Evaluating the Feasibility of Reusing Pre-trained Thermal Models in the Residential Sector

Abstract

Modelling temperature dynamics of a building is necessary to develop control mechanisms for reducing energy consumption of heating and cooling equipment. While Resistance-Capacitance (RC) models can accurately explain how the indoor temperature changes over time, building such models requires the knowledge of the building insulation and thermal mass, which is not readily available for most residential buildings in operation today. In the absence of this information, model parameters can be estimated from coarsegrained data collected by smart thermostats. In this paper we train a Bayesian neural network to establish the RC model for a home equipped with a smart thermostat, and investigate how to reuse this model to predict the temperature inside another home which may not be equipped with a smart thermostat. Leveraging data from ecobee smart thermostats installed in over 4,000 homes in Canada, we validate that a small number of pre-trained neural network models is enough to develop a sufficiently accurate RC model for any home across the country and that this model outperforms a seasonal time-series model that is built using the same amount of data.