Performance investigation of MISO soft sensors in predicting AQI: a comparative analysis

Abstract

ABSTRACT The present investigation proposes a Bayesian Regularized Artificial Neural Network (BRANN) based multi input single output (MISO) soft sensor. It predicts the value of air quality index (AQI) by receiving measured inlet concentration of different air pollutants such as particulate matter-2.5 (PM2.5), particulate matter-10 (PM10), SO2, NO2, O3, and CO. The performance of BRANN is also evaluated comparing with other two soft sensors namely, scaled conjugate gradient trained artificial neural network (SCGANN) and Levenberg-Marquardt artificial neural network (LMANN). The Bayesian Regularization (trainbr), trainscg, and Levenberg-Marquardt (LM) feed-forward backpropagation algorithms are used to train BRANN, SCGANN and LMANN, respectively. Hence, for their training and validation, two sets of 24 hours average of real time data released by the Central Pollution Control Board (CPCB) of India for Talcher coalfield and its surrounding area in Odisha from Jan-2019 to Dec-2020 are used. The training performances of all three soft sensors are reasonable (average R2 values of 0.975452, 0.995533 and 0.998718 for SCGANN, LMANN and BRANN, respectively). Obviously, the LMANN exhibits its superiority over SCGANN (corresponding average R2 values of 0.811955 and 0.760921). Here, the BRANN overperforms (average R2 values of 0.891167) other two soft sensors in predicting the AQI.