Feature Extraction and Clustering of Feed Oil from a S Zorb Unit Based on AE and PCA Algorithms

Abstract

Based on the 5-year data on the feed oil characteristics obtained from the S Zorb unit, the outliers in the data were detected using the boxplot and LOF methods, and 536 modeling samples were obtained. Combining MIC with the Pearson correlation coefficient, six characteristics of feed oil including RON, sulfur content, olefin content, aromatic content, density, and vapor pressure were chosen as input variables for the clustering model. Two features were extracted from the six variables by the autoencoder (AE) characterized by the 6-32-2-32-6 neural network structure and PCA algorithm for clustering. Three clustering models were built using AE+K-means, PCA+K-means, and K-means. The results of evaluation showed that the optimal clustering number in these models was three, and the AE+K-means model provided the best clustering effect. According to the clustering centers and the property distribution, the dividing boundaries between three types of feed oils are obvious indicating that the AE+K-means model is available to classify feed oils from the S Zorb unit. On this basis, prediction models for the RON of refined gasoline were built for different types of feed oils to get the optimal operation conditions for the reduction of RON losses of refined gasoline in the S Zorb unit.