МЕТОД МИНИМИЗАЦИИ ОСТАТКОВ В ПРОЦЕССЕ КОМПАУНДИРОВАНИЯ БЕНЗИНОВ, ОСНОВАННЫЙ НА ИНТЕГРАЦИИ ЛИНЕЙНОГО ПРОГРАММИРОВАНИЯ И НЕЙРОННОЙ СЕТИ

Abstract

Relevance The work is devoted to solving an important economic problem for oil refineries — the selection of the optimal composition of components when compounding motor gasoline brands AI-92 and AI-95, which minimizes residual components during mixing. Optimization of this process is complicated by the deviation from additivity of the physicochemical properties of the components, their large number and the constantly changing composition of the raw materials. Operators responsible for the compounding process typically reproduce blending recipes based on pre-calculated planning data obtained either by linear programming or statistical methods. However, these formulations may be subject to adjustments based on current plant availability. Operators often perform this task manually, relying on their experience. Due to these difficulties associated with the preparation of mixing recipes, the use of mathematical models to automate operator activities is difficult. In this paper, it is proposed to use neural networks to solve this problem, since they are often able to effectively cope with problems that are difficult to algorithmize or traditional mathematical modeling. Aim of research The main aim of the research is to explore a method for selecting the optimal formulation for compounding gasoline of the AI-92 and AI-95 brands in order to minimize residual components, based on the integration of linear programming and neural networks. Research methods The study used compounding data from three different refineries. The data includes 40 mixing recipes for commercial gasoline brands AI-92-K5, AI-95-K5. In total, 15 different components are involved in the mixing recipes, such as: catalytic cracking gasoline, reformate, isomerate, MTBE and others. The first stage of work includes solving the problem of minimizing the remaining components. A linear programming problem with constraints is constructed. The number of restrictions depends on the source data. As a result of solving this problem, mixing recipes will be obtained. However, in order for these formulations to comply with State Standards and other regulatory requirements, they need to be adjusted, since when solving the linear programming problem, the quality indicators of commercial gasoline (octane number, volatility, viscosity, etc.) are not considered. The second stage of work is the use of a neural network (multilayer perceptron) to adjust the recipes obtained from the first stage. For each final product, its own neural network must be created and trained. To create a data set, existing gasoline blending recipes are used that meet all necessary standards. Results The paper proposes a method for solving the problem of creating compounding recipes, combining linear programming and a multilayer perceptron. A method for solving the minimization problem using linear programming methods is described, and a program has been created to automate the search for recipes with minimum remainders. The application of the obtained results for training a multilayer perceptron is described, bringing the compositions closer to real ones, considering regulatory requirements and State Standards.