Research on Fast Prediction Method of Neutron Flux Based on Hybrid Driven Reduced Order Model
-
摘要: 反应堆参数发生扰动后的瞬间,中子注量率和反应堆功率的准确预测对反应堆安全运行至关重要,而现有的本征正交分解(POD)与Galerkin投影相结合的方法存在累积误差而导致精度不高的问题。使用隐式差分法得到一维中子时空扩散的精确解,并作为基准数据,引入2个长短期记忆(LSTM)神经网络项,用于降低POD的累积误差和截断误差,实现物理驱动和数据驱动的混合驱动模型的构建。结果表明,添加神经网络修正项后,对中子注量率、总功率和各阶模态系数预测的均方根误差(RMSE)均降低了1~2个数量级,添加神经网络扩展项后,在预测相同阶数情况下计算时间显著减小,基于2阶和3阶扩展到6阶的改进模型相较于原始6阶模型分别提速了13%和7.6%。混合驱动模型可以很好得改善POD快速预测精度,结果有一定的参考价值。
-
关键词:
- 本征正交分解(POD) /
- Galerkin投影 /
- 长短期记忆(LSTM)神经网络 /
- 降阶模型 /
- 中子注量率预测
Abstract: The accurate prediction of neutron flux and reactor power is very important for the safe operation of the reactor immediately after the disturbance of reactor parameters. The traditional method combining POD and Galerkin projection has the problem of low accuracy due to cumulative error. In this study, the implicit difference method is used to obtain the exact solution of one-dimensional neutron spatiotemporal diffusion. As the reference data, two LSTM neural network terms are introduced to eliminate the cumulative error and truncation error of POD, and to build a hybrid drive model driven by physics and data. The results show that the root-mean-square error of neutron flux, total power and each order modal coefficient is reduced by 1-2 orders of magnitude after adding the neural network correction term, and the calculation time is significantly reduced under the same order of prediction when the neural network extension term is added. The improved model based on 2nd and 3rd order scaling to 6th order is 13% and 7.6% faster than the original 6th order model, respectively. The hybrid drive model can improve the rapid prediction accuracy of POD, and the results have certain reference value. -
表 1 验证算例数据
Table 1. Data of Verification Examples
ν/(m·s−1) D/m 中子扩散长度/m a/m k∞ ϕ0/(cm2·s)−1 2.2×103 0.211×10−2 45.866 2 1.001 cos(2πx/a)+1 表 2 程序误差
Table 2. Results of Code Error
时间/s RMSE MAE 0.05 0.00207 0.00205 0.1 0.00206 0.00263 0.15 0.16371 0.14529 0.2 0.25782 0.22907 RMSE—均方根误差;MAE—平均绝对误差 表 3 物性参数
Table 3. Physical Property Parameters
g ν/(cm·s−1) χg Dg /cm Σɑ, g/cm−1 (νΣf)g /cm−1 Σs,1→2/cm−1 1 107 1 1.4 0.01 0.007 0.1 2 2×105 0 0.4 0.15 0.2 Σs,1→2—快群移出截面 表 4 缓发中子先驱核参数
Table 4. Delayed Neutron Precursor Parameters
参数名 群数 1 2 3 4 5 6 λ/s−1 0.0127 0.0317 0.1150 0.3110 1.4000 3.8700 β 0.000247 0.001385 0.001222 0.002645 0.000832 0.000169 λ—缓发中子先驱核衰变常数;β—缓发中子份额 表 5 POD模型不同阶数重构下的MAE
Table 5. MAE of POD Model under Different Order Reconstruction
