Bayesian calibration of building energy models: Comparison of predictive accuracy using metered utility data of different temporal resolution

Modern smart meters in heating systems offer building energy data of high temporal resolution. Compared to the annually aggregated readings used for conventional billing, the continuous information flow from these smart meters can be made available as time series data containing monthly, daily or even hourly aggregated values. In this paper, the effect of different temporal aggregation levels of commercial smart meter data on building energy model (BEM) calibration is investigated. Four different aggregation levels of a training data set were applied for calibration of six BEM input parameters to set up a Gaussian process emulator of the physical system. The performance of the emulator was subsequently tested on an unseen validation data set. Results reveal a systematic pattern of increasing predictive accuracy as a function of increasing training data resolution.