(196b) Accurate Prediction of Gasoline Octane Number at Molecular Level

With the increasingly fierce competition in the refining industry, the refineries continue to optimize the process route and adjust the product structure through a series of technologies such as oil upgrading, refining and chemical transformation, molecular management, digital intelligence, and so on, which is expected to maximize the economic benefits of the enterprise. At the same time, the molecular composition in the gasoline pool has brought great changes, making the oil blend face severe challenges.

Traditional oil blending is linear blending based on physical properties, which is difficult to accurately predict octane number due to the ignored nonlinear effect between single hydrocarbon molecules (alkanes, cycloalkanes, olefins, aromatics, etc.) in gasoline pool. Therefore, it is necessary to establish an accurate "molecular level octane number prediction model", which can realize the blending precisely, save octane number, reduce production cost and provide strong technical support for refineries to improve quality and efficiency.

At present, there are up to 200 to 300 kinds of single hydrocarbon molecules that can be identified in one oil sample, such as FCC gasoline, reformulated gasoline, etherified gasoline, etc., while the number of oil samples available for analysis in a single production cycle is very small, only dozens of kinds. The quantity variance between the two is an order of magnitude. It forms a mass-feature small-sample research problem, which is difficult to predict by conventional model methods. But machine learning algorithms provide solutions.

In this work, 27 oil samples including single hydrocarbon molecular composition data of different carbon numbers were selected, and three models based on machine learning algorithms of BP neural network, support vector machine and random forest respectively were compared to predict octane number. The model with the smallest absolute deviation would be selected as the optimal model. As the results showed that the absolute deviation range of random forest prediction was the smallest, so the random forest model was selected the optimal model. The prediction accuracy of gasoline octane number at molecular level was 1.5 higher than of traditional physical property method.