William T. Welch, Michael L. Bain
Ashland Petroleum Co.
Russell, Ky.
Steven M. Maggard
Fluid Data Inc.
Angleton, Tex.
Jack M. May
Setpoint Inc.
Houston
Careful design of a near-infrared (NIR) analysis C system can optimize reformulated gasoline blending. The lessons learned by Ashland Petroleum Co., and the companies that designed its NIR systems, will help other refiners set up efficient, accurate NIR analyzers.
NIR has gained a great deal of attention recently as an on line process analysis technique. When properly installed, calibrated, and serviced, NIR can offer attractive returns on investment for refiners.
The major operating advantages of NIR include faster analysis time for key blending properties and increased accuracy and precision, which can allow the refiner to blend closer to specification. And NIR analyzers can determine several analytical test results simultaneously.
With the advent of reformulated fuels, blending has become extremely complex as refiners must meet multiple physical and chemical property targets. On-line analyzers can make the timely adjustments required to provide optimized blends that are routinely in compliance the first time.
NIR PROCESS ANALYSIS
Recent experience demonstrates the reliability, accuracy, and precision of a properly calibrated and maintained NIR process analyzer. Ashland Petroleum Co. successfully uses NIR as a closed-loop, feedback control analyzer at its St. Paul Park, Minn., refinery with an on stream time of greater than 99%.
Table 1 shows the 3-year continuous track record at St. Paul.
The design criteria of Fluid Data Inc.'s InfraTane octane analyzer system included the requirement that the system's octane measurements contribute less than 0.02 octane number (ON) to the total error of the measurement. This value was selected because it is much less than the measurement error achievable with an online knock engine system and is therefore insignificant.
The system performance otherwise would be unacceptable for on-line gasoline blending. This is true because the variance of the analyzer is additive to the variance of the reference method (knock engine). The analyzer therefore contributes to the total system error.
The actual magnitude of the analyzer error is a complex subject because of the mathematics involved in manipulating the spectral data to produce data output. For an analyzer to be adequate for a particular measurement, generally it must have absorbance stability with no more uncertainty than the number of significant figures needed in the measured property.
For example, if a signal of 1 absorbance unit (AU) is processed for an octane output (typically 100 ON), the instrument needs an absorbance stability of at least 0.0002 AU to contribute less than 0.02 ON to the total measurement error.
Additional complexity is introduced into the calculation because many factors (such as wavelength stability) affect absorbance stability. And some terms add to the measurement error while others subtract when different terms are summed over the wavelengths used in the mathematical model.
Unless these terms are analyzed in depth, their net effect on the analyzer error frequently is hidden and the magnitude of the error is indeterminable.
The errors mentioned are the maximum errors tolerable over months of operation. Additional complexity is introduced, however, because the samples exhibit different absorbances in different wavelength regions.
The maximum time a system can be available without user intervention is governed by the time it takes for the source (usually a halogen bulb) to bum out. The equipment should remain stable over the entire lifetime of the bulb, but for critical applications such as gasoline blending the bulb generally is used only a fraction of its expected lifetime.
From a process control standpoint, one important factor affecting analyzer design is the optical design of the light collection system. The optical design affects the presentation of the sample to the analyzer. The sample can pass straight through the analyzer for a transmission measurement or be remotely sensed by fiber optics.
One of the main advantages of NIR analyzers is that they are compatible with conventional quartz fiber optics. With fiber optics, light can pass from the source to the fiber optic, then through the sample before it is focused onto the detector. This separates the electronics of the analyzer from the sample, which is desirable, particularly when the sample is a combustible liquid.
Despite the advantages of fiber optics, their use can cause difficulty for the analytical measurement. First, there is an economic incentive to place the fiber in the stream, blend header, or pipeline. This avoids the need for sample conditioning and the cost of installing additional sample piping.
The primary drawback of this placement is that temperature is introduced as an uncontrolled variable. Temperature affects the analytical measurement because the amount of light absorbed by the sample is a function of the density of the material, and hydrocarbons' densities change with temperature.
