A physic-extract-constrained neural network for performance prediction of finned-tube condenser in air source heat pump system

Abstract

It’s vital to construct a fast and robust simulation model for the digital twins and intelligent control of air source heat pump (ASHP) system, where heat exchangers are the most complex and time-consuming parts. Therefore, a physics-exact-constrained neural network (PeCNN) is proposed to achieve highly generalized prediction of finned-tube condenser performance in the ASHP system under small data sample conditions. General framework of the PeCNN is to construct loss function integrated with physical constraints based on the 1-D physical correlation-based distributed parameter (PCDP) method and attention mechanism based residual network, resulting in the effective solution of the multiple non-smooth convex optimization problems in parallel. Comprehensive data experiments are conducted in the designed dataset. The results show that the proposed model achieves coefficient of determination ($R^2$) of 0.998 for the refrigerant pressure drop prediction. Compared with other data-driven models, abilities of capturing physical details and extrapolation can be improved by 85% and 11.7%, respectively. The method of physical knowledge integration of the PeCNN also provides a new perspective on fast modeling ASHP systems.