群名 MAE/10−3 前1阶 前2阶 前3阶 前4阶 前5阶 前6阶 快群 3.7 3.1 1.1 1.0 1.1 0.98 热群 5.4 4.2 0.8 0.85 0.86 0.89 表 6 LSTM神经网络超参数列表
Table 6. List of LSTM Hyperparameters
神经网络项 优化器 损失函数 隐藏层数 学习率 迭代次数 修正项 Adam MSE 190 0.01 200 扩展项 Adam MSE 120 0.01 150 表 7 不同模型各阶RMSE
Table 7. RMSE of Each Order for Different Models
阶数 6阶 1阶修正+
5阶扩展2阶修正+
4阶扩展3阶修正+
3阶扩展6阶修正 1 1.5×10−1 1.8×10−3 1.8×10−3 1.8×10−3 1.8×10−3 2 7.5×10−5 4.2×10−2 3.0×10−6 3.0×10−6 3.0×10−6 3 1.4×10−4 5.2×10−2 3.2×10−3 1.1×10−6 1.1×10−6 4 4.8×10−6 4.9×10−3 1.2×10−4 7.1×10−5 3.2×10−8 5 1.2×10−5 8.7×10−3 1.4×10−4 1.3×10−5 9.2×10−8 6 9.5×10−6 4.3×10−3 2.4×10−5 1.1×10−5 2.8×10−8 平均 2.5×10−2 1.9×10−2 8.8×10−4 3.1×10−4 3.0×10−4 -
[1] 龚禾林,陈长,李庆,等. 基于物理指引和数据增强的反应堆物理运行数字孪生研究[J]. 核动力工程,2021, 42(S2): 48-53. [2] 寇家庆,张伟伟. 动力学模态分解及其在流体力学中的应用[J]. 空气动力学学报,2018, 36(2): 163-179. [3] 丁鹏,陶文铨. 建立低阶模型的POD方法[J]. 工程热物理学报,2009, 30(6): 1019-1021. doi: 10.3321/j.issn:0253-231X.2009.06.032 [4] 羊俊合. 中子输运计算的OpenFOAM应用及其降阶研究[D]. 哈尔滨: 哈尔滨工业大学,2019. [5] VERGARI L, CAMMI A, LORENZI S. Reduced order modeling approach for parametrized thermal-hydraulics problems: inclusion of the energy equation in the POD-FV-ROM method[J]. Progress in Nuclear Energy, 2020, 118: 103071. doi: 10.1016/j.pnucene.2019.103071 [6] 张伟伟,朱林阳,刘溢浪,等. 机器学习在湍流模型构建中的应用进展[J]. 空气动力学学报,2019, 37(3): 444-454. [7] 张译文,王志恒,邱睿贤,等. 利用长短期记忆神经网络的改进POD-Galerkin降阶模型及其在流场预测中的应用[J]. 西安交通大学学报,2024, 58(2): 12-21. doi: 10.7652/xjtuxb202402002 [8] ZHANG X. The research of method for core on-line monitoring[J]. Nuclear Science and Technology, 2017, 5(4): 216-225. [9] 康伟,张家忠,李凯伦. 利用本征正交分解的非线性Galerkin降维方法[J]. 西安交通大学学报,2011, 45(11): 58-62,67. [10] 谢海润,吴亚东,欧阳华,等. 基于本征正交分解和动态模态分解的尾涡激振现象瞬态过程的模态分析[J]. 上海交通大学学报,2020, 54(2): 176-185. [11] PARISH E J, RIZZI F. On the impact of dimensionally-consistent and physics-based inner products for POD-Galerkin and least-squares model reduction of compressible flows[J]. Journal of Computational Physics, 2023, 491: 112387. doi: 10.1016/j.jcp.2023.112387 [12] 贾续毅,龚春林,李春娜. 基于POD和BPNN的流场快速计算方法[J]. 西北工业大学学报,2021, 39(6): 1212-1221. doi: 10.3969/j.issn.1000-2758.2021.06.006 [13] 冀南,杨俊康,赵鹏程,等. 耦合多变量LSTM与优化算法的铅铋反应堆事故参数预测方法研究[J]. 核动力工程,2023, 44(5): 64-70. [14] CHEN Z F, LIU Z J, JI N, et al. Accident parameter prediction method for lead-bismuth cooled reactor based on a multivariate LSTM network coupled with an optimization algorithm[J]. Annals of Nuclear Energy, 2023, 193: 110027. doi: 10.1016/j.anucene.2023.110027 [15] 史建楠,邹俊忠,张见,等. 基于DMD-LSTM模型的股票价格时间序列预测研究[J]. 计算机应用研究,2020, 37(3): 662-666.