Prediction of hourly indoor Carbon Monoxide concentrations in semiarid regions using Regression and feedforward backpropagation as a hybrid model

Abstract

Accurate site-specific forecasting of indoor hourly carbon monoxide (CO) concentrations in school microenvironments is a key issue in air quality research nowadays due to its impact on children’s health. This paper investigated the improvement prediction of multiple linear regression (MLR) and feed forward back propagation (FFBP) by combining them with principal component analysis (PCA) for predicting indoor CO concentration in Gaza Strip, Palestine. Measurements were carried in 12 schools from October 2012 to May 2013 (one academic year). The results suggested that the selected models are effective forecasting tools and hence can be applicable for short-term forecasting of indoor CO level. The predicted indoor CO concentration values agree strongly well with the measured data with high coefficients of determination (R2) 0.869, 0.870, 0.885 and 0.915 for MLR, PCA-MLR, FFBP and PCA-FFBP, respectively. Overall, results showed that PCA models combined with MLR and PCA with FFBP improved MLR and FFBP models of predicting indoor CO concentration, with reduced errors by as much as 7.14% (PCA-MLR) and 56.6% (PCA-FFBP). Moreover, PCA improved the accuracy of the FFBP model by as much as by 3.3%.