Hierarchical calibration of archetypes for urban building energy modeling

The application of building archetypes is a widespread approach used in urban building energy modeling. Working with archetypes has a range of benefits, but it is important that modelers avoid using oversimplified approaches when establishing the archetype as they lead to loss of uncertainty and, consequently, to models with inferior predictive capabilities. In this paper, we propose a multilevel take on the challenge of establishing archetypes. A simultaneous modeling and calibration framework is formulated using Bayesian inference techniques - a technique that allows for the propagation of uncertainty throughout the calibration process. By means of hierarchical modeling, information from training buildings is partially pooled together to form an optimal solution between separate building energy models and a completely pooled model. This enables the inference of uncertain archetype parameters that are less prone to building outliers than what is achieved using ordinary aggregation of individual building estimates. The proposed framework incorporates dynamic building energy modeling of arbitrary temporal resolution where uncertain parameters are fitted for individual building models and the archetype model simultaneously. The application of the framework is demonstrated using case-study data from the Danish residential building stock, containing 3-hourly measurements of energy use for 50 training buildings. The model is tested for the prediction of 100 out-of-sample test buildings' aggregated energy use time series on a holdout validation period. With a prediction error of only NMBE = 2.9% and CVRMSE = 7.8%, the archetype framework promises well for urban modeling applications. (C) 2018 Elsevier B.V. All rights reserved.