INFERENTIAL CONTROL-PART 1 CRUDE UNIT ADVANCED CONTROLS PASS ACCURACY AND REPEATABILITY TESTS

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

Keith C. Landells
BP Oil International
London

David C. Mackay
BP Oil International
London

Quality control schemes using inferential predictions have reduced product giveaways and enabled throughput increases on a large crude distillation unit in Singapore.

An inferential model is one that provides a quality for which an analyzer is not available. This type of model uses readily available physical measurements-such as temperatures, pressures, and flow rates-to infer a quality such as kerosine flash point.

The No. 2 crude distillation unit (CDU-2) at Singapore Refining Co. Pte. Ltd.'s Pulau Merlimau refinery has a nominal 130,000 b/d capacity. It produces naphtha, kerosine, diesel, and residue products from a wide range of crude blends.

Over the past 12 months, extensive advanced control applications have been implemented on the unit. Increased profits of about $8 million (U.S.)/year have been achieved and verified independently.

This first of two articles will describe the control system and its implementation. The second will outline the project's achievements, including reduced quality giveaway and increased profits.

BACKGROUND

Singapore Refining Co. (SRC) is a joint venture of British Petroleum Co. Ltd., Caltex, and Singapore Petroleum Co. Pte. Ltd. It is located on Pulau Merlimau island off Singapore. BP provides all technical and management support to the refinery.

The nominal crude capacity of the refinery is 220,000 b/d. The crude processed varies widely in terms of distillate yields (52-70 vol % on feed). Each crude cocktail can last 1-6 days, depending on scheduling requirements.

As part of a reinstrumentation project, control of the refinery process units has been moved to a centralized control building and is operated via a distributed control system (DCS). A small team of BP and SRC staff has been responsible for the implementation of about 150 advanced control schemes throughout the refinery.

The CDU2 was extensively modified in 1992, with the addition of a preflash tower to increase throughput. The main economic drives for the unit are to maximize kerosine and diesel yields while maximizing throughput.

Normally, there is little price difference between these two products. Atmospheric residue and overheads - effectively a chemical grade naphtha - have a much lower value.

To meet the economic requirements, a variety of advanced control schemes were implemented, including several quality control schemes that use a combination of inferential quality models and dynamic reconciliation (DR) technology from Procontrol Inc.

PROCESS DESCRIPTION

The process flow for CDU2 is shown in Fig. 1.

Crude oil from various tanks is blended and fed, via the preheat exchanger train, to the preflash tower. This tower was installed to unload the top section of the main fractionator.

The overheads liquid (a naphtha stream) is injected directly back into the overheads liquid from the main fractionator, while the gas is compressed and cooled before reinjection. A small flow of unstabilized naphtha/kerosine material from another unit is injected into the column to recover the kerosine fractions.

The base of the tower has been designed to hold a "buffer" inventory of material to ensure that the furnaces have sufficient flow rate at all times.

Preflashed crude enters the main fractionator via two furnaces in parallel, the load being split about 30:70 between them. The tower generally has four product streams, namely overheads liquid, kerosine, diesel, and residue. A fifth product, heavy naphtha, can be draws between the overheads and kerosine streams (from the No. 1 pumparound) but is not normally taken.

Both the kerosine and diesel streams are steam stripped, and residue stripping steam is injected into the lower section of the tower. Overheads cooling is generally achieved solely via No. 1 pumparound (P/A), but is backed up by a cold external reflux stream from the overheads, and the accumulator, when necessary.

Two further pumparounds are located in the lower sections of the tower.

All overheads material from the main fractionator is combined with preflash tower overheads and fed to the stabilizer, which produces chemical naphtha and LPG. Excess gas is sent to the fuel system.

PROJECT IMPLEMENTATION

Throughout implementation of the controls, the project staff attempted to keep the following guidelines in mind:

Start at the highest level, with regard to unit economics, when designing schemes (focusing on valuable products and key quality targets).
Start at the lowest level when implementing schemes (i.e., the control valve or orifice plate).
Talk to the unit operators as much as possible standardize and simplify the operating interface.
Use the simplest technology that is capable of achieving the economic benefits. (Do not implement anything that is of purely technical interest.)
Apart from the inferential model, all controls should reside at the plant DCS level.
avoid user-written control programs when possible, particularly in the feedback paths of control strategies. The technology used must recognize resource limitations and be maintainable by existing site staff.

Once conceptual design of the advanced control schemes (ACS) had been completed, attention was switched to the regulatory loops, and to other more-advanced loops that were needed to support top-level quality control schemes.

Furnace and stripping-steam schemes were implemented in late 1992, the yield controls and pumparound management in February and March 1993, quality controls in April 1993, and rate maximization in August 1993.

Apart from new loop design and implementation, existing PID control loops were retuned, where necessary, using the Discover software package from Procontrol. After a general review of regulatory loop performance, four main areas, all relevant to the main fractionator, received further attention.