Temperature also can affect the transmission characteristics of the fiber optic. One reason for this is that the fiber, like the sample, absorbs part of the light, and the magnitude of that absorption is a function of temperature.
Temperature is not an important factor in the transmission of digital signals by fiber optics. The user should be aware, however, that NIR sources produce continuous light emissions in an analog fashion and that the magnitude of the output is important.
Partial compensation can be made for temperature effects, but the effects of temperature can greatly exceed the error of the analyzer. The variance of error introduced by temperature, moreover, is directly additive to the variance of the total system error.
Like the analyzer error, errors due to temperature effects must be controlled at least to the number of significant figures desired in the measured property.
Another consideration is that gasoline streams contain contaminants that may adversely affect the in-line probe. And one last design parameter deserving attention is the type of reference scan made by the analyzer.
An absorbance measurement involves the use of a reference sample to calculate a ratio. Systems that use a stored reference scan, or systems where the reference fiber optic does not experience the same environmental conditions as the sample fiber (length of fiber run, temperature changes, etc.), can add even more error to the absorbance measurement.
This error is in addition to the instability of the power supplies controlling the source and detector and cannot be avoided when a stored reference is used.
DATA TREATMENT
Analyzers utilizing NIR technology are different from most other types of refinery analyzers because they do not directly measure a chemical or physical property but rather establish a correlation between a chemical or physical property and a sample's NIR spectrum.
A wide variety of statistical techniques are available to correlate a property to the sample spectrum, including multiple linear regression (MLR), partial least squares, and principal component regression. These techniques, or a hybrid version of them, are generally used with data pretreatment.
Typical forms of data pretreatment include mean centering, a baseline offset correction, taking a derivative of the raw absorbance data, or a combination of these techniques.
Data pretreatments are designed to remove factors from a sample's spectrum that affect the sample's baseline, reduce spectral variance arising from extraneous factors, or in some cases improve the resolution of overlapping bands present in the sample's NIR spectrum.
The magnitude of a measured error can be a direct result of the statistical treatment used to develop the equation, or it can be caused by the choice of data pretreatment. For example, errors in octane determinations can vary as much as 50%, depending on the statistical method and pretreatment used.
It is preferable, therefore, to use MLR when a direct (or nearly direct) relationship exists between the measured property and the sample's NIR spectrum.
It should be pointed out that normal MLR techniques, such as forward and reverse stepwise regression and all possible pairs, do not attempt to develop the best direct relationship between a property and the spectral data. Rather, then, attempt to develop a correlation based solely on a statistical criterion.
Generally, only a skilled person can develop the calibration equations required for a statistical model to be based on a direct relation-ship.
Octane is a good example of a property for which the relationship between chemical structure and a physical property can be directly correlated. This is true because octane, a physical property, is governed by chemical phenomena such as bond breaking and thermal fragmentation of bonds.
Even though the relationship between the physical property and the chemical phenomena is complex, direct correlation is possible because these chemical phenomena are governed by the chemical structure of molecules, which can be measured directly by the NIR analyzer.
Knocking occurs when glowing hot carbon deposits in the piston cause ignition of the fuel mixture before the piston reaches full compression. Structural groups that enhance octane, such as methine, t-butyl, and benzylic groups, delay preignition by momentarily stabilizing the secondary combustion zone.
The structural groups giving rise to high octane, found in isoparaffinic and aromatic-rich blending components, serve as free radical sinks that momentarily stabilize the secondary combustion zone.
This occurs because these groups are thermodynamically more stable than other types of structural groups present and thereby inhibit free radical propagation of the secondary flame front until full compression is achieved and spark ignition occurs.
Developing an equation that directly models these groups significantly reduces modeling error. Other properties where a direct (or nearly direct) correlation is possible include aromatics, olefins, oxygenate, and benzene contents.
REFORMULATED GASOLINE
The U.S. Environmental Protection Agency (EPA) emission equations for reformulated gasolines use benzene, oxygen, Rvp, sulfur, 90% distillation temperature, olefins, and aromatics.
Specific limits are set for benzene and summertime Rvp. The other properties are used to calculate volatile organic compounds, toxic air pollutants, and NO, emissions using the EPA equations.
