Comparison between different refrigerant charge level predictive methods in water-to-water heat pump

Heat pumps performance is negatively impacted by non-optimal refrigerant charge levels, often resulting from improper installation or leakage. This study compares different non-invasive methods to predict the charge levels in a R513A water-to-water heat pump. Theoretical approaches available in the literature fail to provide accurate results when the heat pump operating parameters are affected by the presence of a liquid receiver and two-phase refrigerant at the condenser outlet. To overcome this limitation, two empirical approaches using artificial neural networks (ANN) are developed. The first ANN, which uses the same inputs as the best-performing model in the literature, shows slight improvements but fails below 95% of the nominal charge. The second ANN utilizes inputs from commonly found heat pump sensors. This approach achieves 1.0% uncertainty in determining charge levels between the nominal charge and the level where subcooling collapses to zero. Beyond this threshold, the approach remains capable of detecting charge reductions with a 9.1% uncertainty of the nominal charge.