Abstract:

Machine learning (ML) is promising for heating, ventilation, and air conditioning (HVAC) modeling because it can provide fast surrogate models that capture complex nonlinear behavior. Most existing ML-based vapor-compression system (VCS) models, however, adopt a monolithic formulation that learns cycle-level behavior directly. Such models are tied to a fixed system topology, require retraining when the refrigerant circuit changes, and may violate mass and energy conservation, leading to drift and instability in long-horizon simulations. We propose a modular, deep learning framework in which individual VCS components are learned independently during their respective training phases and then assembled into a cycle-level simulator. The resulting deep learning framework is modular, allowing various VCS cycle configurations to be constructed from pretrained component models. Within this framework, dynamic components are represented in continuous time using neural ordinary differential equations. To ensure physical consistency, we enforce mass and energy conservation at component interconnections as exact algebraic constraints during system assembly, rather than as penalty terms during training. This design en- sures that the assembled cycle satisfies fundamental conservation laws and results in a numerically stable cycle-level simulation over long time horizons, even when individual component models are imperfect. We evaluate the framework on two air-source heat pump configurations with different topologies (single- and dual- compressor) using a high-fidelity Modelica reference. Across thermodynamic and air-side variables, the assembled cycle achieves a maximum mean absolute percentage error of 2.15% while satisfying mass and energy conservation by construction. The proposed simulator runs 8.7× faster (single-compressor) and 5.54× faster (dual-compressor) than the Modelica baseline, enabling stable long-horizon rollout suitable for control-oriented simulation and topology variation studies.