The variable use of components such as olefins in the EPA equations provides the refiner some flexibility in optimizing blends. Better knowledge of component properties will allow refiners to consider more possibilities in their blend planning models.
Fluid Data's InfraTane NIR analyzer is capable of accurately measuring most of the properties needed to calculate emissions of volatile organic compounds, toxic air pollutants, and NO. in blend components and finished gasoline blends, as specified by the EPA equations.
Table 2 lists the octane, aromatics, olefins, Rvp, and benzene correlations for conventionally blended gasolines (and the NIR correlation coefficient and standard errors of estimate for each).
Table 3 lists the same chemical and physical NIR property correlations for MTBE-blended gasolines, in addition to a primary NIR method for determining wt % oxygen.
Table 4 lists the NIR correlation coefficients and standard errors of estimate for the octane determinations of ethanol blended gasolines.
Figs. 1 and 2 illustrate the excellent accuracy and precision of the NIR research and motor octane correlations to the ASTM knock engine test methods for conventional, MTBE blended, and ethanol-blended gasolines. Similar results were achieved with road octane Correlations.
These blends were made by varying the blend compositions, octane ranges, volatility ranges, and crude sources to obtain robust models. Except for the ethanol blends, the octane values used in the correlations were an average of results from four knock engine laboratories.
The NIR predicted vs. primary method results are shown in Figs. 3-5, respectively, for benzene, aromatics, and Rvp. The graphs illustrate expected deviations about the regression line, as dictated by the accuracy and precision of the primary methods. Similar results were achieved for olefins analysis.
The GC-Piano (paraffins, isoparaffins, aromatics, naphthenes, olefins) technique is the primary test method used for the determination of benzene, aromatics, and olefins contents of the blended gasolines. The method incorporates state of the art, high-resolution gas chromatography.
The sample is injected into a capillary column-which is temperature-programmed then separated into its components (possibly 400 or more compounds.)
The data-acquisition system and software integrate the area under each peak and normalize the data, after adjusting for detector response factors for each compound. Each peak is then grouped into one of the five Piano hydrocarbon types. Finally, a detailed list of the hydrocarbon groups and each component, in wt % and vol %, is produced.
Rvps were determined with ASTM D-5191, also known as the Grabner method. This method is more rapid and precise than the traditional method.
Fig. 6 shows the NIR predicted vs. the gravimetric oxygen concentration for ethanol and MTBE blends. Future work will include NIR correlations for distillation data, as required by the EPA protocol.
The property targets are measured and recorded simultaneously at Ashland Petroleum's Catlettsburg, Ky., refinery, which has a 105,000 b/d blending operation. The blender capabilities include multivariant property control.
Several refiners have used on-line NIR analysis for many years. Ashland Petroleum's St. Paul refinery has used many of these correlations for conventional and ethanol-blended gasolines in feedback closed-loop control with good results.
Ashland has found the use of a "protofuel" validation system invaluable, especially when measuring multiple targets of gasoline blends.
At the Catlettsburg refinery, three protofuels with different levels of known property targets are used to check the calibration equations of each gasoline product. The system has integrated software, which allows the protofuel sequence to be run on demand or anytime during a blend.
For example, the system can be preset to run before, during, and near the end of the blending operation. If the calibration has changed, as indicated by statistical analyses, the system will correct the equations automatically and produce a printed report. This system, which is an integral part of the NIR analyzer setup, validates the on-line NIR analyzer according to ASTM D-3764.
Protofuels provide a high degree of quality assurance. Under the best conditions, contaminants may deposit to some extent on the probe, the lamp sources may degrade slightly, and the instrument electronics or optical system may change over time. Validation with protofuels will alert the system and automatically correct if any of these factors start to have an effect.
BLEND COMPONENTS
Properties of blend components also correlate well using NIR analysis. Tables 57 list the correlations for octanes, aromatics, olefins, benzene, and temperatures at 50 and 90% distillation (T50 and T90) for blend components.
The correlations are equivalent to or better than the standard error of performance of the primary test method. The NIR correlations also are accurate for T50, T90, and benzene content measurements. Future work will include reporting the distillation data as percentage distilled at 200 and 300 F.
