Optimization of Hydrotreating Unit Operating Mode Control Based on Machine Learning Models and a Heuristic Method

Optimizing the operating modes of hydrotreating units is difficult because process parameters are numerous and interdependent, and product-quality indicators are often available only as fuzzy expert assessments. This study integrates fuzzy mathematical modeling, a heuristic decision-making method, and machine learning into a unified framework for controlling a hydrotreating unit. A system of models was developed for the reactor, three columns, and the furnace, combining nonlinear regression equations for measurable parameters, fuzzy regression equations for quality indicators, and a heat balance equation. The reactor control problem with fuzzy quality constraints was solved using the heuristic AorRA+PO method (absolute or relative allowance with Pareto optimality), which takes the decision-maker’s priorities into account. In parallel, eight regression algorithms were compared on a synthetic dataset, and the best gradient-boosting models (R² = 0.989–0.992) were used as surrogates in a two-stage optimization. The fuzzy approach increased the hydrotreated product volume to 78.0 m³/h, which is 2.0% higher than the deterministic method and 0.6% higher than the production mode; the machine learning approach yielded 79.6 m³/h with the quality constraints satisfied to a degree of 0.96–1.00. The lower reactor temperature in the recommended mode also reduces the furnace heat duty by about 2.5%. Because the machine learning models were trained on synthetic data, the recommended mode requires experimental verification at the unit.