(345s) An Integrated Machine Learning and Optimization Approach for Octane Optimization

A refinery produces a wide variety of products. Some of them are high value light products such as gasoline, jet fuel and diesel and the others like residual oil can be considered as by products. Among them gasoline is one the most common transportation fuel for cars, trucks etc. The most important quality specification of gasoline is the octane number. It is a measure of the knock resistance and defines the behaviour of gasoline in the engine during combustion. Refineries adjust operation of the units so that octane in the gasoline meets the specifications.

One of the biggest contributors to gasoline pool is isomerization unit. Isomerization is the principal reaction which takes place in Isomerization reactors. It is the conversion of normal paraffins to iso-paraffins so that it increases the octane number of light straight naphtha (feed). The reaction takes place in a hydrogen atmosphere, over a fixed bed of catalyst, and at operating conditions which promote isomerization and minimize side reactions such as hydrocracking.

In this work, our objective is to find the operating conditions that maximizes octane number in isomerization unit in Tupras. The main operating process variable is the reactor inlet temperatures. An upper limit exists for the amount of iso-paraffins which can exist in the reactor product at any given outlet temperature (equilibrium imposed by thermodynamics). As reactor temperature is raised to increase the rate of isomerization, the equilibrium composition will be approached more closely. At very high temperatures, the concentration of iso-paraffins in the product will actually decrease because of the downward shift in the equilibrium curve, even though high temperatures give a higher reaction rate. Therefore, the main question of this work is “what the reactor inlet temperature should be?”

The main operating process variable is determined as the reactor inlet temperature, and it is left constant until a significant change in the feed conditions are observed. However, with this project we achieved to predict the octane number and by using this prediction, we back calculated the optimum reactor temperature. Among the different machine learning methods we employed, Xgboost is identified as the best performing algorithm for this purpose with MAE 0.9. Then, we decrease the reactor temperature by 0.1 Celsius and check the octane value. If the octane value is decreased, then we do it reverse. Our move size for temperature is 0.1. After that we wait for two hours to see the effect. Since this is reactor and there is a chemical and physical delay between octane value sample and reactor. We choose this time delay as two hours however, this is an also decision variable to be calculated. Currently we assumed it to be two hours from the operation experiences. Overall, we achieved setting the reactor temperature value to optimum one. The effect of this change on octane in terms of numeric value is close to 0.5. And it yields to 100.000$/month. The utility cost to change the reactor temperature is close to 25.000 $/month. So overall profit for this project is 75.000 $/month, or 900.000 $/year.