Abstract:

Data-driven dynamic models typically offer faster execution than their physicsbased counterparts described by large systems of nonlinear differential-algebraic equations (DAEs) with quantitatively reasonable accuracy. Therefore, development of such models can be extremely useful for design optimization, controls, fault detection and diagnostics of vapor compression based building
Heating, ventilation and air conditioning (HVAC) systems. As the complexity and scale of vapor compression systems (VCS) increase rapidly across the industry, a modular approach of generating and interconnecting data-driven component models enables model reuse and efficient adaption to arbitrary system layouts. Despite the flexibility, the modular integration for system model generation can suffer from nonphysical behaviors of violating conservation laws due to inevitable prediction errors associated with each component.
This paper presents a data-driven modeling framework that exploits state-of-the-art deep learning techniques for constructing component models, while enforcing physical conservation for system simulations. A general-purpose system solver is developed to handle arbitrary configurations by automatically integrating data-driven or interchangeable physics-based component models into a system model. Results of an air-source heat pump system reveal a significant speedup with good agreement, compared to high-fidelity first-principles models.