TR2022-030

Scalable Bayesian Optimization for Model Calibration: Case Study on Coupled Building and HVAC Dynamics


    •  Chakrabarty, A., Maddalena, E., Qiao, H., Laughman, C.R., "Scalable Bayesian Optimization for Parameter Estimation of Coupled Building and HVAC Dynamics", Energy and Buildings, DOI: 10.1016/​j.enbuild.2021.111460, Vol. 253, pp. 111460, March 2022.
      BibTeX TR2022-030 PDF
      • @article{Chakrabarty2022mar2,
      • author = {Chakrabarty, Ankush and Maddalena, Emilio and Qiao, Hongtao and Laughman, Christopher R.},
      • title = {Scalable Bayesian Optimization for Parameter Estimation of Coupled Building and HVAC Dynamics},
      • journal = {Energy and Buildings},
      • year = 2022,
      • volume = 253,
      • pages = 111460,
      • month = mar,
      • doi = {10.1016/j.enbuild.2021.111460},
      • url = {https://www.merl.com/publications/TR2022-030}
      • }
  • MERL Contacts:
  • Research Areas:

    Machine Learning, Multi-Physical Modeling, Optimization

Abstract:

Model calibration for building systems is an key step to achieving accurate and reliable predictions that reflect the dynamics of real systems under study. Calibration becomes particularly challenging when integrating building and HVAC dynamics, due to large-scale, nonlinear, and stiff underlying differential algebraic equations. In this paper, we describe a framework for calibrating multiple parameters of coupled building/HVAC models using scalable Bayesian optimization (BO), whose advantages include global optimization without requiring gradient information, and its ability to perform calibration in a data-efficient manner. The proposed methodology is improved online via two additional steps: domain tightening and domain slicing, both of which leverage the surrogate calibration cost function. We demonstrate effectiveness of the proposed algorithm by simultaneously calibrating 17 parameters (including emissivities, heat transfer coefficients, and thickness of walls/floors) of a Modelica model of joint building and HVAC dynamics, with 2 weeks worth of training data. This high-dimensional calibration task is solved via our proposed method, which yields parameters that are > 90% accurate with < 1000 model simulations, and the outputs of the final calibrated model on unseen testing data complies with standard ASHRAE calibration guidelines.