NEAR-INFRARED OFFERS BENEFITS AND CHALLENGES IN GASOLINE ANALYSIS

Rocco DiFoggio, Maya Sadhukhan, Martha L. Ranc
Core Laboratories
Houston

Near-infrared (NIR) analysis has enormous potential to save refiners money by helping them monitor, in real time, many parameters in such process units as reformers, alkylation units, and gasoline blenders.

The savings arise from realizing optimal process configurations and conditions, using fewer instruments, and cost-effectively measuring parameters specified by regulatory agencies.

Although widespread use of NIR has been delayed by some technical challenges, these challenges are being actively addressed. Some - advances in the technology will be described here, along with recent NIR experience in a refinery and assessment of the current state of the art.

FASTER METHOD

NIR is a fast spectroscopic technique that offers low cost per analysis. With NIR, all analyses are completed in the time it takes to obtain a spectrum (about 1 min), regardless of how many properties are being analyzed.

A single NIR analyzer can substitute for many traditional analyzers and provide faster results. NIR can be used to monitor physical properties, such as gasoline octane number or diesel cetane number, or chemical properties such as the percentage of aromatics or saturates. 6

Calibration of an NIR analyzer requires a training set of samples for which the properties of interest have been measured by conventional methods. For many applications, a new calibration must be done for each refinery and, sometimes, for each instrument.

NIR CALIBRATION

Traditional analytical techniques usually depend on a single variable. NIR, on the other hand, is an example of a chemometric technique. In chemometric techniques, the property to be determined depends on many variables, often in complex and unknown ways.

The process of NIR calibration begins with obtaining the NIR spectra of a training set of samples for which the properties of interest have been measured by traditional means. In the case of a set of gasoline samples, octane numbers would be measured by octane engines.

Models (equations) that relate the measured properties of the samples to their spectra are generated using a computer and regression-analysis software. These models allow prediction of the properties of unknown samples directly from their spectra.

MODEL GENERALITY

The model predictions are accurate, as long as the unknown samples are sufficiently similar to the training-set samples. Inaccurately predicted samples can be added to the training set to develop a revised model that has greater generality.

Model generality is less of a concern when predicting chemical composition than when predicting physical properties. This is because chemical composition is observed directly as peaks in a spectrum, whereas physical properties are inferred from a complicated correlation to their chemical composition.

Model generality is also less of a concern when modeling an individual process unit, such as a reformer, because the composition of the product does not vary as widely as do finished gasolines.

Even when unknown samples are too different from the training set to permit accurate (true) predictions, NIR predictions can still be quite precise (repeatable). For those trying to control and optimize a process unit, the immediacy and precision of NIR predictions can be more important than their accuracy.

This is particularly true when trying to determine changes in process stream properties following a change in operating parameters.

CALIBRATION TRANSFER

NIR equations often "split hairs" to distinguish between similar-looking spectra. This high sensitivity to subtle differences between spectra means that NIR regression equations can also be sensitive to subtle differences in instrument response.

Any instrument instabilities, such as slight drifts or jumps in an NIR instrument's response, can result in unacceptable deviations in the NIR predictions.

Similarly, if there is even a small difference in response between two instruments, the NIR equation developed on the first instrument will probably not work on the second without modification (i.e., calibration transfer), even if both instruments are the same make and model.

Some instrument vendors feel they have resolved the calibration transfer issue in their software or hardware. Others continue to research the subject. just as with model generality, the simpler applications are less affected by the issues of instrument stability and calibration transfer.

INSTRUMENT INSTABILITY

Although today's instruments are considerably more stable than their predecessors, NIR is now being applied to more demanding applications that have higher stability requirements.

The two most-common types of NIR-instrument instabilities are wavelength shifts (spectral X-axis) and absorbance nonlinearities (spectral Y-axis), which are called transmittance shifts.

Core Laboratories has developed a method for automatically compensating for these and other types of instrument instabilities in the software.7

For example, with use of a training set of 121 U.S. gasolines, an equation was derived for road octane number (also known as [R+M]/2, pump octane number, or antiknock index) that was simultaneously self-compensating for both wavelength and transmittance shifts (including those much larger than are likely to be encountered).

