GA-BP算法应用于核事故源项反演的研究  

GA-BP neural network algorithm and its application to the nuclear accident source term inversion

作　　者：侯闻宇 凌永生[1] 赵丹[1] 单卿[1] 黑大千[1] 贾文宝[1] HOU Wen-yu LING Yong-sheng ZHAO Dan SHAN Qing HEI Da-qian JIA Wen-bao(College of Materials Science and Engineering, Nanjing University of Aeronautics and Astronautics, Collaborative Innovation Center of Radiation Medicine of Jiangsu Higher Education Institutions, Nanjing 210016, China)

机构地区：[1]南京航空航天大学材料科学与技术学院江苏省高校放射医学协同创新中心,南京210016

出　　处：《安全与环境学报》2016年第6期24-28,共5页Journal of Safety and Environment

基　　金：国防基础科研项目(B2520133007);江苏高校优势学科建设工程项目

摘　　要：为了更准确地估算核事故源项,提高核事故后果评价的准确性,将用遗传算法优化的BP神经网络算法应用于核事故源项反演,改善BP神经网络学习算法应用于核事故源项反演时易陷入局部极小的缺陷。应用Matlab软件编码实现用遗传算法来优化BP神经网络的权值和阈值,将这些优化值赋给网络,得到优化的BP神经网络,即GA-BP神经网络。用7 852组训练数据训练GA-BP神经网络,训练结束后,用500组测试数据评估两种算法的性能。经统计计算得到GA-BP神经网络的平均训练误差为7.98%,小于原始BP神经网络的10.25%;GA-BP的平均测试误差为9.78%,小于原始BP神经网络的10.21%。训练和测试结果表明,经遗传算法优化的BP神经网络误差明显小于原始BP神经网络算法,GA-BP神经网络算法能有效地避免原始BP神经网络容易陷入局部极小值的缺陷,且缩短了训练时间,提高了源项反演的准确度。This paper is aimed at improving the assessment accuracy of the releasing rate of iodine-131 in the source term inversion process by optimizing the backflash propagation （BP）neural network algorithm with the genetic algorithm （GA）. As is known, it is difficult to detect and characterize nuclear accidents due to their nature of uncertainty of the nuclear accident sources and in turn may lead to serious negative effects on the emergency protection measures to be taken. Therefore, it remains a myth on how to obtain the accurate source when it is necessary to assess the consequence of such nuclear accidents, which has aroused broad international concern. It is just starting frmn this need that, in this paper, we have proposed that the optimized GA - BP neural network can help to overcome the inherent defects that may lead to the fall of the assessment and the inefficiency of the BP neural network operating process. Besides, we have also adopted a software named Matlab to optimize the weighing faetors and the thresholds along with all the training and testing calculations by coding the genetic algorithm. Furthermore, the optimized GA- BP neural network has enabled us to turn the three-layer structure into 6 inputs, 65 nodes and 1 output. In addition, we have set up the bacteria population size, the evolutional generation, the crossover probability and the mutation probability at 200, 80, 0. 2 and 0. 1 in the GA - BP neural network, respectively. Furthermore, the performance of the GA - BP neural network has also been tested and assessed via the GA - BP algorithm with the 7 852 sets of data through 100 times of training, with the training error restricted in less than 0.000 01 and 0. 01 rate in every 500 testing sets. Moreover, we have also gained the experimental data based on the international radiation evaluation results along with some real monitoring data from Fukushima accident. It has thus been found that the final training error by using the said GA - BP neural network can be limited to 7.98% , much les

关 键 词：安全学 遗传算法 神经网络算法 GA-BP 源项反演 核事故后果评价 

分 类 号：X945[环境科学与工程—安全科学]