Comparing Gray Box Methods to Derive Building Properties from Smart Thermostat Data

The development of quantitative techniques for determining the amount of heat lost through the building envelope is essential for targeted retrofits. This type of evaluation is traditionally a resource intensive process that involves onsite appraisal and in-situ measurements. In order to build more efficient and scalable methods for retrofit analysis, new sources of data could be used. Smart thermostat data, for example, provide a valuable resource, however they often lack detailed information about the building characteristics and energy loads. This paper presents and compares three methods for assessing heating characteristics of households using a dataset that does not contain heating power. The three methods are based on: (1) balance point plots, (2) the extraction of indoor temperature decay curves, and (3) the classic differential equation for indoor temperature. These methods all take a gray box approach in which physics-based and machine learning models are combined. The dataset used for this study consists of over 4,000 houses in Ontario and New York. The three methods are applied to each building and the resulting data is analyzed to determine whether the results are statistically sound. It is found that there is a positive linear correlation between characteristics derived for each method, although there is uncertainty about absolute values. This result indicates that the methods can be used to ascertain relative values for the thermal characteristics of a building. The methods suggested in this paper may therefore be used to filter heating profiles to target potential retrofit measures or other stock-level decisions.