Abstract:

Data-driven dynamic models typically offer faster execution than their physics-based counterparts described by large systems of nonlinear and stiff differential-algebraic equations with satisfactory accuracy. Therefore, development of accurate but computationally efficient models directly identified from data constitutes a solution path for investigating controls, fault detection and diagnostics of vapor compression systems (VCS). A modular approach of generating and interconnecting data-driven component models enables reuse of readily trained models and adaption to arbitrary system configurations. Despite the flexibility, a modular integration for system model generation can suffer from nonphysical behaviors of violating conservation laws due to inevitable prediction errors associated with each component model. This paper presents a data-driven dynamic modeling approach for VCS that exploits state-of-the-art deep learning methods for constructing component models while enforcing physical conservation for system simulations. Specifically, gated recurrent unit (GRU) and feedforward neural network models are employed for heat exchangers and mass-flow devices. Predictive capabilities and conservation properties of the proposed modeling approach is demonstrated via a case study of an air-source heat pump system. Simulation results reveal a significant speedup with negligible discrepancies compared to high-fidelity physics-based models.