A physics–neural network hybrid framework for accurate modeling of direct expansion refrigeration systems

• Neural network models electronic expansion valve throttling from six inputs. • Physics–neural hybrid model is built for a direct-expansion refrigeration system. • Valve model reaches 1.14% mean relative error, with <4.5% maximum error. • External test gives 1.39% mean relative error, with ≤5% maximum error. • Hybrid model predicts cooling capacity and sensible heat ratio within ±7%. Accurate modeling of direct expansion (DX) refrigeration systems is critical for optimizing energy efficiency, control performance, and system design. Traditional physics-based models, although grounded in thermodynamics, often fail to capture the nonlinear behaviors of complex components such as the electronic expansion valve (EEV), while purely data-driven models lack interpretability and generalizability. To address these limitations, this study developed a hybrid modeling framework that integrated physics-based formulations with an artificial neural network (ANN). The ANN component modeled the EEV’s nonlinear throttling behavior, whereas the compressor, condenser, and evaporator were represented by validated physical sub-models. The hybrid model was trained and validated using high-resolution experimental data from a variable-speed DX test rig and was further assessed with external datasets for generalization. Results showed that the ANN-based EEV model achieved a mean relative error below 5%, and the integrated hybrid model predicted system cooling capacity and sensible heat ratio with overall errors within ±6%, representing an approximately 50% improvement in accuracy compared with conventional physics-based models. This hybrid framework effectively combines physical interpretability with data-driven flexibility, providing a robust and scalable basis for applications in performance prediction, model predictive control, and intelligent fault diagnosis of refrigeration systems.