Application of Non-Intrusive Reduced Order Models for Turbulent Reacting Devices

Abstract

Data-driven surrogate modelling has received significant attention over the past decade in various fields. Surrogate models with the machine learning (ML) technique are applied to develop digital applications or solve optimization problems. These techniques also have been applied in combustion applications of simple laminar and turbulent jet flames, oxy-pilot coal boiler and flameless furnace. Digital twins in energy facilities using surrogate models are utilized by training 2D or 3D image or simulation data for model order reduction to reduce the computational time. This is called a reduced order model (ROM), which instantly gives an internal solution of industrial devices. In this dissertation various ROMs are developed for turbulent reacting devices in the non-intrusive way. First, ROMs are constructed by proper orthogonal decomposition (POD) and regression by Kriging and Radial Basis Neural Network (RBFN) for a 500 MWe tangentially fired pulverized coal boiler. POD is performed toextract low-dimensional basis vectors to reproduce 3-D distribution of reacting scalars with respectto the operation parameters of total secondary air (TSA) and burner zone stoichiometric ratio (BSR).The ROMs by Kriging and RBFN both reproduced the scalar fields within 6% averaged relative L2 normerror at three validation points in the parameter space. It is possible to reproduce a 3-D scalarfield at any unexplored operation condition within a few seconds through parallel computation of the ROM. It allows fast evaluation of the effects of different operation parameters in the design stage and real time response of a digital twin based on the ROM for smart operation and maintenance of industrial combustion facilities. Second, the convolutional autoencoder (CAE) is applied to the ROM fora turbulent methane jet flame. Autoencoder is a machine learning algorithm, which reduces the problem dimension by non-linear projection. It has an advantage in reconstruction of significantly non-linear data. Additionally, the characteristics of original data can be trained using a convolutional layer with a relatively small number of hyper-parameters. To check accuracy of the ROM by CAE, we applied it to surrogate model and sparse reconstruction problem, and compared it with other dimension reduction techniques. For model training, five parameters are selected as the model training parameters and 20 and 40 sensor data are extracted for a sparse reconstruction problem. The proposed CAE showed better accuracy than the linear projection-based dimension reduction technique. Finally, hybridROMs are developed with variational autoencoder (VAE) models and gappy interpolation (GI) from three-dimensional CFD results for a semi-industrial scale coal boiler. Non-linear prediction near the burner section is performed through the VAE with the remaining domain covered by the GI method. The VAE-GI results for the validation set showed the averaged relative L2 norm error of 0.36% for temperature, 2.58% for O2 mass fraction, 3.26% for CO mass fraction and 0.92% for velocity magnitude. It outperformed the proper orthogonal decomposition (POD) method in all scalar predictions, showing differences of 1.0%, 9.2%, 12.2% and 4.6%, respectively. The following results show that the VAE-GI methodcan improve the prediction performance especially in prediction of the species mass fractions fora reacting device.