基于核电厂高加测点数据实时预测模型的样本迁移学习方法研究
Sudy on the Instance-based Transfer Learning Method Based on the Real-time Prediction Model of the High-voltage Heating Measurement Point Data of Nuclear Power Plant
习家轩 1张大志 2肖云龙 3王志会 2周华兵1
作者信息
- 1. 武汉工程大学计算机科学与工程学院,湖北 武汉 430205;武汉工程大学智能机器人湖北省重点实验室,湖北 武汉 430205
- 2. 中核武汉核电运行技术股份有限公司 中核核工业仿真技术重点实验室,湖北 武汉 430040
- 3. 中核武汉核电运行技术股份有限公司,湖北 武汉 430040
- 折叠
摘要
针对核电厂高压加热测点数据实时预测在工程实际中存在故障数据缺失,观测数据量少等问题,提出基于数字孪生仿真数据的样本迁移学习方法.首先,利用数字孪生技术产生的特征丰富、样本充足的核工业相关仿真数据构建源域数据集,利用核工业设备运行过程中各监测点获取的实际观测数据构建目标域数据集;然后,通过交叉验证得到的分类模型,测得源域数据与目标域数据的相似度,将源域数据中与目标域数据相似的数据赋予高权重,使源域数据与目标域数据更匹配;最后,利用带权重的源域数据训练核工业数据实时预测模型.实验结果表明,样本迁移学习方法在小样本与特征缺失的情况下,具有良好的迁移学习效果,并且相较于传统迁移学习方法具有较高的实时预测精度.
Abstract
To address the problems of missing fault data and insufficient observation data in real-time prediction of the high-voltage heating measurement point data of nuclear power plant in engineering practice,an instance-based transfer learning method based on the digital twin simulation data is proposed.Firstly,using digital twin to generate rich features and sufficient samples of nuclear industry-related simulation data to construct the source domain dataset,and using the actual observed data obtained from various monitoring points during the operation of nuclear equipment to construct the target domain dataset.Then,the classification model obtained through cross-validation is used to measure the similarity between the source domain data and the target domain data.Data in the source domain that is similar to the target domain data is given a high weight,making the source domain data more compatible with the target domain data.Finally,a real-time prediction model for nuclear industry data is trained using weighted source domain data.The experimental results show that the instance-based transfer learning method has better performance under small sample and feature missing conditions,and has higher accuracy compared to traditional transfer learning methods.
关键词
核电厂/高压加热/实时预测/数字孪生/样本迁移Key words
Nuclear power plant/High-voltage heating/Real-time prediction/Digital twin/Instance-based transfer learning引用本文复制引用
出版年
2024
NETL
NSTL
万方数据