To test the effectiveness of the self-compensating equation, five gasolines representing five different levels of road octane were studied (Table 1).

Tables 1 and 2 and Figs. 1 and 2 compare the performance of the self-compensating equation to that of an ordinary (noncompensating) equation for two sets of simulated spectra that were generated from the spectra of these five gasolines.

One set simulates wavelength shifts in 0.2 nm increments from -2.0 nm to 2.0 nm. The other set simulates transmittance shifts in 0.1% transmission increments from -1.0% to 1.0%.

The octane number predictions of the self-compensating equation (Fig. la) are insensitive to wavelength shift and remain close to the perfect-prediction line. The predictions of the ordinary equation (Fig. lb), however, deviate considerably with wavelength shift for each of the five gasolines.

Similarly, the predictions of the self-compensating equation (Fig. 2a) are insensitive to transmittance shift, but the predictions of the ordinary equation (Fig. 2b) are adversely affected.

NIR VS. OCTANE ENGINES

Core Laboratories installed an NIR analyzer on a finished gasoline stream in a refinery, alongside a pair of on-line octane-rating engines. Both the spectra and the octane engine ratings were automatically recorded by a computer in real time.

The data base produced was used to determine how well all grades of gasoline, observed over a period of several months, could be modeled using a single NIR equation for each type of octane number (road, research, and motor).

CALIBRATION

For the calibration set, 13,700 engine ratings (and corresponding spectra) were obtained over a 100-day period. Core developed selfcompensating prediction equations for (R+M)/2, research octane number (RON), and motor octane number (MON). A separate prediction equation was developed for (R+M)/2 because this was found to be more accurate than averaging the RON and MON predictions.

For this 13,700-sample calibration set, the standard deviations of the residuals (prediction errors) were 0.27 for (R + M)/2, 0.41 for RON, and 0.30 for MON, with good R2 values and F-statistics, as listed in Table 3.

Bear in mind that each NIR prediction is being compared to a single engine rating and, therefore, the data can never be fitted better than the engine's uncertainty for a single measurement.

The observed NIR prediction "errors" are consistent with the inherent uncertainty of a single-engine rating (0.30 to 0.40 octane numbers). A histogram of the residuals for this calibration set is shown in Fig. 3a.

VALIDATION

Next, the calibration equations were applied to a validation set (Table 3). This set consisted of 2,071 samples obtained during a 10-day period immediately following the collection of the calibration-set data.

For this validation set, the standard deviations of the residuals were about the same (0. 29 for [R + M]/2, 0.26 for RON, and 0.38 for MON). The following slight biases, however, did develop (Table 3, Fig. 3):

0.15 for (R+M)/2
0.12 for RON
0.07 for MON.

The biases for specific blends varied from these average biases. A histogram of the residuals appears in Fig. 3b.

These results imply that the models, which fit the calibration set very well, were not general enough to predict the validation set without some bias. Perhaps the blend compositions used in the validation set were combinations unseen in the calibration set or perhaps the crude oil feedstocks differed. Other factors may also play a role.

PRECISION TESTS

As a test of precision, iso-octane was run through the NIR analyzer probe automatically, once a day, at the refinery (Fig. 4). Over 100 days, the predictions (using refinery-calibration equations) were repeatable to within 0. 134 octane numbers for (R + M)/2, 0.135 for RON, and 0.134 for MON.

Ignoring one 3-week period (Days 24-44) when the iso-octane may have been contaminated with gasoline from the blends, the prediction repeatability was even better (0.096 for [R+M]/2, 0.098 for RON, and 0. 107 for MON).

It should be noted that Core did not expect its models to be general enough to accurately, predict iso-octane because pure iso-octane had not been included in the calibration set. Thus, the average predictions for iso-octane were 105.134 for RON, 100.669 for (R+M)/2, and 98.020 for MON, instead of the ASTM-defined value of 100.00 (Fig. 4).

