제철소 부생 가스 보일러에 대한 차수 축소 모델 구축

Abstract

Reduced order models (ROMs) were constructed to predict the scalar fields in real time for a POSCO's industrial byproduct gas boiler. A temporary ROM was constructed using the proper orthogonal decomposition (POD) known as principal component analysis (PCA) and the regression by Kriging. The top 100,000 cells were selected in the order of the highest error predicted by the temporary ROM. A variational autoencoder (VAE) was trained on selected 100,000 cells, and the cells were reconstructed. Cell data except for selected cells were predicted using POD basis vectors obtained by performing POD and Gappy-POD which is known as a sparse data reconstruction technique. Six normal flow rates were selected as operating variables for ROM construction and a 6D parameter space was obtained. Latin hypercube sampling (LHS) and adaptive sampling were used as sampling methods, and a total of 70 training sample points were extracted. Prediction errors were compared for three types of ROMS by VAE with Gappy-POD (VAE-GPOD), POD with Kriging (POD-Kriging) and POD with deep neural network (POD-DNN). VAE-GPOD showed the best performance and a prediction accuracy evaluated by a R2 score for scalar fields was 0.94 ~ 0.97. As a result of the study, ROMs were constructed to predict the temperature, CO2, CO and O2 of the internal field in real time for any operating conditions in the 6D parameter space.