A physics-data fusion framework for fast modeling of variable refrigerant flow system using exact and trend-based physics-constrained learning

Constructing digital models of variable refrigerant flow (VRF) systems requires massive amounts of operational data, including system states, control parameters, and ambient conditions. However, acquiring these data is expensive. The objective of this study is to develop a fast, accurate, and robust modeling platform for VRF systems to enable the low-cost acquisition of operational state data. This modeling platform is built using a multi-scale physics-data fusion (MS-PDF) framework, which employs an exact and trend-based physics-constrained neural network (ET-PCNN) for component-level modeling, and utilizes a two-phase fluid network for automated system-level assembly. To validate the speed, accuracy, and robustness of the MS-PDF method, tests were conducted across a wide range of operating conditions, and the model was subsequently applied to optimize operational parameters. The quantitative results of this study demonstrate that the ET-PCNN component models achieve an accuracy of at least 98.3%. At the system level, the MS-PDF framework maintains a mean relative error of 0.80% to 3.81% while accelerating the simulation time from 2 min to 5 s. Furthermore, applying this framework to real-time operational optimization achieves a 19.7% reduction in power consumption during a typical operating day. These results provide valuable insights for the digital modeling of VRF systems and establish a pathway for deploying intelligent optimization in buildings.