Multi-Objective Optimization of Catalytic Reforming Using Deep Neural Network Surrogate Modeling

Abstract

Process optimization in catalytic reforming is essential to maximize high-octane reformate yield and reduce energy consumption while maintaining reformate Research Octane Number (RON) as a constraint. However, process optimization initiative for this unit encounters several challenges, such as complex reaction mechanism, large option for decision variables, and conflicting objective functions. Moreover, characteristics of naphtha are very variative from different oil fields and different operating parameters of crude distillation unit (CDU), so they are difficult to generalize. This study intends to optimizes the energy costs and the reformate yield for the reactor and purification section by considering different feed characteristics of naphtha. A detailed first-principles model of the reforming and separation sections was constructed using UniSim Design, and 200 simulated operating points were generated for training and testing the surrogate. The DNN achieved strong predictive accuracy (MSE < 0.354, MAE < 0.0467, R² > 0.93), successfully replicating the behavior of the reactor and downstream separation units. The framework simultaneously minimized utility cost and maximized reformate yield, subject to the constraint of RON ≥ 92. SPEA2 produced 11–19 Pareto-optimal solutions across three distinct naphtha feeds. Visualization through parallel-coordinate plots and correlation heatmaps revealed key operational trends, including the positive influence of H₂/HC ratio, flash pressure, and boil-up ratio on yield, and the negative effect of high reactor and flash temperatures. The results demonstrate that combining surrogate modelling with evolutionary optimization provides a robust, computationally efficient approach for identifying optimal operating windows and understanding trade-offs in catalytic reforming units.