Oil circulation rate prediction by transformer based deep learning model in R290 heat pump systems

R290 (propane) has emerged as a leading candidate for next-generation heat pump systems due to its negligible ozone depletion potential (ODP), very low global warming potential (GWP), and excellent thermophysical properties. However, its relatively low density makes oil transport within the cycle less favorable, highlighting the importance of accurate Oil Circulation Rate (OCR) control. Excessive oil migration can degrade performance and hinder compressor lubrication. This study develops and validates a framework that combines indirect sensing with deep learning for real-time, non-invasive OCR prediction in R290 heat pumps. A total of 54 heating-mode experiments were conducted under diverse conditions, yielding OCR values from 0.31 % to 1.51 %. Excessive OCR caused up to a 10.3 % reduction in refrigerant mass flow rate, a 15.5 % drop in COP, a 5.7 % decline in heating capacity, and a 20 % decrease in condenser pressure at 0 °C. Deep learning models (MLP, LSTM, Transformer) were trained on 44 datasets and tested on 10 unseen cases, achieving mean absolute percentage errors below 10 % for OCR prediction. Among them, the Transformer showed the best performance, with an RMSE of 0.0179 % under transient compressor frequency variations. Moreover, the self-attention analysis captured feature dependencies consistent with oil entrainment mechanisms, highlighting the Transformer's capability to provide physically meaningful explanations beyond pure prediction. Finally, Transformer-predicted OCR was applied as a virtual sensor to estimate component-wise oil retention using a corrected mixture law, yielding 3–12 g.