Application of artificial neural networks to predict prevalence of building-related symptoms in office buildings

Abstract

Artificial neural networks (ANN) were constructed to predict prevalence of building-related symptoms (BRS) of office building occupants. Six indoor air pollutants and four indoor comfort variables were used as input variables to the networks. A symptom metric was used as the measure of BRS prevalence, and employed as the output variable. Pollutant concentration, comfort variable, and occupant symptom data were obtained from the Building Assessment and Survey Evaluation study conducted by the US Environmental Protection Agency, in which all were measured concurrently. Feed-forward networks that employ back-propagation algorithm with momentum term and variable learning rate were used in ANN modeling. Root mean square error and R2 value of the simple linear regression between observed and predicted output were used as performance measures. Among the constructed networks, the best prediction performance was observed in a one-hidden-layered network with an R2 value of 0.56 for the test set. All constructed networks except one showed a better performance than the multiple linear regression analysis.