Integrating Neural Networks with Mathematical Programming for Gasoline Blend Formulation

This paper tackles a key refinery-economics problem: optimizing component blends to produce RON-92 and RON-95 motor gasoline (AI-92/AI-95 in Russian practice). The task is challenging due to nonadditive physicochemical effects, a large pool of candidate components, and variability in feedstock composition. We explicitly formulate a mathematical model of gasoline compounding as a constrained optimization problem, where the decision variables are the component shares in each product and the objective minimizes total leftovers under material-balance, component-availability, and recipe-window bounds. This optimization model is coupled with feed-forward neural-network validators that assess GOST compliance; the classifiers are integrated either as hard feasibility constraints or via the objective function, turning the baseline linear program into a nonlinear mathematical model that captures nonadditive interactions. The workflow comprises: (i) a linear program that minimizes leftovers; (ii) dataset construction and training of grade-specific classifiers; and (iii) two integration schemes for the classifiers in the optimizer. We implemented a real-time Python application with a graphical interface that generates formulations for RON-92-K5 and RON-95-K5 (AI-92-K5/AI-95-K5) from up to 15 components. On refinery data, the classifiers achieved high precision (97–100%), and the resulting formulations were robust and operationally realistic.