Construction and Analysis of Octane Number Loss Prediction Model

In this research work, we introduce an octane loss prediction model to solve the problems of difficult product quality control and untimely process optimization response in the catalytic cracking process. A decision tree algorithm is first used to regress the raw data to generate an optimal cut-off variable and cut-off point. Then, the variables with high correlation are screened according to Pearson coefficients. A four-layer BP neural network was constructed to predict octane loss. The model was validated by a cross-validation method, with 70% of the data selected as the training set and 30% of the data selected as the test set. The experimental data show that the mean error (MAE) of the prediction results is 16.23%, the mean squared error (MSE) is 4.27%, and the root means squared error (RMSE) is 20.66%. Finally, the model was optimized by the particle swarm algorithm to obtain the final operating ranges of the main operating variables. The model has a high degree of fit for the prediction of the target values, which helps to optimize the operating conditions and improve the quality of gasoline.