5 estimation, selection of appropriate forward features from the input variables is carried out using FFS technique and Bayesian regularization is incorporated to the neural network system to avoid the overfitting problem. The comparative evaluation of performance of BRNN/FFS estimation system with various other methods shows that our proposed estimation system has the lowest mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE). Moreover, the coefficient of determination (R-squared) is around 0.95 for the proposed estimation method, which denotes a good fit. Evaluation of the SVM classifier showed good performance indicating that the proposed air quality warning system is efficient.