一种基于变分自编码器的核反应堆堆芯功率降阶计算方法OA

Three-dimensional relative power distribution is an important physical quantity for characterizing the neutronic state of a reactor core and the spatial uniformity of core power.However,the high-fidelity calcu-lation of three-dimensional relative power distribution heavily relies on large-scale numerical simulations and is computationally expensive,which makes it difficult to satisfy the demands of rapid evaluation,repeated-query tasks and uncertainty quantification in practical applications.To address this issue,a nonlinear reduced order method is developed for the reconstruction and uncertainty characterization of relative power fields by combining the truncated linear modal representation with the variational autoencoder.In the offline phase of the method,high-fidelity snapshot fields are first projected onto a truncated modal space by proper orthogonal decomposition(POD),to obtain the low-dimensional modal coefficients that retain the dominant physical structures of the original power field.Then,a variational autoencoder is employed to learn the nonlinear latent distribution of the modal coefficients,such that the high-dimensional relative power field can be represented and reconstructed in a more compact latent space.On this basis,a parameter-to-latent prediction network is further trained to establish the mapping from physical parameters to latent variables,and the reconstructed power field is obtained through the decoder together with the truncated modal basis.For the online application of the method,by inputting the parameters of physical model,the distribution of latent variables is first ob-tained and then sampled in the latent space.The method is able to propagate uncertainty from the latent repre-sentation to the reconstructed power field,thereby providing uncertainty estimates and confidence intervals for the prediction results.Simulation experiments are carried out for the steady-state three-dimensional rela-tive power distribution of the HPR1000 reactor core of Hualong One,where high-fidelity samples are gener-ated by CORCA-3D and 5％Gaussian noise is added to test the robustness of the method.It is shown that,in comparison with traditional linear reduced order modelling and neural-network-based methods,the method can maintain high reconstruction accuracy with much lower latent-space dimension,the relative error can be controlled close to 1％in the representative settings,while the uncertainty estimates are generally consistent with the local prediction quality and can reflect the reliability of the reconstructed results by providing the cor-responding confidence intervals.These results demonstrate that the method provides an effective approach for rapid reactor-core power restruction and uncertainty quantification.