87 / 2024-08-28 15:25:08

Modeling Transient Thermal Stratification in the Upper plenum of a MONJU Reactor Using Entropy Production Analysis Combined with CNN-LSTM

Thermal Stratification; CFD; Entropy Generation Analysis; Convolutional Neural Networks; Long Short-Term Memory networks

The phenomenon of thermal stratification affects the core residual heat dissipation capability and causes thermal fatigue and thermal stress in components. Detailed three-dimensional Computational Fluid Dynamics (CFD) analysis methods are time-consuming, resource-intensive, and challenging in terms of convergence during long transient simulations. This paper proposes a rapid and intelligent prediction model for thermal stratification based on entropy production analysis. The model first analyzes the impact of temperature difference heat transfer entropy production rates and convective heat transfer entropy production rates on transient thermal stratification, and defines a dimensionless number for the self-preservation of thermal stratification (Ssp) in the upper chamber during accident transients to evaluate the stability of the thermal stratification state. By integrating Convolutional Neural Networks (CNNs) with Long Short-Term Memory Networks (LSTMs), and employing entropy generation analysis to emphasize thermal stratification features while disregarding non-stratified regions, the difficulty of feature extraction for CNNs is reduced. This approach avoids extensive temperature feature extraction, thereby lowering computational complexity and mitigating the risk of overfitting, while enhancing the performance of time series analysis for rapid prediction of large-scale accident transient thermal stratification. Additionally, the impact of the transition from inertial force dominance to buoyancy force dominance in the flow field, due to poor self-preservation of thermal stratification, on the LSTM's ability to capture sequential dependencies is analyzed. Compared to three-dimensional CFD methods, this model achieves accurate prediction of thermal stratification's location, contour, and intensity at different times, while simultaneously reducing computational resources and time.