Machine Learning Enabled Large-Scale Estimation of Residential Wall Thermal Resistance from Exterior Thermal Imaging

Abstract

Traditional building energy audits are both expensive, in the range of USD$1.29/m²-$5.37/m², and inconsistent in their prediction of potential energy savings. Automation to reduce costs of evaluating the energy effectiveness of buildings is strongly needed. A key element of such automation is a means to estimate the building envelope energy effectiveness. We present a method that addresses this need by using infrared thermography to characterize building wall envelope effectiveness.  To date, thermal imaging approaches for estimating wall R-Values, based upon thermal-physical models of walls, require additional manual measurements and analysis which prohibit low-cost, large-scale implementation. To overcome this implementation challenge, a machine learning approach is used to predict wall R-Values for a set of residences with known thermal resistance by utilizing the measured wall imaging temperature, prior weather conditions, historical energy consumption data, and available building geometrical data. The developed model is shown to predict wall R-Values with a maximum test-set root mean squared error of 7% using as few as nine training houses. This result has significant implications for low-cost large-scale envelope energy effectiveness characterization.