Model order reduction by radial basis function network for sparse reconstruction of an industrial natural gas boiler
Jinwoo Park, Woojin Lee, Kang Y. Huh
Abstract
The radial basis function network (RBFN) is compared with gappy interpolation for sparse reconstruction of a reduced order model (ROM) for an industrial natural gas boiler. It is a non-intrusive method based on proper orthogonal decomposition (POD) and sensor measurements substituted by full order model (FOM) results at specified locations. The FOM was formed by steady-state computational fluid dynamic simulation at training and validation points selected by Latin hypercube and adaptive sampling in the 2D parameter space. Parametric study was performed for a varying number of sensors located on the boiler walls as a realistic measurement option. The optimal numbers of training samples and truncated eigenmodes were determined according to the relative L2 norm error between FOM and ROM by the truncated eigenmodes. Results showed RBFN outperforming gappy interpolation with a lower relative L2 norm error and less dependence on the number of sensors for both temperature and nitric oxide concentration fields. The RBFN may be a better choice for reconstruction of multidimensional scalar fields as a digital twin through fusion of simulation and online data for smart operation of complicated thermofluid facilities.