✨ 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.
What the paper shows.
The proposed framework achieves multi-fold speedups over high-fidelity simulations while maintaining mean absolute percentage errors (MAPE) below a few percent. The approach successfully scales to large systems with up to 32 compressor-condenser pairs, demonstrating both computational efficiency and accuracy across dual-compressor configurations and scaled network studies. Bayesian optimization of solver parameters enables effective accuracy-efficiency trade-offs, and the corrector network successfully reduces residual system-level bias in the predictions.
The paper does not explicitly detail limitations, but implicit constraints include the need for training data from high-fidelity simulations to learn component models, potential challenges in generalizing to HVAC configurations significantly different from training conditions, and the requirement for careful tuning of solver parameters through Bayesian optimization. The corrector network approach suggests that some system-level bias persists after component integration, requiring additional correction. Scalability beyond 32 compressor-condenser pairs and performance on highly transient or extreme operating conditions are not discussed.
✨ Generated by Claude · Apr 21, 2026 · Read the PDF for authoritative content.