Scalable tuning of building models to hourly data

Abstract

Energy models of existing buildings are unreliable unless calibrated so that they correlate well with actual energy usage. Manual tuning requires a skilled professional, is prohibitively expensive for buildings below 50,000 ft 2 , imperfect, non-repeatable, and not scalable to the dozens of sensor channels that smart meters, smart appliances, and sensors are making available. A scalable, automated methodology is needed to quickly, intelligently calibrate building energy models to all available data, increase the usefulness of those models, and facilitate speed-and-scale penetration of simulation-enhanced capabilities into the marketplace for actualized energy savings. The "Autotune" project is a novel, model-agnostic methodology that leverages supercomputing, large simulation ensembles, and big data mining with multiple machine learning algorithms to allow automatic calibration of simulations that match measured experimental data in a way that is deployable on commodity hardware. This paper shares several methodologies employed to reduce the combinatorial complexity to a computationally tractable search problem for hundreds of input parameters. Accuracy metrics are provided that quantify model error to measured data from a highly instrumented, emulated-occupancy research home.

Building simulation: Ten challenges

“…Mihai and Zmeureanu proposed a bottom-up calibration technique based on Building Automation System trend data, which starts with zone level calibration with supply air flow rate to each zone, indoor air temperature and cooling load, followed by AHU level calibration. The results show that the AHU model was calibrated naturally on top of most calibrated zones, which avoids any additional tuning through the trial-and-error method (Mihai and Zmeureanu 2017 (Garrett and New 2015).  Several studies address the computational cost of auto-calibration, which has slowed the adoption of such techniques.…”

“…The regression emulator calibrates more quickly while maintaining similar performance compared to the standard Gaussian process emulator (Li et al 2016b). Lastly, for optimization-based auto-calibration approaches, feature selection and sampling is a very important step; emerging methods include Latin Hypercube Sampling (Kim and Park 2016;Yun and Song 2017), Markov chain Monte Carlo (MCMC) (Garrett and New 2015) and No-U-Turn-Sampler MCMC (Chong et al 2017), etc.  Optimization is based on mathematical methods and typically lacks critical inputs from physics and engineering perspectives, thus sometimes leading to unreasonable calibrated results.…”

Evaluation of “Autotune” calibration against manual calibration of building energy models

“…Therefore, there is a need to mitigate this increasing complexity by relying on cost-effective, intelligent algorithms to calibrate BEPS models while using as much data as is available. [115] Figure 1 illustrates Autotune's workflow for replacing the inaccurate and expensive manual calibration process with software intelligence informed by machine learning to fully-automate calibration for the EnergyPlus simulation engine. The Autotune project [113] aims to mitigate increasing complexity with an automated process and has previously demonstrated calibration results for envelope parameters using monthly utility data [48].…”

“…This is much larger than the estimated 10 80 number of subatomic particles in the observable universe. However, the actual size of the search space is effectively infinite, because many of the parameters are continuous-valued [115]. A common approach to such search problems, successfully applied in previous building model tuning efforts, [48] is evolutionary computation (EC).…”

“…A previous study [115] introduces Autotune and shares 5 of the 19 experiments used to refine the calibration methodology as it was being developed; this was done by reducing combinatorial complexity to a computationally tractable search problem that could scale to handle hundreds of parameters and dozens-tomillions of measured data points. Particularly, the study explores methods for speeding up the calibration through parallelization and simulating only portions of a year in the context of overall calibration accuracy for two models (primitive and refined) generated for a single building during a manual calibration.…”