The Long Beach Convention Center
Long Beach, CA
1:30pm, April 27th

Brian Rohrback
Infometrix, Inc.
Will Warkentin
Chevron Richmond Refinery

KEYWORDS
Best Practices, Calibration, Cloud Computing, Database, Gasoline Blending, Optical Spectroscopy, PLS, Process Control

ABSTRACT
Optical spectroscopy is a favored technology to measure chemistry and is ubiquitous in the hydrocarbon processing industry. In a previous paper, we focused on a generic, machine-learning approach that addressed the primary bottlenecks of mustering data, automating analyzer calibration, and tracking data and model performance over time. The gain in efficiency has been considerable, and the fact that the approach does not disturb any of the legacy (i.e., no changes or alterations to any analyzer or software in place) made deployment simple.

We also standardized a procedure for doing calibrations that, adheres to best practices, archives all data and models, provides ease of access, and delivers the models in any format. What remains is to assess the speed of processing and the quality of the models. To that end a series of calibration experts were tasked with model optimization, restricting the work to selecting the proper samples to include in the computation and setting the number of factors in PLS.  The amount of time and the quality of the models were then compared.  The automated system performed the work in minutes rather than hours and the quality of the predictions at least matched the best experts and performed significantly better than the average expert.  The conclusion is that there is a large amount of recoverable giveaway that can be avoided through automation of this process and the consistency it brings to the PLS model construction.

INTRODUCTION
There is a lot of mundane work tied to the assembly of spectra and laboratory reference values to enable quality calibration work.  There is also insufficient guidance when it comes to the model construction task.  How much time should be spent on this task?  How to best assess whether a spectrum-reference pair is an outlier or not? How many cycles of regression-sample elimination make sense? Where do we switch over from improving the model by adding PLS factors to overfitting and incorporating destabilizing noise?

For more information or the full paper, contact us.