State prediction and analysis of 3D upper plenum of lead–bismuth fast reactor based on model order reduction under transient accidents

Accurate prediction of three-dimensional (3D) thermal–hydraulic parameter evolution during transients in lead–bismuth fast reactors is important for safety. Although high-fidelity computational fluid dynamic (CFD) models are accurate, they are computationally expensive for real-time use. Model order reduction (MOR) techniques can alleviate this cost while retaining accuracy. In this work, the upper plenum of the lead–bismuth fast reactor NCLFR-Oil is taken as the object of study. Using the proper orthogonal decomposition (POD)-based MOR method and artificial neural networks (ANN), two different 3D transient analysis frameworks are proposed for different data scenarios. 1) A time-series hybrid model (THM) framework designed for time multiple-query tasks, which enables rapid prediction of future three-dimensional physical fields through nonlinear temporal extrapolation of reduced-order modal coefficients. 2) A hybrid data assimilation (HDA) framework aimed at situations with limited sensor data, where the full 3D field distribution is reconstructed using only sparse temperature measurement points by integrating real-time sensor observations with the MOR. The frameworks enhance computational efficiency significantly, with maximum errors around 0.05. Speed-up ratios of 940 and 713 are achieved for THM and HDA frameworks, respectively. Using only three noisy temperature sensors, the HDA framework accurately reconstructs pressure, temperature, and velocity fields, demonstrating robustness and practical applicability. Sensitivity analyses further confirm reliability under varying sensor numbers and noise levels. This work provides an effective tool for real-time monitoring and safety evaluation under accident conditions, offering high practical value.