有一定物理基础的核素浓度预测神经网络模型  

ANN model for prediction of concentration change of radionuclide based on physical mechanism

作　　者：胡铁松[1] 周彦辰[1,2] 王先甲[3] 

机构地区：[1]武汉大学水资源与水电工程科学国家重点实验室,武汉430072 [2]长江科学院水资源所,武汉430010 [3]武汉大学经济与管理学院,武汉430072

出　　处：《系统工程理论与实践》2016年第1期263-272,共10页Systems Engineering-Theory & Practice

基　　金：国家自然科学基金面上项目(71171151)~~

摘　　要：为解决突发核电事故后使用机理模型预测放射性液态流出物迁移扩散,需长时间迭代计算的不足,提出了一种新型混合神经网络模型,该模型耦合了描述液态流出物在受纳水体中迁移扩散的组分输运方程和神经网络模型,采用并行多种群混合进化粒子群算法计算神经网络权值与阈值.论文以湖北咸宁大畈核电站受纳水体富水水库为研究对象,对事故工况下长半衰期核素迁移扩散进行模拟预测,研究结果表明有一定物理基础的神经网络模型是一种有效、可行的预测模型,预测结果与机理模型的模拟输出拟合度较好,新模型较传统的黑箱神经网络模型以及基于单调型先验知识的神经网络模型具有更强的泛化性能改善能力.It needs long time to predict radioactive contaminant diffusion in receiving water by using mechanism model based on computational fluid dynamics, which is not applicable in emergency situation under accident condition. In order to shorten the computation time, a new artificial neural network model that combines species transport equation which governs contaminant diffusion and neural network model is proposed, and an improved particle swarm optimization algorithm is used to obtain the weight and threshold values of neural network. In this paper, long half-life radionuclide diffusion in Fushui reservoir after a postulated accident happened in Xianning nuclear power station in Hubei Province is studied as a case. The result shows that this proposed model can basically predict the contaminant diffusion trend, and the prediction result fit well with CFD simulation output. Compared with the conventional black box neural network model and the ones with priori knowledge obtained from data monotone, the priori knowledge obtained from equation of physical mechanism is a stronger constrain, which can make the prediction result more close to the simulation output.

关 键 词：水动力学 神经网络 先验知识 物理机理 核电事故 

分 类 号：X591[环境科学与工程—环境工程]