New paper in Energy and Buildings: “Physics vs. Structure: A Systematic Benchmark of Learning Strategies for Multi-Zone Building Thermal Dynamics.”
We provide an extensive benchmark that systematically compares a range of modeling approaches, from linear models to neural ODEs and physics-constrained models, for data-driven modeling of thermal dynamics in buildings. Besides prediction accuracy, we thoroughly quantify computational cost and physical consistency of the models, which is critical for HVAC control applications.
paper:
sciencedirect.com/science/ar…
All the experiments have been made possible through the Neuromancer scientific machine learning library:
github.com/pnnl/neuromancer
Thanks to my PNNL colleagues Soumya Vasisht, Cary Faulkner, Elad Michael, Aaron Tuor, and Draguna Vrabie for this fruitful collaboration, as well as the funding from
@ENERGY that supported this work.