MULTIVARIATE STATISTICAL ANALYSIS OF DATA FROM AN INDUSTRIAL FLUIDIZED CATALYTIC CRACKING PROCESS USING PCA AND PLS

Abstract

<p>Principal Components Analysis (PCA) and Partial Least Squares (PLS, or Projection to Latent Structures) were used to evaluate the process history of a fluidized catalytic cracking unit (FCCU). Specifically, the goals of the work were to identify interesting periods in the process history, identify relationships amongst process variables, develop a predictive model of the product yields and selectivities, and to create a monitoring space to detect process changes and disturbances.</p> <p>Major process changes of feed rate, feed quality and production modes were easily modelled by the first few latent variables (LVs) in both the PCA and PLS analyses. Later LVs highlighted transients obvious to operations. Plots of the process behaviour in the space of these latent variables were able to clearly reveal where major changes occurred in the process, implying that this approach is useful for the post analysis of historicaldata bases. Diagnosing the reasons for changes, however, was much more difficult.</p> <p>PLS was quite successful in obtaining predictive models for the product yields and selectivities. A linear model of eleven dimensions was able to predict 81.3% of the cross-validated sum of squares in the Y space and 78.3% of the sum of squares in the X space. The hierarchical PLS approach of Wold et al. (1987) was also applied to the data set and generated results of similar predictive ability and interpretation.</p>