Research on fine analysis and accelerated prediction technology for thermal stratification in the upper plenum of the “Monju” reactor

Abstract

Thermal stratification phenomena in SFRs and machine learning applications in nuclear system R&D have captured the community's attention. Existing methods, such as CFD, demand significant time and computational resources, hindering the timely prediction and management of thermal stratification in reactors. In response, this paper introduces a novel approach employing data-driven models for super real-time prediction of thermal stratification. Themodel aims to enhance the efficiency and speed of thermal stratification predictions. Refined flow field data were first obtained via CFD, revealing the transition mechanism between the dominant roles of inertial and buoyancy forces during thermal stratification formation. To select the optimal data-driven model, this paper employs the proper orthogonal decomposition (POD) method to evaluate three algorithms - k-nearest neighbor (KNN), classification and regression tree (CART), and least absolute shrinkage and selection operator (LASSO) - basedon temperature data from a temperature probe. Evaluation reveals that the POD-KNN model offers notable improvements in prediction speed while ensuring consistent accuracy, outperforming traditional numerical simulation methods for super real-time thermal stratification prediction.

Keywords

• Thermal stratification
• CFD
• Machine learning
• K-nearest neighbor
• Proper orthogonal decomposition