Integrating CFD and Non-intrusive Reduced-order Method for Digital Twin Construction of PWR Upper Plenum and Hot Leg

The upper plenum and hot leg of pressurized water reactor (PWR) exhibit significant thermal stratification in the hot leg due to the complex internal structure of components and the three-dimensional flow characteristics of the coolant exiting the core. The spatial variability of temperature measurements at the same cross-section in the hot leg makes it difficult for traditional average temperature methods to accurately characterize the actual operating conditions, thereby directly affecting the precision of core physical parameter calculations. Therefore, constructing a digital twin model of the reactor upper plenum and hot leg to obtain precise and online temperature distribution is of great significance for reactor operation monitoring. However, the highly complex internal components and flow field in the upper plenum make traditional computational fluid dynamics (CFD) methods based on spatial-temporal discretization unsuitable for the online computational requirements of digital twins. To address this issue, this study proposed an efficient method for constructing a digital twin model that could rapidly predict and accurately reconstruct the temperature field of the upper plenum and hot leg by combining non-intrusive reduced-order techniques. First, transient full-order field data under various inlet temperature distributions were generated through CFD simulations. Fast Fourier transform was then applied to extract the mean temperature field, amplitude, and centroid frequency across the entire space, which were combined into a snapshot matrix to characterize the transient flow behavior and serve as the prediction target for the digital twin model. Subsequently, proper orthogonal decomposition (POD) was employed to extract the reduced-order basis and coefficients from the full-order snapshots. A mapping between the inlet temperature distribution parameter and the reduced-order coefficients was established using Gaussian process regression (GPR), resulting in a reduced-order model with rapid prediction capabilities. Finally, with the temperature information from the temperature bypass as input, the reduced-order model was used as the forward computational model. A hybrid global-local optimization algorithm combining grid search and gradient descent was applied to invert the inlet temperature distribution parameters, further reconstructing the mean temperature field, amplitude, and centroid frequency across the entire space. The average temperature of the coolant in the bypass was calculated based on the reconstructed temperature field. The results indicate that compared with the full-order model, the reduced-order model developed in this study achieves a computational speed improvement of over four orders of magnitude, with a maximum prediction relative deviation not exceeding 22.05%. The average temperature of the coolant at the hot leg cross-section obtained through inversion has a deviation of less than 0.05 K compared with the CFD results, which is significantly better than the error level based on average measurements. The model effectively meets the real-time computational requirements of digital twins while maintaining prediction accuracy, providing a methodological reference for reactor operation monitoring.