Physics-constrained neural network for chiller guided by reverse Carnot cycle analogy model with extrapolation ability

Amid growing global energy demand and rising building energy consumption, chillers, which are the core energy-consuming equipment in heating, ventilation, and air conditioning (HVAC) systems, require accurate modeling for energy efficiency optimization and carbon emission control. However, traditional white-box models are hardly practical due to complex structures and difficult parameter acquisition, while purely data-driven models lack physical mechanism support, leading to poor extrapolation beyond training data. Moreover, physics-constrained neural network methods used in existing studies are limited to simple physical information such as monotonicity and energy conservation. To address these problems, this study proposes a Physics-Constrained Neural Network Guided by reverse Carnot cycle analogy model (PCNN-GC). A simplified mechanism model with correction factor is built and integrated into multiple physical constraints, forming a framework balancing physical consistency and data adaptability. The CoolProp library is connected with PyTorch for differentiable refrigerant thermophysical calculations and stable training. Validated via four extrapolation scenarios using actual operational data and benchmarked against Multilayer Perceptron (MLP), Random Forest (RF), and Support Vector Machine (SVM), the PCNN-GC demonstrates superior extrapolation capabilities and learning efficiency, achieving high accuracy with only three or four external data augmentation samples. Compared to the next-best model, it reduces the Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) by 26.7 % and 33.3 %, respectively, and increases the coefficient of determination (R2) by 1.76 times. Furthermore, standard deviations of the PCNN-GC's performance metrics across all extrapolation scenarios are significantly lower than those of the MLP, demonstrating its superior robustness.