Single-channel Temperature Prediction of Heat Pipe Reactor Based on Deep Neural Network
-
摘要: 热管反应堆因其设计独特性和高效的热传导性能成为核能发电的有力候选者。然而，其堆芯温度场的准确监测仍是关键挑战。本文基于深度学习技术，探索了一种全新的堆芯温度快速预测方法。通过建立反向传播神经网络（BPNN）模型，训练大量堆芯模拟数据，可实现利用6个温度测点预测堆芯单通道截面温度场。BPNN模型训练结果表明，选择合适的神经元数量和隐藏层层数，可有效提高预测精度并且减少过拟合风险。本研究的BPNN模型在测试集上的平均绝对误差为1.06 K，显示出较好的预测能力和较低的误差水平，且误差较为集中在角燃料棒以及换热剧烈的区域。
-
关键词:
- 热管反应堆 /
- 固态堆芯 /
- 深度学习 /
- 反向传播神经网络（BPNN） /
- 温度预测
Abstract: Heat pipe reactors have become a strong candidate for nuclear power generation due to their unique design and efficient heat conduction performance. However, accurate monitoring of the core temperature field remains a key challenge. This paper explores a novel method for rapid prediction of the core temperature field based on deep learning technology. By establishing a backpropagation neural network (BPNN) model and training a large amount of core numerical simulation data, it is possible to predict the temperature field of a single channel core section using 6 temperature measurement points. The training results of the model show that selecting the appropriate number of neurons and hidden layers can effectively improve prediction accuracy and reduce the risk of overfitting. The neural network model in this study has an average absolute error of 1.06 K on the test set, demonstrating good predictive ability and a low level of error. Errors are primarily concentrated in corner fuel rods and regions with intense heat exchange. -
表 1 不同神经网络结构测试集的$ L\mathrm{_{MAE}} $
Table 1. $ L\mathrm{_{MAE}} $ of Test Sets of Different Neural Network Structures
神经元数量 $ L\mathrm{_{MAE}} $/K 隐藏层数1 隐藏层数2 隐藏层数3 隐藏层数4 隐藏层数5 隐藏层数6 32 11.59 4.44 1.51 1.15 1.06 1.07 64 5.04 1.18 1.06 1.06 1.07 1.06 128 1.80 1.06 1.06 1.06 1.06 1.07 256 1.12 1.06 1.06 1.05 1.07 1.05 512 1.06 1.06 1.06 1.07 1.08 1.06 
下载: 导出CSV
表 2 神经网络超参数设置
Table 2. Neural Network Hyperparameter Settings
参数 参数值 参数 参数值 输入层维度 6 学习率 3×10−5 输出层维度 8538 优化器 Adam 隐藏层层数 3 训练批次大小 32 每层神经元数量 256 正则化系数 0 Dropout 0 训练批次 200
下载: 导出CSV
-
[1] 马誉高,刘旻昀,余红星,等. 热管冷却反应堆核热力耦合研究[J]. 核动力工程,2020, 41(4): 191-196. doi: 10.13832/j.jnpe.2020.04.0191. [2] 薛秀丽,杨红义,杨福昌. 中国实验快堆堆芯出口区域温度监测点布置计算验证[J]. 原子能科学技术,2008, 42(5): 428-433. [3] BAE J W, RYKHLEVSKII A, CHEE G, et al. Deep learning approach to nuclear fuel transmutation in a fuel cycle simulator[J]. Annals of Nuclear Energy, 2020, 139: 107230. doi: 10.1016/j.anucene.2019.107230 [4] 李仕鲜,刘井泉,沈永刚. 基于神经网络方法的LOCA事故诊断[J]. 核技术,2017, 40(8): 080604. [5] 刘振海,齐飞鹏,周毅,等. 基于机器学习的燃料棒温度分布代理模型构建方法研究[J]. 核动力工程,2023, 44(S2): 1-5. doi: 10.13832/j.jnpe.2023.S2.0001. [6] 胡攀,陈立新,王立鹏,等. 热管冷却反应堆燃料组件稳态热分析[J]. 现代应用物理,2013, 4(4): 374-378. [7] RUMELHART D E, HINTON G E, WILLIAMS R J. Learning representations by back-propagating errors[J]. Nature, 1986, 323(6088): 533-536. doi: 10.1038/323533a0 [8] 陈小前,罗世彬,王振国,等. BP神经网络应用中的前后处理过程研究[J]. 系统工程理论与实践,2002, 22(1): 65-70,88. doi: 10.3321/j.issn:1000-6788.2002.01.010. [9] ZHAO W P, LI J C, ZHAO J, et al. Research on evaporation duct height prediction based on back propagation neural network[J]. IET Microwaves, Antennas & Propagation, 2020, 14(13): 1547-1554. DOI: 10.1049/iet-map.2019.1136. [10] 徐德江,史泽林,罗海波. 零均值归一化互相关跟踪算法特性研究[C]//第九届全国光电技术学术交流会论文集(下册). 北京: 中国航天科工集团公司,2010: 9-13. [11] 黄颖,顾长贵,杨会杰. 神经网络超参数优化的删除垃圾神经元策略[J]. 物理学报,2022, 71(16): 160501. -
图(12) / 表(2)
计量
- 文章访问数: 4
- HTML全文浏览量: 1
- PDF下载量: 1
- 被引次数: 0
