Multi-Fidelity Predictive Modeling for Residual Oil Hydrotreating Process

Abstract

The residual oil hydrotreating process presents challenges in input–output modeling due to its complex compositions, inaccurate mechanisms, and limited available data sets. Previous efforts indicate that single-fidelity modeling based on first-principles or actual data is inadequate for predicting effluent compositions. This work proposes an improved multifidelity modeling method, termed gradient addition and factor selection based nonlinear Gaussian process (GFNGP), which effectively integrates prior mechanisms and industrial data. By incorporating gradients and selecting factors, GFNGP outperforms the traditional multifidelity nonlinear autoregressive Gaussian process, low-fidelity neural network, and high-fidelity Gaussian process. Taking the low-fidelity neural network as the baseline, GFNGP reduces prediction error by at least 27% across seven output variables. Its robustness and applicability are verified by testing different training sets, yielding median performance improvements ranging from 12% to 64%. Consequently, GFNGP is a practicable modeling strategy for the residual oil hydrotreating process and prompts the petrochemical industry to operate intelligently and efficiently.