Local Surrogate Modeling for Spatial Emulation of Gas-Turbine Combustion Via Similarity-Based Sample Processing

Abstract

This work proposes an accurate and efficient surrogate modeling method for predicting combustion field in a gas-turbine combustor. The method integrates proper orthogonal decomposition-based dimensional reduction, and Gaussian process regression, in conjunction with the similarity-based sample processing technique. The design parameters of concern include fuel mass flowrate and swirler vane angle. Global surrogate models (GSMs) based on proper orthogonal decomposition and kriging produce significant errors for spatial emulation of methane concentration and turbulent kinetic energy (TKE), which is found to be largely attributed to the feature disparity of sample data at different design points. The Tanimoto coefficient is introduced to identify the similarity relation of the sample design points. The similarity-based sample processing method leverages the techniques of radial partitioning, azimuthal rotation, and sample similarity clustering to enhance the similarity among samples. The radial partitioning divides the physical fields into subzones according to the peak and trough characteristics along the radial direction. Local surrogate models (LSMs) are then adaptively constructed in the subzones, through azimuthal rotation for the methane concentration field and sample similarity clustering for the TKE field. The results show that the LSMs reduce the average prediction error of the CH4 concentration field from 19.56% to 8.16% and the TKE field from 93.75% to 9.12% compared to the GSMs. The present method can effectively support the surrogate modeling of combustors with complex variations of geometric structures and flow physics.