Two-stage stacking heterogeneous ensemble learning method for gasoline octane number loss prediction

Abstract Gasoline is the main fuel for small vehicles, and the exhaust emissions from its combustion have a major impact on the atmospheric environment. In the cumbersome process of gasoline refining, the aim is to reduce the sulfur and olefin contents within the raw material while maintaining its octane number as much as possible. In general, the research octane number (RON) loss is measured by using an instrument in the laboratory, which is time-consuming and expensive. Therefore, the use of algorithms to build RON loss prediction models has become a hot topic. Considering that machine learning has a good ability in fitting the non-linear complex data, we propose a stacking-based heterogeneous ensemble method for RON prediction. First, we propose a fusion algorithm of sequence forward search (SFS) and feature importance score to reduce the dimension of data set. Later, the data after feature selection will be used to construct a two-stage stacking heterogeneous ensemble learning model. Finally, the differential evolution (DE) algorithm is used to optimize multiple sensitive parameters involved in the model. Experiments with data obtained in the actual gasoline refining process show that the proposed method can accurately predict the RON loss in the product. Compared to the popular machine learning methods such as support vector machines, random forests, and XGBoost, the proposed method achieves the smallest mean square error. Furthermore, we analyze the important features that affect the RON loss to promote the development of the gasoline refining industry.