An effective method based on multi-model fusion for research octane number prediction
Abstract
The octane number is one of the important indicators in crude oil processing, and it is related to the anti-knocking performance of gasoline engines. The reduction of the octane number in the gasoline refining process is closely related to the economic benefits. According to the actual needs of chemical production, a research octane number (RON) prediction model combining a random forest algorithm, a BP neural network and a genetic algorithm is proposed. First, a constrained random sample selection strategy and method is designed, then this is combined with the sulfur content and octane content, using an improved random forest algorithm to screen the main operating variables. This is then combined with the BP neural network to optimize the hidden layer nodes to reduce overfitting, and the octane number and residual value are predicted. After that, the main operating variables are optimized through the genetic algorithm based on the constrained fitness function to make the average drop loss of all samples 34.7%. It has been verified by practice that the prediction and optimization of the model are effective and meet the requirements of actual industrial applications.