BLIND TESTS

Unfortunately, most of the published reports of measuring octane number by NIR are based on small sample sets of about 50 samples. 1 3 6 Only a few have used several hundred samples.

Core has some additional insights into the generality of octane-number models because it has provided blind validation samples to other laboratories determining octane number spectroscopically.

These laboratories were given National Exchange Group (NEG) gasoline samples (stored at 2 C.). The octane numbers of the samples were known quite accurately because they had been rated on approximately 30 different engines, and the results averaged (excluding outliers).

Table 4 shows that these outside laboratories were unable spectroscopically to predict the octane numbers of blind samples within the accuracy they anticipated (0.2 to 0.3 octane numbers). All the laboratories determined (R+M)/2 by averaging their RON and MON predictions instead of predicting (R+M)/2 directly.

Some laboratories used near-infrared analysis, others used mid-infrared. Some used discriminate analysis to decide which samples could be predicted with their models and which could not.

Table 4 only includes samples that a laboratory believed it could predict with its model. An interesting point is that there was little difference in prediction accuracy for samples that the laboratories felt they could predict compared to those they felt they could not.

Core believes other laboratories thought their models were more general than they were because they simply had not seen enough samples of different gasolines. Core has looked at over 1,000 finished gasolines from all regions of the U.S., and about 20,000 samples from one refinery. Core observed that octane response can be quite complex. Spectra that are nearly identical can have very different octane numbers, and spectra that are very different can have essentially the same octane number.

Results of these blind tests suggest that a universal octane-number equation for all gasolines that is applicable at all refineries has yet to be developed. However, many people in this field have long felt that, for best results, instruments at each refinery should be individually calibrated.

Core's refinery test for finished gasolines suggests that even this approach may require large training sets and periodic model updates to make NIR a true substitute for octane engines. Meanwhile, NIR can be used to monitor octane number of finished gasoline on-line by obtaining several engine ratings for each blend to correct for any NIR bias resulting from a lack of model generality.

FUTURE

Near-infrared and other spectroscopic analyses hold considerable promise for helping refiners stay on specification and optimize their processes. Initially, the technology will find more applications on individual process units, such as reformers, where precision and immediacy are often more important than high accuracy.

Core has helped move NIR technology forward by developing prediction equations that are self-compensating for various types of instrument instabilities.

As this technology matures and other issues, such as calibration transfer and model generality, are more satisfactorily resolved, it will have greater impact on more complex applications such as blended gasoline.

REFERENCES

Hirschfeld, Tomas, "Near-Infrared Reflectance Spectrometry: Tip of the Iceberg," Analytical Chemistry, Vol. 56, No. 8, July 1984, pp. 933A-34A.
Honigs, D. E., Hirschfeld, T. B., and Hieftje, G. M., "Near Infrared Determination of Several Physical Properties of Hydrocarbons," Analytical Chemistry, Vol. 57, No. 2, February 1985, pp. 443-45.
Kelly, Jeffrey J., Barlow, Clyde H., Jinguji, Thomas M., and Callis, James B., "Prediction of Octane Numbers from Near Infrared Spectral Features in the Range 660-1215 nm," Analytical Chemistry, Vol. 61, No. 4, Feb. 15, 1989, pp. 313-20.
Kelly, Jeffrey J., and Callis, James B., "Non-Destructive Analytical Procedure for Simultaneous Estimation of the Major Classes of Hydrocarbon Constituents of Finished Gasolines," Analytical Chemistry, Vol. 62, No. 14, July 15, 1990, pp. 1444-51.
Asker, N., and Kokot, S., "The Application of NIR Spectroscopy for the Prediction of Properties of Australian Refined Reformate," Applied Spectroscopy, Vol. 45, No. 7, 1991, pp. 1153-57.
Swarin, Stephen J., and Drumm, Charlene A., "Predicting Gasoline Properties Using Near-IR Spectroscopy," Spectroscopy, Vol. 7, No. 7, September 1992, pp. 42-49.
DiFoggio, Rocco, and Sadhukhan, Maya, U.S. Patent pending.