分类： 物理学 >> 核物理学 提交时间： 2024-01-02
摘要： The aging of operational reactors leads increased mechanical vibrations of reactor internals. The vibration of the in-core sensors near their nominal locations is a new issue for the neutronic fields reconstruction. Current field reconstruction methods fail to handle spatially moving sensors. In this work, we proposed a Voronoi tessellation techinque in combination with convolutional neural networks (V-CNN) to handle this challenge. The observations from movable in-core sensors are projected to the same global field structure, this projection is achieved with Voronoi tessellation, holding the magnitude and location information of sensors. The general convolutional neural networks were used to learn the map from observations to the global field. The proposed method is able to reconstruct the multi-physics fields (e.g., the fast flux, thermal flux and power rate) using observations from single field (e.g., thermal flux). Numerical tests based on IAEA benchmark proved its potential for real engineering usage, particularly, within an amplitude of 5 cm around nominal locations, the field reconstruction leads to average relative errors below 5 % and 10 % in $L_2$ norm and $L_{ infty}$ norm, respectively.
分类： 核科学技术 >> 辐射物理与技术 提交时间： 2024-05-08
摘要： Accurate and efficient online parameter identification and state estimation are crucial for leveraging Digital Twin simulations to optimize the operation of near-carbon-free nuclear energy systems. In previous studies, we developed a reactor operation digital twin (RODT). However, non-differentiabilities and discontinuities arise when employing machine-learning-based surrogate forward models, challenging traditional gradient-based in verse methods and their variants. This study investigated deterministic and metaheuristic algorithms and developed hybrid algorithms to address these issues. An efficient modular RODT software framework that incorpo rates these methods into its post-evaluation module is presented for comprehensive comparison. The methods were rigorously assessed based on convergence profiles, stability with respect to noise, and computational performance. The numerical results show that the hybrid KNNLHS algorithm excels in real-time online applications, balancing accuracy and efficiency with a prediction error rate of only 1% and processing times of less than 0.1 s. Contrastingly, algorithms such as FSA, DE, and ADE, although slightly slower (approximately 1 s), demonstrated higher accuracy with a 0.3% relative L2 error, which advances RODT methodologies to harness machine learning and system modeling for improved reactor monitoring, systematic diagnosis of off-normal events, and lifetime management strategies. The developed modular software and novel optimization methods presented offer pathways to realize the full potential of RODT for transforming energy engineering practices.
分类： 物理学 >> 核物理学 分类： 数学 >> 计算数学 提交时间： 2023-09-29
摘要： Machine learning-based modeling of reactor physics problems has attracted increasing interest in recent years.Despite some progress in one-dimensional problems, there is still a paucity of benchmark studies that are easyto solve using traditional numerical methods albeit still challenging using neural networks for a wide rangeof practical problems. We present two networks, namely the Generalized Inverse Power Method Neural Net#2;work (GIPMNN) and Physics-Constrained GIPMNN (PC-GIPIMNN) to solve K-eigenvalue problems in neu#2;tron diffusion theory. GIPMNN follows the main idea of the inverse power method and determines the lowesteigenvalue using an iterative method. The PC-GIPMNN additionally enforces conservative interface condi#2;tions for the neutron flux. Meanwhile, Deep Ritz Method (DRM) directly solves the smallest eigenvalue byminimizing the eigenvalue in Rayleigh quotient form. A comprehensive study was conducted using GIPMNN,PC-GIPMNN, and DRM to solve problems of complex spatial geometry with variant material domains fromthe field of nuclear reactor physics. The methods were compared with the standard finite element method. Theapplicability and accuracy of the methods are reported and indicate that PC-GIPMNN outperforms GIPMNNand DRM.
点击量
通过
点击量
下载量
评论
