Industrial units modeling using self-attention network based on feature selection and pattern classification

Data-driven modeling method is an effective method for large-scale refinery industrial units. However, due to the long operational cycle, large number of measurement points, complex spatial and temporal correlations and dynamic switching of operation modes of refinery industrial units, utilizing a single data-driven model for process modeling may not be optimal. To tackle with the issues above, a self-attention model based on feature selection and pattern classification (FS-PC-SA) is develop. This approach uses Random Forest algorithm for feature selection and extract different operating modes from industrial data through K-means clustering algorithm. To capture complex spatial and temporal correlations in refinery units, sub-models based on self-attention for each mode are built for prediction. Finally, the method is used to predict the gasoline product yield in fluid catalytic cracking process and wax oil product yield in ebullated bed residue hydrocracking process, which demonstrates that the proposed method is better than other traditional methods.