On-line NIR analysis of blend components presents an opportunity to improve blend control by giving frequent real time measurements of critical properties. Ashland plans to use such a system at its Catlettsburg refinery and expects the installation to improve advanced gasoline property control.
The control of gasoline blend properties depends heavily on known or assumed properties of blend components. Whether that control is accomplished by a person, trim controller, or advanced model-based controller, the basis is always an expectation that a certain change in the recipe will give a calculated change in the blend property.
The effect of a component on a blend property is determined by:
- The percentage of the component in the blend
- The property of the blend
- The property of the component
- Nonlinear or interaction effects between various components.
The percentage of each blend component is calculated from flow measurements and is generally known and controlled in real time. The property of the blend is commonly measured by on-line analyzers in near-real time. The other two items on the list are not measured in real time.
Nonlinear or interaction effects are modeled by several methods. Most of these models depend in some way on the property of the component. The portion of these effects not assigned to properties of the individual components is expected to be constant for long periods.
For example, binary-pair octane interaction coefficients, which are assumed to be valid for months or years, are multiplied by the percentages of the components in the blend, then added to linear averages of the component properties.
The component property values used in real time control generally are taken from a data table. The values in the table may be from recent or older laboratory tests, or educated guesses based on experience.
The age and accuracy of these values have a major effect on the control of blend properties.
PROPERTY MEASUREMENT
The most common approach to component property measurement is to sample the component tank, measure the properties, and enter the values in a data table. This can take a day or more and occurs weekly or less often in most plants. At least one refinery, however, samples and tests its component tanks daily.
A second, less common approach is to use results from the streams as they are released from the process units. These results are then averaged or filtered for use in blending. This process makes double use of unit analyzers or routine samples, but the tests frequently exclude important properties such as octane.
If done less often than daily, the tank-sampling method assumes that component properties are reasonably steady over several days. Both approaches assume tanks are well mixed. unfortunately, these assumptions frequently are invalid. Component tanks are often active and usually stratified. An active tank is one into which current production goes at the same time the blend component is drawn off. Tank stratification results from the physical size of the tanks and the common practice of putting inlet and outlet nozzles near each other.
Properties of real components vary constantly. This variability comes from feed changes, unit upsets, day-to-night temperature swings affecting overhead condensers, and the thousand natural shocks to which units are subject.
The reason these assumptions are used is that realistic assumptions do not yield a useful calculation method. Most refiners, therefore, have decided that it is better to have useful numbers based on poor assumptions than no numbers.
ON-LINE MEASUREMENT
The on-line measurement of component properties has long been recognized as desirable, but the costs have made it impractical.
The most important property is octane, the measurement of which with on-line knock engines is slow and capital and labor intensive.
The other most critical properties-Rvp and distillation-are less expensive to measure on-line, but the payback for these is not worthwhile. Octane, therefore, is the key.
One economic incentive for on-line component property measurement is savings of laboratory manpower and resources. Component tank sampling requires someone to climb a tank; collect samples from the top, middle, and bottom; mix the samples; and run a battery of tests. Even with economies of scale, this costs about $200/sample. If six tanks are sampled weekly, the cost is more than $60,000/year. There also can be significant personnel hazard in tank sampling.
NIR analysis may be the breakthrough that will make on-line component property measurement practical. And NIR's speed and flexibility make it economically feasible.
MULTIPLEXING
The use of a single analyzer to measure multiple streams generally has been called "multistreaming" because several streams are brought to the analyzer. It also is possible to bring multiple signals to an NIR analyzer without piping those streams to the analyzer. This typically is called "multiplexing."
For the balance of the discussion, both approaches will be called multiplexing, but will be differentiated by using the terms "optical" and "physical." In optical multiplexing, fiber optic waveguides are brought from multiple sample cells to a single analyzer, which switches among them. In physical multiplexing (multistreaming), multiple streams are piped to the analyzer sample system, which switches among them.
