基于组合神经网络模型的快堆堆芯瞬态热工水力参数预测方法研究  

Combined neural network-based transient thermal hydraulic parameter prediction method for fast reactor core

作　　者：赵梓炎 赵鹏程[1] 刘紫静[1] 李卫 于涛[1] ZHAO Ziyan;ZHAO Pengcheng;LIU Zijing;LI Wei;YU Tao(School of Nuclear Science and Technology,University of South China,Hengyang 421001,China)

机构地区：[1]南华大学核科学技术学院,衡阳421001

出　　处：《核技术》2025年第4期99-111,共13页Nuclear Techniques

基　　金：核反应堆系统设计技术重点实验室运行基金(No.KFKT-05-FW-HT-20220014)资助。

摘　　要：对于反应堆热工水力参数的预测,现有的研究多使用单一神经网络的预测方法,但在噪声较大的情况下,单一神经网络不能很好地剔除噪声的影响。本文使用基于经验模态分解法(Empirical Mode Decomposition,EMD)与奇异谱分析法(Singular Spectrum Analysis,SSA)结合自适应径向基神经网络(Radial Basis Function Neural Network,RBF)的组合模型提高堆芯热工参数瞬态预测的精度。采用1/2中国实验快堆(China Experimental Fast Reactor,CEFR)为研究对象,使用快堆子通道程序SUBCHANFLOW生成瞬态堆芯热工水力参数的时间序列,并利用组合神经网络模型对堆芯质量流量和包壳表面最高温度时间序列进行单步预测和连续预测。结果表明:相对于单一RBF神经网络,EMD-RBF组合神经网络和EMD-SSA-RBF组合神经网络对质量流量的单步预测误差分别下降41.2%和86.7%,对包壳表面最高温度的单步预测误差分别下降44.7%和60.5%,明显地降低了连续预测误差,且计算时间较短。该方法相比于深度神经网络有一定的优势,对于提高反应堆在工程应用中的安全性有一定的参考价值。[Background]The inner working conditions of a reactor are complicated and affected by many factors.Accurate prediction of the key thermal parameters of the reactor core under various working conditions can greatly improve reactor safety.Most of the existing research focuses on the prediction method that uses a single neural network.In the case of excessive noise,a single neural network cannot sufficiently eliminate noise and accurately detect data change.[Purpose]This study aims to propose a novel transient thermal hydraulic parameter prediction method for fast reactor core,making use of a model that is based on the empirical mode decomposition(EMD)and singular spectrum analysis(SSA)combined with an adaptive radial basis function(RBF)neural network.[Methods]Firstly,the 1/2 China Experimental Fast Reactor(CEFR)was used as the research object,and the fast reactor subchannel program SUBCHANFLOW was employed to generate a time series of transient core thermal hydraulic parameters.Then,two combined models,i.e.,EMD-RBF and EMD-SSA-RBF,were used to predict the core mass flow rate and time series of the maximum temperature on the surface of the cladding.Both the single step prediction and continuous prediction were performed.[Results]The results show that compared with a single RBF neural network,the single-step prediction errors of mass flow rate with the EMD-RBF combined neural network and EMD SSA-RBF combined neural network are reduced by 41.2%and 86.7%respectively,whilst the single-step prediction errors of temperature are reduced by 44.7%and 60.5%respectively.Not only the prediction errors are significantly reduced,but also the calculation time for parameter prediction is shortened.[Conclusions]The combined neural network models proposed in this study can make fast and high-precision predictions,providing advantages over the deep neural network.Hence have certain reference value for improving the safety of the reactor in engineering applications.

关 键 词：经验模态分解 奇异谱分析 径向基神经网络 热工参数预测 快堆 

分 类 号：TL433[核科学技术—核技术及应用]