A. Barsamian and D. B. C. SON, Refinery Automation Institute, New York City, New York (U.S.)
Analyzers have proven to be crucial to many industries, providing accurate product property measurements and confirming a product’s conformity with specifications and regulations. In the oil and gas industry, one best practice is to check the properties of a final product by using an analyzer online during the inline blending process in real time and within the laboratory.1
Spectroscopic analyzers are very reliable—with very few moving parts (if any)—and may replace numerous different property analyzers due to their modest installation requirements and affordability.2
Many users have bought these multi-property analyzers instruments and, after using them unsuccessfully to control the properties of a blend to certify a shipment, find no further use for them other than to let them gather dust on a shelf. Why?
The problem lies in the nature of the instruments: the translation of a “spectral fingerprint” into an actual desired property (e.g., octane). The most frequently used methods, such as classical chemometrics, require frequent updates involving the gathering of plant information to update the prediction models, model validation, online performance monitoring, etc. These are tedious and expensive processes requiring specialized personnel, which many blenders do not have, at least for the long term.
The use of artificial intelligence (AI) has revolutionized this expensive and time-consuming step, bringing it close to a “plug-and-play” system. This article compares the two approaches—classic chemometrics and AI—highlighting the superiority and practicality of using AI.
The basis of using AI to predict properties. The microcomputer revolution made it possible to economically code large scientific programs in personal computers—e.g., a linear program refinery simulator and embedding AI into well-known mathematical algorithms to predict properties from the spectral fingerprints. The decision on what to use to determine the predicted properties depends on the cost.
There are two contemporary methods to achieve a prediction from spectral fingerprints (the analyzer provides the raw data): chemometrics or various AI methods. As an example, partial least square (PLS) is the cheapest implementation using Microsoft Excel. The more sophisticated systems that utilize improvements in computer technology include AI, which requires more complicated code—naturally, this makes the system more expensive.
An example is the use of AI in the form of statistical analysis and pattern recognition to extract the prediction models. For example, a commercial online analyzera uses embedded AI in the form of topological analysis and a Monte Carlo simulation, a mathematical technique that uses random sampling to predict the likelihood of different outcomes in an uncertain event.
Spectrometers in blending. Spectrometers are popular and effective instruments to determine gasoline blend properties in real time. They are fast, reliable, do not have moving parts and can measure multiple properties. They are much cheaper than conventional analyzers in terms of initial acquisition cost, ongoing spare parts and periodic maintenance.
Their function is to acquire a spectral fingerprint of the test liquid by shining infrared light through gasoline samples. The resulting absorbance pattern is called a “spectral fingerprint” (FIG. 1).
Converting spectral fingerprints to lab values for blend property control. The spectral fingerprint acquired from the spectrometer cannot be used directly, since it does not provide a direct number. Therefore, it requires a stage where the spectral fingerprint is translated into actual properties data (octane numbers, distillation points and others), as shown in FIG. 2.
There are two main methods to this data processing stage: Method 1 uses traditional chemometrics-based regression; while Method 2 uses special fingerprints recognition software using AI to do the recognition.
Chemometrics modeling. The first data processing method is a traditional chemometric modeling method. In this method, each wavelength (in the x-axis) is converted to a number format in terms of the height of the wavelength (in the y-axis).3
The chemometrics method is then used to tie those numbers to match the properties from the lab data (FIG. 3). The chemometric methods include thousands of commercial software embedding modern techniques (partial lest square, principal component regression, etc.).
This proven traditional method has been utilized for many years. Multiple vendors are familiar with this method and have chemometricians ready to provide the service along with the analyzers. Once the model has been set up—and if no drastic changes have been made—it will continue to provide accurate predictions of the properties.
However, the result is a spectrometer time-snapshot frozen in time. Any changes in the process variable value are not captured, and this can introduce an error in the prediction (e.g., octane). When a calibration check fails based on the discrepancy with a reference, it is time to update the chemometric model by adding new spectral fingerprints. The updated model must then be validated.
The frequency of updates depends on the stability of the blending process [e.g., fluidized catalytic cracking (FCC) severity, reformer severity, catalyst change, process changes (adding a new blend component)]. For this reason, every time new changes in recipes, grades, seasons and other variables are introduced, chemometricians must update the model, which costs time and money.
Spectral fingerprint direct match. The spectral fingerprint direct match method is based on the premise that “same fingerprint” means “same properties.” This method uses AI tools, such as Monte Carlo simulation, to modify the raw database to make it easier to use pattern recognition tools, such as topological analysis and neural networks chains.
This method requires a large fingerprint reference database to compare and match raw spectral data with actual laboratory data. In this case, the initial size of the comparison database was > 10,000 samples, and commercial vendors must provide updated versions.
To improve prediction accuracy, some versions of the software include automatic “densification,” (FIG. 4) which combines existing reference samples with Monte Carlo Mahalanobis distance samples. The purpose of the densification is to improve the precision of prediction, which is improved by the square root of independent measurements.
If there is no direct match within the densified database, an additional Monte Carlo simulation run conducts a sophisticated extrapolation, looking for the “nearest statistical neighbors” as the expected answer.
Within the spectral fingerprint method, the spectral and lab-measured properties are entered into a database, building up the reference points of product properties. Using these reference points, it automatically incorporates the new sample in a direct-match database.
Comparing PLS with the AI analyzera. A U.S. refinery and a European refinery set up a premium gasoline blender to compare results from an AI analyzera with results from a commonly used chemometric PLS software (FIG. 5).4
Takeaways. Direct spectral comparison using AI frees the end user from spending time to generate, validate, document and update the database every 6 mos–9 mos, saving time and money. In addition, the self-learning feature of the spectral fingerprint method eliminates the need for chemometricians to be present at the plant. HP
NOTES
FTIR TopNIR online analyzer, a product of TopNIR Systems LLC
LITERATURE CITED
Plock Blending Site Acceptance Report, Plock Refinery, Poland, No. PL610188201/96-1686/ZZC/JN.
“New gasoline-blending unit started at Poland Refining, Oil & Gas Journal, July 5, 1999, online: https://www.ogj.com/home/article/17230811/new-gasoline-blending-unit-started-at-poland-refining
Descales, B., D. Lambert, et al., “Property determination of petroleum refinery products,” U.S. Patent No. 5,712,797, Jan 27, 1988.
Lambert, et al., “TopNIR vs. PLS comparison report,” Hydrocarbon Engineering, May 2006.
ARA BARSAMIAN is the President of Refinery Automation Institute (RAI) LLC in New York City, New York. He has more than 54 yr of experience in gasoline, diesel and biofuels blending operations and technology. He is a fellow of the American Institute of American Engineers for his contributions to blending technology, analyzers and process control computers.
Earlier in his career, Barsamian was a group head with Exxon Research and Engineering Co., President of 3X Corp., and Vice-President of ABB Simcon, all in the area of fuels blending. Barsamian holds BS and MS degrees in electrical engineering from City University of New York.
DANIEL BYEONG CHAN SON is a Project Engineer at Refinery Automation Institute LLC (RAI). He was involved in the feasibility study of blending improvement project for refineries and terminals, an inline blender and online analyzer upgrade study, the implementation of a crude blend compatibility predictor calculator, and participated in the research of the IMO 2020 bunker availability study update. Son is a member of AIChE and holds a BS degree in chemical engineering from New Jersey Institute of Technology.
