Thermal modeling for control applications of 60,000 homes in North America using smart thermostat data

Abstract

As smart thermostats become increasingly available in residential buildings, there is an opportunity to use measured building data to calibrate models for community and district applications, instead of relying on high-fidelity simulations. This study used smart thermostat data from 60,000 houses in North America to create single-zone models. The model structure was defined with an automated forward selection procedure. 61% of the final models were classified as good fits and the structure of 80% of them was of 5th-order (5 thermal capacitances). An investigation of the 24-hour prediction error of the models showed that the ones classified as good fits are accurate enough for day-ahead predictions and Model Predictive Control (MPC) applications. An analysis of the model parameters suggested no strong correlation between them and the available metadata. The time constants of the houses were estimated, providing valuable information about the houses thermal inertia. Building models that can accurately capture and leverage building thermal inertia are ideal candidates for MPC and energy flexibility applications.