Optical multiplexing is a feature of the analyzer and is accomplished by equipment and software. The system switches the input signal, calibration curve, and output signal between the multiplexed points according to an internal program. The stream's measured, the sequence, and the frequency of measurement can be changed by modifying the program and adding new sample cells as needed, within the limits of the multiplexer.
The advantage of optical multiplexing is that it avoids the lag times associated with sample transport and cell flushing. Some disadvantages of optical multiplexing have been described, including sample conditioning systems, signal attenuation, and fiber optic waveguide alignment.
The final disadvantage is cost. Multiple NIR probes may be more costly than the piping and valves needed for a physically multiplexed system.
Physical multiplexing relies on features of the control system, rather than the analyzer itself, to take advantage of analyzer speed. It is accomplished by equipment and software external to the purchased analyzer system.
In a physically multiplexed system, the DCS (or supervisory computer), switches solenoid or motorized valves to change the sample flowing through a single sample cell. The DCS "tells" the analyzer which calibration to use and is responsible for matching the measurement to the correct stream. The streams measured, the sequence, and the frequency of measurement can be changed by modifying the external tubing and DCS logic as needed, with no fixed limits.
Physical multiplexing has a number of advantages. It can be used with any NIR analyzer, provided only that the DCS can signal it to switch between preloaded calibrations. Physical multiplexing relies on standard sample transport and switching techniques. The analyzer's task thus is kept simple.
Physical multiplexing can be done with a single sample conditioning system. It also can provide more flexibility at a lower cost per stream measured. Finally, a physically multiplexed system can be used with other fast analyzers, such as Rvp -analyzers, if desired.
Physical multiplexing also has disadvantages, such as the lag time associated with bringing a sample to the analyzer and flushing the sample cell.
The time required to transport the sample depends on the location of the analyzer and the arrangement of the piping. The time to flush a well-designed NIR analyzer conditioning system can be as little as 15 sec.
The other disadvantage of physical multiplexing is that custom logic must be developed to do the switching. such switching is beyond the scope of blend ratio controllers. This is not a simple task and must be done with care by skilled personnel.
COMPONENT PROPERTIES
The arrangement of components in a typical in-line gasoline blender lends itself to multiplexed analysis. The component lines generally are close to one another just before they enter the header.
All the streams are at relatively high pressures to allow easy transport, and power and cable trays are in place for the existing controls. An analyzer shelter generally is located nearby.
Two approaches suggest themselves for this type of measurement. In the first, an analyzer can be dedicated to component property measurements. In the second, the existing analyzer or analyzers can be piped to share time between components and the blend header.
In a typical in-line blending system, a fast sample loop draws blended gasoline from the header and makes it available to the analyzers. Each analyzer has a motorized valve to obtain a fresh sample when needed. Spent samples are returned to a sump and pumped back into the fast sample loop, which returns to the header.
Fig. 7 shows a blender with a dedicated component-analysis system. It is the same as a typical blend-sampling system, with the addition of a component sampling system and another NIR analyzer dedicated to measuring component properties. The component sampling system comprises a sample loop and component selector valves to control the selection of the component being sampled.
Fig. 8 shows a shared analyzer system. It is the same as Fig. 7, except the dedicated NIR analyzer has been replaced with a system of sample selector valves to control which sample source (component or blend) is available to each analyzers This approach saves the cost of an analyzer but adds complexity in the sample selection system.
The operation of a component sampling system is straightforward in theory. For example, if Component 2 is to be sampled, its selector valve is opened and it flows through the loop to the sample return system. After a certain time, the loop will contain only Component 2 and will be available to the analyzer.
Of course, the actual piping and selection logic must be designed to prevent cross-contamination of streams by methods such as "double block and bleed" or self-bleeding three way valves. Provisions also may be necessary for on-line calibration samples and the logic they require.
With either a dedicated or shared analyzer system, the order in which components are sampled and the frequency of sampling can be set according to the user's requirements, within the limitations of the analyzers. This might result in a schedule of frequent samples for some components and infrequent samples for less important components. All the control elements are manipulated by the DCS, a personal computer, or a supervisory computer. The flexibility and ease of use are determined by the design of the control software.
