Abstract:

Large-scale vapor-compression HVAC systems involve strongly coupled thermo-fluid dynamics and algebraic junction constraints, making high-fidelity simulation too expensive for real-time control, optimization, and design exploration.
This paper presents a scalable hybrid simulation framework that combines physics-informed neural ordinary differential equations (PINODEs) for component-level heat-exchanger dynamics with differential-algebraic equation (DAE) solvers for system-level coupling. A key feature of the proposed PINODE formulation is that conserved quantities, namely refrigerant mass and internal energy, are predicted as outputs rather than prescribed as inputs. This enables physicsinformed training through automatic differentiation of mass and energy balances while preserving the continuous-time structure needed for long-horizon prediction. At the system level, learned component models are coupled through algebraic pressure constraints and integrated with DAE solvers that enforce junction pressure equilibrium and mass-flow consistency. Case studies on dual-compressor and larger parallel HVAC configurations demonstrate that the proposed framework achieves low prediction error while significantly reducing computational cost relative to high-fidelity reference simulation. In the dual-compressor study, the best DAE-based configuration achieves 2.04% overall MAPE with a simulation time of 58.91 s, while the fastest setting achieves 15.43 s runtime. Across larger networks, the framework scales to systems with up to 16 compressor-condenser pairs, demonstrating its potential for fast and thermodynamically consistent simulation of complex HVAC systems.