INFERENTIAL CONTROL-CONCLUSION CRUDE UNIT CONTROLS REDUCE QUALITY GIVEAWAY, INCREASE PROFITS

Yip Poh San
Singapore Refining Co. Pte. Ltd.
Pulau Merlimau, Singapore

Keith C. Landells, David C. Mackay
BP Oil International
London

Singapore Refining Co. Pte. Ltd. (SRC) is using quality-control schemes based on inferential predictions to increase throughput and profits from its No. 2 crude distillation unit.

[For an explanation of inferential control, see Part 1 of this series, OGJ, Nov. 28, p. 76.]

Since implementation of the controls, the unit-originally designed to process 130,000 b/d-now can process as much as 146,000 b/d. This increased throughput, along with reduced quality giveaway is saving the refiner as estimated $8 million/year.

The first article in this two-part series described the control system and its implementation. This concluding article outlines the project's achievements.

PROJECT RESULTS

The tuning process was to be carried out in two stages.

First, all crude-independent qualities were to be rigorously tested against laboratory and on-line analyzer results (where possible). Once the "Infer" model had passed this stage, crude-dependent qualities (such as kerosine freeze point) were to be introduced.

As soon as results from the first stage were available, however, SRC and BP decided to implement the kerosine freeze-point prediction immediately, albeit in a much simplified form.

The new form effectively gave a nonlinear freeze-point prediction that needs to be biased by laboratory results for each crude blend. The tuning results discussed below, therefore, do not include kerosine freeze point.

REPEATABILITY

Infer cut-point and product quality predictions, for both the preflash and main fractionator columns, appeared stable under all conditions except serious process upsets. This analysis was based on 30 sets of plant data, each containing 2 hr of 6-min averaged data over 30 consecutive days.

For the most stable plant data, Infer predictions exhibited standard deviations on the order of 0.1 C. (Under normal conditions, standard deviations of 0.2-0.8 C. were typical for cut points and other qualities.)

DIRECTIONAL RESPONSE

Because of the project time scale, climatic variations, other work, and increasing confidence in Infer, SRC and BP carried out only a limited number of unit tests in this category. The results are shown in Table 1.

The observed steady-state gains from Infer were reasonable in both magnitude and sign, and showed some variation with crude type. They did not, however, always coincide with gains determined from laboratory results.

In this context, it must be noted that the changes predicted by Infer were sometimes of the same magnitude as the recognized repeatability of the laboratory test. On-line analyzer results, when available, generally agreed with Infer, but they exhibited long dead times of an hour or more.

STABILITY

Table 2 shows the results of the comparisons of Infer with laboratory data. A normalized distribution was assumed and standard deviations were calculated for each quality, based on the observed ';lab minus Infer prediction" difference.

The final column in Table 2 shows the percentage of values further than two target standard deviations from the mean. (While more samples should have been analyzed to allow a valid statistical analysis, this had be to compromised, given the finite time and resources available.)

Diesel 90% point and flash point showed significant differences between laboratory and Infer results. All of the other values were acceptable.

ACCURACY

The observed repeatability range of the sampling and laboratory analysis process was wider than expected. This gave rise to a larger spread of "lab - Infer" values, for any specific quality, than might otherwise have been seen.

By plotting the results of all of the comparisons, it was clear there existed lab-vs.-Infer biases that were dependent on the magnitude of the quality value. This should not be so, and is probably caused by a combination of insufficient model tuning and systematic errors in the sampling and analysis methods, although the former is more likely.

This skewing was "extracted" from the model predictions. The results in Table 2 were modified (via simple bias corrections) to give the results shown in Table 3.

The key prediction (diesel 90% point) still exhibited a great deal of variation. After closer inspection of its dynamic behavior, however, SRC and BP were convinced that it would be of use in the control scheme.

Fig. 1 shows the 90% prediction and on-line sulfur analysis of the diesel stream during an 18-hr period using a constant crude blend. It is clear that, for a given crude blend, the Infer prediction is valid, even when viewed over a short time.

CONTROL SCHEME USAGE

As stated in Part 1, the project philosophy was to start at the lowest levels of control and build methodically up to the high-earnings schemes. For this reason, simple monitoring methods were developed to allow engineers to see quickly the percentage of time that each advanced control scheme (ACS) was in closed-loop.

