Hybrid Strategy Integrating Variable Selection and a Neural Network for Fluid Catalytic Cracking Modeling

Different from traditional modeling methods for maximizing iso-paraffins (MIP), a hybrid approach that integrates the least absolute shrinkage and selection operator (LASSO) method for variable selection and an output-focused back-propagation neural network (BPNN) method for predictive model construction is proposed in this paper. LASSO was used to reduce the dimensionality of the factors influencing the yield and property of products and eliminate the correlations among factors to obtain the feature variables that were used as BPNN input vectors. The combined LASSO-BPNN models were trained and tested using industrial production historical data and were further compared with principal component analysis (PCA)-BPNN models and BPNN models. The selection results of the LASSO method, which are superior to the results attained according to traditional knowledge and experience, quantitatively show that the production information varies in its effect on product yields and properties. The prediction results of a validation dataset analyzed by comparing model values with industrial values indicate that the intelligent LASSO-BP models have good prediction accuracy. Comparative results of the average relative errors show that the LASSO-BPNN models have the highest generalizability among LASSO-BP, PCA-BPNN, and BPNN models. Sensitivity analysis results of feed carbon residue content and operating conditions indicate that the combined LASSO-BP neural network models effectively predicted the yields and properties of MIP products and provided a good reference for industrial MIP process optimization.