The thermal stratification in the upper plenum of lead-bismuth fast reactor after emergency shutdown has a significant impact on the structural integrity of the reactor and the residual heat removal capacity of the natural circulation. The research on thermal stratification based on Computational Fluid Dynamics (CFD) method has the problems of large computational overhead and time-consuming whilst the existing standard dynamic mode decomposition (DMD) method has poor forecasting results on thermal stratification.

This study aims to solve this problem by proposing a thermal stratification model reduction method for the upper plenum of lead-bismuth fast reactor.

Firstly, the high-precision full-order snapshot was obtained on the basis of the CFD program FLUENT. Then, based on the truncated DMD, the time step samples were compressed according to the characteristic frequency, and the thermal stratification reduction model was constructed by combining the Long Short-Term Memory (LSTM) neural network with DMD. Finally, three methods, i.e., standard DMD, improved DMD and improved DMD-LSTM, were comparatively analyzed in terms of temperature oscillation error and computation time.

Computational results show that the thermal stratification model reduction method based on improved DMD and LSTM in the upper plenum of the lead-bismuth fast reactor achieves best performance, with root mean square error reduced by 46.60% and 30.45% respectively, compared to standard DMD and improved DMD. The computational time of the improved DMD and LSTM is only 4.4% of FLUENT's, significantly improving efficiency and enabling faster emergency response in lead-bismuth reactors.

Results of this study verify that the thermal stratification model reduction method proposed in this paper can better simulate the temperature distribution in the upper plenum and realize the rapid prediction of the thermal stratification phenomenon.