While this does not ensure that the loop has the correct set-point, it does indicate how stable the process should be.

Table 4 shows 1993 monthly usage factors for each scheme. (The data for March is missing because of an archiving problem.) Several problems were encountered:

In June, the refinery commissioned the heavy naphtha draw from the No. I pumparound. Infer needed to be tuned to this mode and, hence, was off-line much of the time (the effects of this can be seen later in the SRC middle distillate yield performance shown in Fig. 6).
In July, and throughout the remainder of the year, oil firing of the heaters was prevalent and fuel gas compensation was not required.
In August, crude supply problems led to severely reduced throughputs and unusual plant behavior.
In November, the crude unit was shut down because of a leak, and hydraulic problems were experienced at the top of the main fractionator. The data from this period were removed from Table 4.

UNIT STABILITY

Hour-to-hour unit stability is illustrated in Figs. 2, 3, and 4. These figures show each of the Infer quality-control loops and on-line analyzer trends, where appropriate. '

The SRC unit achieves this stable performance through tropical storms and crude changes as long as minimal, sensible, manual intervention steps are taken when necessary.

Because SRC products generally are certified on the basis of laboratory results, Table 5 has been included. This shows target run-down qualities, laboratory averages, and standard deviations for those qualities, for three periods:

Before ACS commissioning
Immediately after ACS commissioning (when there was a high level of control engineer input)
During normal operation (with very little control engineer involvement).

It should be noted that the kerosine freeze-point giveaway often is very small, or even negative (i.e., diesel is worth more). The freeze point, therefore, is not always tightly controlled.

QUALITY GIVEAWAY

To reduce giveaway and improve yields, Infer predictions must be consistently correct during a crude run.

SRC observed that the TBP slope parameter sometimes drifted higher during a crude run, causing Infer to adjust yields accordingly. This allowed continuous adjustment of yields in the face of known, but previously undefinable, changes in single-tank crude composition, and possibly in crude tank ratios, as levels decreased. Fig. 5 shows a typical trend for this parameter.

It was explained in Part 1 of this series how giveaway in kerosine flash could be reduced. Using typical LP-based gains, in terms of "quality / yield," it was possible to estimate the quality giveaway in $/bbl.

Because the production plan predicted product yields, it was possible to calculate the difference between the planned and actual yields and develop another measure of unit performance. These values are given in Table 6 for the pre and post-ACS cases, indicating a unit margin improvement of 9/bbl.

While these data gave the project team confidence that, however the data was analyzed, definite improvements had been made on the unit, the final confirmation needed to come from refinery product sales.

Each month, the difference between "planned" and "produced" middle distillate (kerosine + diesel) is calculated, including all refinery units. Because Infer was successfully implemented on the other crude unit CDU1, in July 1993, ii should be possible to infer the "global" effects of this model using this calculation. The results are shown in Fig. 6.

FEED RATE MAXIMIZATION

Fig. 7 shows the difference between the actual feed rate to CDU2 during 1993 and the feed rate based on the pre-ACS capacity factor. Note that operation exactly according to plan has been assumed for the first 3 months because feed rate tests were ongoing then. Those tests led to the conclusion that maximum feed rate should be constrained to 138,000 b/d.

Informal work on feed rate maximization commenced in June 1993, was interrupted in August because of supply problems, then continued through September and October. In November, unrelated hydraulic problems and a unit shutdown affected performance.

Maximum feed rates of 146,000 b/d have been achieved with no yield losses. The constraint controller commonly runs on furnace oxygen or pass control valve opening limits, or both, for the heavier crudes.

The maximum vapor rate (calculated from distillate yields and the Infer over/lash estimate)-or, for lighter crudes, the main fractionator pressure-drop constraint-is hit simultaneously.

ECONOMICS

A variety of analyses of the daily average data over several months showed no evidence of total distillate yield loss because of the increased throughput. For the heavier crudes, distillate yield was found to be consistently lower than planned, but this "loss" did not increase significantly as throughput increased.

