基于粒子群优化BP神经网络的核事故源项反演  

Nuclear accident source term inversion based on BP neural network with particle swarm optimization

作　　者：游清悦 曹博[1,2] 彭丁萍 李中昊 缪学伟 陈洲亮 YOU Qingyue;BO Cao;PENG Dingping;ZHONGHAO Li;MIAO Xuewei;CHEN Zhouliang(School of Nuclear Science and Engineering,North China Electric Power University,Beijing,102206,China;Beijing Key Laboratory for Passive Safety Technology of Nuclear Energy,Beijing,102206,China)

机构地区：[1]华北电力大学核科学与工程学院,北京102206 [2]非能动核能安全技术北京市重点实验室,北京102206

出　　处：《核电子学与探测技术》2025年第3期371-381,共11页Nuclear Electronics & Detection Technology

基　　金：国防基础科研计划资助(JCKY2022110C073)。

摘　　要：核事故发生后,快速准确地估算源物质的释放速率对于提升核应急响应速度及确保决策的可靠性至关重要。本文选择碘-131(^(131)I)核素的释放速率作为源项反演的目标值,利用课题组开发的放射性核素大气扩散模拟程序RADC生成神经网络训练所需的数据集。利用Matlab构建了粒子群算法(Particle Swarm Optimization,PSO)优化误差反向传播(Back Propagation,BP)神经网络的核事故源项反演模型,同时考虑了粒子群算法中超参数和适应度函数的不同对算法优化性能的影响。结果表明:PSOBP模型源项反演测试结果的平均绝对百分比误差为2.14%,平均绝对误差为0.011437,均方差为0.000685,各个评价指标明显优于BP神经网络,验证了该模型的可行性,有助于快速核应急响应。After a nuclear accident,the rapid and accurate estimation of the release rate of radioactive materials is critical for enhancing the speed of nuclear emergency response and ensuring decision⁃making reliability.This study focuses on source term inversion of the release rate of ^(131)I and employs the atmospheric dispersion simulation program RADC,developed by the research team,to generate the dataset for training a neural network.A source term inversion model is established using MATLAB,integrating a Particle Swarm Optimization(PSO)algorithm to optimize the Back Propagation(BP)neural network.The study delves into the impact of parameter optimization,including hyperparameters and the fitness function,on the performance of the PSO algorithm.Results indicate that the PSO⁃BP model achieves an average absolute percentage error of 2.14%,an average absolute error of 0.011437,and a mean squared error of 0.000685 in source term inversion tests,significantly outperforming the BP neural network across all evaluation metrics.This demonstrates the feasibility of the model and its potential for improving the rapid response in nuclear emergency scenarios.

关 键 词：源项反演 BP神经网络 粒子群优化 参数优化 适应度函数 

分 类 号：TL732[核科学技术—辐射防护及环境保护]