基于决策树的堆芯物理参数预测研究

Prediction of Core Physical Parameter Based on Decision Tree

反应堆堆芯核设计涉及大量方案的搜索与详细计算,缩短方案搜索时间有利于提高核设计效率。数据挖掘技术通过对大量数据进行学习与模式识别,可实现核设计方案物理参数的快速预测,更快地筛选出可行的备选堆芯方案。本文基于数据挖掘的决策树4种算法:C4.5、RepTree、Random Forest及Random Tree,在计算时以燃料富集度、含可燃毒物燃料棒数量及含量作为自变量,以寿期内keff不均匀系数偏差(KUCD)、径向功率不均匀系数偏差(RPNCD)、径向中子通量不均匀系数偏差(RFNCD)、堆芯寿期(CL)作为目标函数,构成目标函数符合度(CPF),利用大量已知核设计参数的组件及堆芯设计方案作为数据挖掘训练集,构建数据挖掘模型,并用于对未知核设计参数的组件方案集合(测试集)进行CPF快速预测。结果表明,4种算法利用训练集构建数据挖掘模型的时间在0.6 s以内,各算法的交叉验证精度均在0.7以上,其中C4.5算法对CPF预测精度最高;对测试集方案的核设计参数预测中,单个方案的预测时间均在0.9 s以内,而Random Forest算法对CPF等于4的预测效果最好。

The reactor core design involves search and detailed calculation of a large number of schemes. Shortening the search time is beneficial to improve the efficiency of nuclear design. Through learning and pattern recognition of large data, data mining technology can realize rapid prediction of physical parameters for nuclear design schemes, and fast screening of feasible alternative core schemes. Four decision tree algorithms were used in this paper, C4.5, RepTree, Random Forest and Random Tree, respectively. Fuel enrichment, and the number of fuel rods containing burnable poison and content of burnable poison were taken as independent variables in the calculation. The number of core parameters fulfilled(CPF) was a combination of the four parameters, keff unevenness coefficient deviation during lifetime(KUCD), radial power non-uniformity coefficient deviation(RPNCD), radial flux non-uniformity coefficient deviation(RFNCD) and core life(CL). Data mining models were constructed through learning the training set consisted by large number of assembly and core schemes whose nuclear design parameters were already known. Using the models, the CPF values for unknown assembly set(the testing set) were quickly predicted. The results show that model learning time of the four data mining algorithms are less than 0.6 s for the training set, the cross-validation accuracy of each algorithm is above 0.7 and C4.5 algorithm has the highest accuracy for the overall prediction of CPF values. In addition, for CPF value prediction of testing set, time for each scheme is within 0.9 s, while Random Forest algorithm has the highest prediction accuracy for CPF=4.