TR2024-113

Assessing Building Control Performance Using Physics-Based Simulation Models and Deep Generative Networks


    •  Chakrabarty, A., Vanfretti, L., Bortoff, S.A., Deshpande, V.M., Wang, Y., Paulson, J.A., Zhan, S., Laughman, C.R., "Assessing Building Control Performance Using Physics-Based Simulation Models and Deep Generative Networks", IEEE Conference on Control Technology and Applications (CCTA) 2024, DOI: 10.1109/​CCTA60707.2024.10666585, August 2024.
      BibTeX TR2024-113 PDF
      • @inproceedings{Chakrabarty2024aug,
      • author = {Chakrabarty, Ankush and Vanfretti, Luigi and Bortoff, Scott A. and Deshpande, Vedang M. and Wang, Ye and Paulson, Joel A. and Zhan, Sicheng and Laughman, Christopher R.}},
      • title = {Assessing Building Control Performance Using Physics-Based Simulation Models and Deep Generative Networks},
      • booktitle = {IEEE Conference on Control Technology and Applications (CCTA) 2024},
      • year = 2024,
      • month = aug,
      • doi = {10.1109/CCTA60707.2024.10666585},
      • url = {https://www.merl.com/publications/TR2024-113}
      • }
  • MERL Contacts:
  • Research Areas:

    Control, Machine Learning, Multi-Physical Modeling

Abstract:

Despite the increasing sophistication of building simulation models and digital twins, some components of the building energy system remain challenging to model using first-principles. For instance, internal heat loads generated by occupants in buildings and their usage of the building are often assumed rather than learned. Consequently, building control systems are designed based on average human behavior, and the assessment of the closed-loop system is often limited to a small set of handcrafted scenarios. In this paper, we propose the use of deep generative networks to complement the physics- based simulation platforms by learning time-series distributions from real occupancy and usage data. The learned distribution can subsequently be sampled to construct scenarios that can drive the building simulation model as a disturbance input. This capability enables the more systematic construction of a large set of scenarios for controller performance assessment. Due to the expense of the simulator and the unknown relationship between the disturbance inputs and performance, we provide a sample-efficient algorithm for extracting ‘limiting scenarios’, i.e., disturbance inputs that are most likely to result in the best and worst closed-loop performance. We demonstrate the potential of our proposed framework using Mitsubishi Electric’s SUSTIE building data, on a building simulation benchmark model implemented in Modelica. We report that the generated scenarios preserve signal features observed in the true data while enabling the automatic identification of low-probability scenarios for controller evaluation, and our sampling method determines these limiting scenarios using only 300 simulations.