Abstract:

Next-generation vapor compression cycles will rely upon multicomponent refrigerant mixtures to reduce the climate impact of the working fluids, but the computation of thermodynamic property data for these mixtures is numerically challenging and often results in non-physical anomalies that are present in the output of standard calculation tools. In this paper, we explore two alternative techniques for mitigating the effect of these anomalous points in a reference dataset. The first of these approaches is based upon heteroscedastic Gaussian processes, and builds a statistical model of the property to identify outliers in the reference data. The second uses an estimation method based upon constrained optimization to first detect these outliers and then compute optimal perturbations to the reference data so that the resulting target dataset satisfies domain-informed constraints on the reference data. We demonstrate the efficacy of these methods in computing a target dataset for the refrigerant R454C that is free of anomalies, and which can then be used to build computationally efficient models for use in dynamic cycle simulations.