Correlating data over a 2-month period produced the following results:

Distillate yield, % 97.4 - (8.31)(crude slope) (0.06)(feed rate, in 1,000 b/d).
Correlation coefficient, R = 0.95.
0.06% (on crude) loss in distillate yield for each 1,000 b/d increase in throughput was used as a basis for the sensitivity, tests.
Investigation of the effects of feed rate on overheads separation was carried out in two ways:
First, the monthly separation gap (overheads liquid 93% point vs. kerosine 5% point) was compared with average throughputs throughout 1993. There was no change in the separation gap, with average monthly throughput varying from 117,000 to 142,000 b/d.
Second, laboratory results for the overheads liquid 95% point, over a 6-month period during which the kerosine flash point target was essentially constant, were compared with feed rates for the sample period. No correlation was found between the data.

An off-line sensitivity analysis was carried out to ensure that throughput maximization was always analogous to maximizing daily operating profit. The conclusion was that, only if the straight-run margin decreased to less than 33% of its normal levels, would there be any possibility of losing money by pushing the unit to its physical constraints.

ADVANCED CONTROLS

Implementation of advanced controls on CDU2 has led to much more stable unit operation. Even the highest-level schemes achieve service factors of 9097%. Performance of the quality-control loops during tropical rainstorms and crude changes exceeds expectations.

The technology used (inference software combined with distributed-control algorithms, supported by multiple-input, single-output dynamic modeling software) is simple to maintain. And given the economics of CDUs in the region and SRC resources, it is hard to see how the application of a more-complex technology could achieve significant extra benefits at this site.

The performance of Infer has been consistently good for more than a year. Its availability is greater than 99% and it copes well with the frequent crude changes, rainstorms, and less-frequent mode changes.

There is no reason why its success could not be repeated elsewhere, provided the fundamentals of plant measurements and laboratory sampling and analysis are dealt with rigorously and correctly. This aspect may not be glamorous or high-profile, but it is essential to success.

While maintenance of the controls has not required more than a few hours per month since June 1993, a great deal of effort has been expended in the areas of laboratory data-handling and analyzer use. Reducing give-aways via proper management of controller set-point changes is now the primary focus of the unit control engineer.

The relatively healthy straight-run margin (compared to Europe or North America) enabled simplification of the controls. If the margin decreases significantly in the future, secondary factors-such as fuel use and yield loss at higher throughputs-will need to be considered in more detail.

FUTURE WORK

The project team concentrated on implementing controls in which the unit operator has confidence. This was a vital and successful step toward achieving high usage factors.

In the third quarter of 1993, however, emphasis shifted toward effective use of the advanced controls. This was not unexpected, but illustrates the need to continuously review and question plant performance at all levels, and in all refinery departments.

Because of other commitments, the control engineer's attention to CDU2 was greatly reduced in December 1993. This may be one of the factors in the performance deterioration for that month (Figs. 6 and 7).

Ensuring that the benefits demonstrated during this work can be sustained is the next challenge for SRC management.

The analyzer group at SRC has made dramatic improvements to the performance of the kerosine flashpoint analyzer over the past 6-9 months. While it still has a lower availability than infer and exhibits slower dynamics and occasional spikes, long-term accuracy appears to be excellent. Using the analyzer as a slow "trim" to the Infer prediction, in closed-loop control, is being investigated.

All laboratory flash-point analyses of the process rundown will cease if the analyzer is used in the control loop (except when required for calibration purposes). But, whatever the outcome, there is clearly a need to investigate statistical techniques of handling laboratory results. The current situation can lead to unnecessary knee-jerk reactions to single results, which inevitably leads to higher giveaways.

The furnace coil outlet temperature (COT) is currently under manual control. During crude changes, this can lead to flooding or dumping of the column. The possibility of adjusting the COT automatically, based on Infer data, is being considered.

The diesel 90% controller has been tuned to maximize yield under normal operating conditions. During crude changes, however, the controller can be too slow.

This can lead to a short period of either underdrawing the diesel stream (a small financial loss) or overdrawing it (which can lead to an off-specification product). To combat this, operators generally break the scheme if the new crude is significantly heavier than the previous blend.

Other methods of improving control performance are being investigated.

ACKNOWLEDGMENTS

The authors wish to thank all BP, project, and SRC staff involved in this work, particularly the instrument, operations, and laboratory technicians. The input provided by Dr. R.Y. Bartman of Procontrol Inc. is gratefully acknowledged.