TR2025-043
Time-Series Generative Networks for Synthesizing Realistic Scenarios in Occupant-Centric Building Simulation
-
- "Time-Series Generative Networks for Synthesizing Realistic Scenarios in Occupant-Centric Building Simulation", Building Simulation, April 2025.BibTeX TR2025-043 PDF
- @article{Chakrabarty2025apr,
- author = {Chakrabarty, Ankush and Vanfretti, Luigi and Wang, Ye and Mineyuki, Takuma and Zhan, Sicheng and Tang, Wei-Ting and Paulson, Joel A. and Deshpande, Vedang M. and Bortoff, Scott A. and Laughman, Christopher R.},
- title = {{Time-Series Generative Networks for Synthesizing Realistic Scenarios in Occupant-Centric Building Simulation}},
- journal = {Building Simulation},
- year = 2025,
- month = apr,
- url = {https://www.merl.com/publications/TR2025-043}
- }
,
- "Time-Series Generative Networks for Synthesizing Realistic Scenarios in Occupant-Centric Building Simulation", Building Simulation, April 2025.
-
MERL Contacts:
-
Research Areas:
Abstract:
Occupant-centric building simulation models rely on two key factors: our understanding of the underlying physics that govern thermal dynamics, and realistic modeling of occupancy patterns and energy use within the zone of interest. While current physics-oriented building simulation models predict thermal dynamics accurately, a systematic and scalable way to generate occupancy and energy use patterns remains an open challenge despite the large amount of data collected from building sensors across academic and industry efforts. In this paper, we leverage deep generative net- works capable of learning from real building data for generating realistic occupant-centric scenarios to inform building simulations. Our ultimate goal is to assess building performance over a wide range of generated scenarios, which is currently done either by taking a small set of ‘nominal scenarios’ or by handcrafting specific scenarios, both of which restrict the quality of building performance assessment to a few biased use-cases. For the purpose of generating scenarios automatically, we employ a recently proposed architecture called RAFT-VG (regularized adversarially finetuned VAE-GAN) that combines the benefits of variational autoencoders (VAEs) and generative adversarial networks (GANs), and demonstrate its capacity for synthesizing a variety of signals including occupancy patterns, internal heat loads, and ambient conditions. A key feature of this neural architecture is that the generative process depends solely on a conditional decoder network. Distilling the deep RAFT-VG model to a simpler decoder for inference allows us to propose a general framework for integrating the generative model directly in Modelica. The closed-loop building performance with various generated scenarios, along with the Modelica integration, is demonstrated via simulation use-cases using the Modelica BESTEST repository.