Deep learning-based prediction of oil reversal in R290 heat pump systems

Recently, the R290 refrigerant has attracted significant attention due to its low global warming potential (GWP) and excellent thermal performance. To evaluate the reliability of R290 heat pump systems influenced by oil behavior of Polyalkylene glycol (PAG), this study introduces a novel oil reversal index (ORI). This index is defined as the ratio of the oil film thickness at the top and bottom of vertical pipes, providing a method to determine the occurrence and intensity of oil reversal. ORI is a metric that is not only easy to measure but also capable of accounting for the effects of oil viscosity and refrigerant solubility. It was experimentally measured under both transient and steady-state conditions, influencing factors were analyzed, and it was subsequently modeled using deep learning. The long short-term memory model with batch normalization (LSTM + BN) achieved a mean absolute percentage error (MAPE) of 12.64 % in predicting oil film thickness under transient conditions. Furthermore, by selecting top 10 most impactful parameters through feature importance analysis and retraining the model, this error was reduced to 8.81 %. Additionally, the model predicted ORI under steady-state conditions with an error of 2.21 % using 20 input features.