✨ TL;DR
This paper develops a hybrid simulation framework for large HVAC systems that combines physics-informed neural networks with traditional differential-algebraic equation solvers, achieving multi-fold speedups over high-fidelity simulations while maintaining low errors. The approach scales to systems with 32+ compressor-condenser pairs by learning component-level dynamics and enforcing system-level physical constraints.
Simulating large-scale HVAC systems is computationally expensive when using high-fidelity physics-based models, limiting their use in real-time control, optimization, and design exploration. Traditional simulation approaches struggle to balance accuracy with computational efficiency, especially as system complexity grows. Purely data-driven methods often fail to respect fundamental physical laws like mass and energy conservation, leading to unstable predictions over long time horizons. There is a need for scalable simulation methods that can handle large HVAC networks while maintaining physical consistency and computational tractability.
The framework operates at two levels. At the component level, physics-informed neural ordinary differential equations (PINODEs) learn heat-exchanger dynamics by predicting conserved quantities (refrigerant mass and internal energy) as outputs, with physics constraints enforced through automatic differentiation of mass and energy balance equations. Stability is achieved using gradient-stabilized latent evolution with gated architectures and layer normalization. At the system level, learned components are integrated with differential-algebraic equation (DAE) solvers (IDA and DASSL) that explicitly enforce junction constraints such as pressure equilibrium and mass-flow consistency. Bayesian optimization tunes solver parameters to balance accuracy and efficiency. A lightweight corrector network trained on short trajectory segments reduces residual system-level bias.