희소벡터 복원 기법 구축을 위한 방사형 기저 함수망 설계 및 방사형 기저 함수의 비교 연구

Abstract

Three radial basis functions (RBFs): Gaussian RBF, multiquadric RBF, and inverse multiquadric RBF, are compared for the radial basis function network (RBFN) as sparse reconstruction of a non-intrusive reduced order model (ROM) based on proper orthogonal decomposition (POD). Steady-state temperature distributions of an industrial-scale gas boiler by computational fluid dynamic (CFD) simulations are applied as training dataset for the ROM. The optimal number of training samples and truncated eigenmodes are selected by adaptive sampling and projection error, respectively. Sparse data of 16 sensors located in the inner wall of the boiler was applied as an input for the RBFNs of this study for the consideration of realistic application. Parameter study was performed for the shape factor on each RBF and reconstruction error are analyzed as the performance of each RBFN. The RBFN with Gaussian RBF showed the best predictability upon the optimal shape factor. Gaussian RBF is recommended for RBFN as sparse reconstruction under the premise of prior search for optimal shape factor.