Development of a non-intrusive ROM for 5 × 5 rod bundles of PWR using small sample data

Abstract

This study aims to develop an efficient Reduced-Order Model (ROM) for rod-type fuel assemblies in pressurized water reactors (PWR), essential components of typical PWR cores. This paper integrates Proper Orthogonal Decomposition (POD) with machine learning (ML) for the rapid prediction of the thermal–hydraulic (TH) parameter field in the reactor core, aiding in real-time monitoring and optimization. This paper focuses on validating the model’s feasibility and accuracy for PWR core components operating at 15.5 MPa pressure. To verify the applicability of the POD method in predicting small sample data under actual operational conditions, this study uses actual operational boundaries from nuclear power plants and 121 sets of small sample data as boundary conditions. Extensive Computational Fluid Dynamics (CFD) calculations are conducted to collect data for constructing a snapshot matrix and further developing the ROM. A data-driven approach is employed to explore the relationship between POD coefficients and boundary conditions, achieving rapid prediction of the TH parameter field under any combination of boundary conditions. The study compares the predictive performance of four machine learning algorithms: Radial Basis Function Neural Networks (RBFNNs), K-Nearest Neighbors Regression (KNN), Random Forest Regression (RF), and Gradient Boosting Regression (GBR). Results show that all ML algorithms can construct ROM and predict the TH parameter field with reasonable accuracy for test data. Among them, Random Forest Regression achieves the highest precision, with an average Mean Square Error (MSE) of approximately 3.24 × 10⁻⁴ and a coefficient of determination (R²) of around 0.96. Furthermore, the average prediction time for the detailed three-dimensional parameter field per component is only 15.0 s.