In a shared analyzer system, any or all of the analyzers can be switched from the blend sample to the component sample. The analyzer and its conditioning system are flushed with the component and the sample is analyzed. The analyzer then switches back to the blend header, is flushed with blend, and resumes ordinary operation.
Physical multiplexing can be used with any analyzer, but is most practical for fast analyzers. With an NIR analyzer, the analysis time is less than I min. If 15 sec is allowed to flush the analyzer after each switch, the NIR analysis of a component would take 11/2 min away from the blend analysis in a shared system.
Six components can be sampled each hour and still leave time for 50 blend analyses. Slower systems, such as distillation analyzers, probably should be used only for critical components and only when the blend recipe is stable.
In a dedicated analyzer system, the time to flush the sample loop is a significant consideration. The analyzer is not available for use from the time of a component switch until the sample loop has been adequately flushed. A sample loop flush time of 1-5 min probably is acceptable.
The shared analyzer approach eliminates the need for the component sample loop to have fast transport times because the analyzers can analyze the blend while the component sample loop is flushing. The difference between component loop flush times of 1 min and 15 min becomes trivial in a shared analyzer system.
The real-time measurement of component properties is certainly worth consideration. The choices of optical or physical multiplexing and dedicated or shared analyzers should be made with both cost and complexity in mind.
BIBLIOGRAPHY
Dunbar, D. N., Tallett, M. R., and Leather, J., "Linear Blending Values Produce Accurate Results for EPA Gasoline Emission Equations," Fuel Reformulation, July/August 1993, p. 35.
Bain, Michael L., Mansfield, Kent W., Maphet, James G., Szoke, Robert W., Bosler, William H., and Kennedy, J. Patrick, "Gasoline Blending with an Integrated On-Line Optimization, Scheduling and Control System," AIChE spring national meeting, Mar. 30, 1993, Houston.
Perino, Joseph O., "Blending Control Upgrade Projects," Fuel Reformulation, July/August 1993, p. 20.
Bain, Michael L., Mansfield, Kent W,, Maphet, James C., Szoke, Robert W., Bosler, William H., and Kennedy, J. Patrick , "Sustain Refinery Profitability and Achieve Proper Yields of Marketable Products," Fuel Reformulation, January/February 1994.
Maggard, Steven, "Octane Measured by Near Infrared," Fuel Reformulation, May/June 1992, p. 71.
Welch, William T., "Using NIR with Multistreaming for Optimization of Feed Forward Gasoline Blending," Paper FL-93-114, NPRA Fuels and Lubricants Meeting, Nov. 4-5, 1993, Houston.
U.S. Patent 4,963,745. Maggard, S. M. (to Ashland Oil Inc.).
1986 Annual Book of ASTM Standards, Volume 05-04, Test Methods for Rating Motor, Diesel, Aviation, Fuels, American Society for Testing and Materials, Philadelphia, 1986.
Maggard, Steven M., "Oxygenates in Gasoline by Near Infrared Spectroscopy," ASTM Committee D-2: Analysis of Oxygenates Workshop, Westin Crown Center Hotel, Kansas City, June 16, 1992.
U.S. Patent 5,145,785. Maggard, S. M., and Welch, W. T. (to Ashland Oil Inc.).
U,S. Patent 5,243,546. Maggard, S. M. (to Ashland Oil Inc.)
Maggard, Steven M. "NIR On-line Gasoline Blending Based on a Chemometric Model," Belgian Institute for Automatic Control (BIRA), New Analytical Techniques, Mar. 25, 193, p. I.
Maggard, Steven M., and Don N. Campbell, "The Advantage of Blending Reformulated Fuels Using Near Infrared Octane and Composition Analysis with Closed Loop Feedback," Forty-Eighth Annual Symposium On Instrumentation for the Process Industries, Rayford C. Anthony and Leo Durbin, eds., p. 61.
Maggard, Steven M., "A Near Infrared Regression Model for Octane Measurements in Gasoline Which Contain MTBE," 199th American Chemical Society Meeting, Boston, Division of Fuel Chemistry, Vol. 35, No. 1, April 1990, p. 266.
Copyright 1994 Oil & Gas Journal. All Rights Reserved.
