Improvements and intelligence integration of virtual refrigerant charge (VRC) sensor.

Refrigerant leakage significantly undermines the energy efficiency and operational safety of variable refrigerant flow (VRF) systems, making accurate prediction of refrigerant charge critically important. Conventional diagnostic approaches are often costly, reliant on complex models, non-quantitative, and lack generalization, which restricts their practical deployment. To address these limitations, a series of enhancements and intelligent integration were introduced to the virtual refrigerant charge (VRC) sensor. An operating-condition matching strategy was first employed to establish an exVRC sensor for condition extending. An exponentially weighted moving average (EWMA) control chart was then incorporated to construct an exVRC-E sensor for oscillation mitigation. Finally, a deep learning-based Residual Neural Network (ResNet) was established and coupled with the exVRC-E sensor to produce an AI-knowledge dual-driven intelligent sensor, exVRC-ER. Experimental validation on a 33.5 kW VRF system under one rated and fourteen off-rated conditions showed that compared with the original VRC sensor, the exVRC sensor reduces MAPE by 16.21 % under extreme off-rated conditions, corresponding to a 74 % relative reduction. The exVRC-E sensor further lowers oscillation amplitude by 84 % and reduces false-alarm risk during normal operation. Across all conditions, the final exVRC-ER intelligent sensor integrated with deep learning achieves the best performance, with a 71.8 % relative reduction in MAPE and a 1.21 kg decrease in root mean square error (RMSE) compared with the original VRC sensor. These results indicate a significant potential for precise quantification of refrigerant leakage, highlighting their importance for enhancing the efficient operation and intelligent maintenance of HVAC systems.