Four new schemes were implemented on each furnace, with the principal objective of maintaining coil outlet temperature (COT) control in the face of load and supply-side disturbances, but also to improve operating efficiency. Additionally, a throughput balancing scheme for the two furnaces was commissioned. The schemes covered:

Heater pass balancing
Heat load compensation
Fuel gas compensation
Excess oxygen control

Because of its importance in supporting all other schemes, extra effort was expended on the overheads temperature control. Additionally, a dew point calculation was installed as an override controller, introducing external reflux to boost hydrocarbon vapor pressure when necessary.

Ratio controls were implemented on the kerosine, diesel, and residue streams. The kerosine scheme, however, was later modified to improve plant performance.

Initial attempts to implement heat load control on the pumparounds were abandoned after repeated plant tests produced inconsistent results. Instead, a pumparound management scheme was successfully commissioned.

This scheme allows the operator to set target flow-rates for Nos. 1 and 2 pumparounds, with the balance of heat removal being absorbed by varying No. 3 pumparound (the base pumparound). By setting suitable targets, the operator is able to shift heat removal up or down the column as required. At SRC, moving heat up the tower is the normal mode of operation.

Classical downward decoupling techniques were used to implement two controllers, namely "overheads liquid + kerosine" volume yield, as percent on feed, and "overheads liquid + kerosine + diesel" volume yield, as percent on feed. Minor modifications were made during commissioning to cope with loss of diesel-stripper level and the possibility of taking a product draw from No. 1 pumparound.

Extensive plant testing was carried out, with support from the BP inferential model and Discover, to ensure that the schemes were robust under normal conditions for all crude types processed.

INFER MODEL

After a great deal of debate on the relative merits of on line analyzers and inferential calculations, the refinery opted for the latter on both crude units. The inferential model used to support these quality control schemes is the BP proprietary software, Infer.

In this application, Infer was required to predict several ASTM distillation points, as well as flash and freeze points, for the kerosine product.

In the first stages of Infer, measured values of temperature, pressure, and flow rate from the distillation tower are used to calculate a series of cut-point temperature and yield pairs on the crude true boiling point (TBP) curve.

The TBP curve is a standard, arbitrary definition of crude type, in terms of boiling point curve (normally expressed as wt % distilled at a given temperature, although vol % is more common in the U.S.).

Once a curve has been fitted through these discrete, inferred points, a fractionation model is invoked. This model generates TBP curves for the overheads, side-stream, and residue products.

From these product TBP curves, the distillation qualities required by the , crude unit advanced control schemes are derived using standard correlations. Additional properties can be derived using data from crude assay packs. (These packs are generated regularly at the BP Research Centre.)

There are two properties predicted at SRC that use blended pack data. The first is flash point. This property is essentially crude-independent. To calculate a product flash point, however, it is necessary to have a table of data relating flash point index to narrow-fraction boiling point. This table is held in every crude assay pack.

The second property derived from assay packs is kerosine freeze point. Rather than blend up a series of crude packs for every crude cocktail run at SRC, a single typical cocktail was blended once, and this is used continuously as a source for freeze and flash data in this application.

As far as freeze point predictions are concerned, this approach probably only works at SRC because the company is interested only in kerosine. For heavier side streams, full pack blending for each crude run may be required. As part of the calculation sequence, the model estimates the average "slope" of the crude TBP curve, and the over-flash. The former is a good indicator of crude change. Fig. 2 shows this slope parameter through a crude change.

The debottlenecked unit can be run in a variety of distinct operating modes, and the inferential model has to be able to handle each one. So far, the model has been commissioned to control the column in two different modes.

The model is implemented as a Fortran application program, running every 6 min on the DEC VAX computer associated with the DCS. The model is scheduled by a BP real-time applications environment that also provides logic constructs, which ensure that any failures in prediction are handled properly and consistently at the control scheme level, whether the errors are caused by bad input data, program failure (in conjunction with logic at the TDC level), or failure of the supervisory computer.

MODEL TUNING

In previous applications of Infer, the project team felt that insufficient attention was paid to specifying the conditions under which suitable data for tuning the models were to be collected, and to defining a reasonable set of acceptance tests.

To tune the inferential model to the plant, it was necessary to gather several sets of complete plant data and corresponding laboratory results. Before any set of data was used in model tuning, however, it had to satisfy previously specified conditions:

The model would only be tested on draw modes for which sufficient tuning data were available.
Only plant operations within the specified range of operating variables and stability criteria were to be used during tuning of the model.
Unsatisfactory mass balance errors or unstable unit operation led to rejection of the data gathered.
The data were accepted only if enough laboratory results were available to completely specify the products in terms of their distillation curves.

It is impossible to overstate the importance of obtaining good quality tuning data at this stage in the project. A great deal of effort was put into ensuring that the mass balance error over the unit was acceptably small and, more significantly for control purposes, that it was not crude dependent.

If all the unit prerequisite conditions were satisfied, the following tests were per- formed:

Repeatability-This was effectively, an on-line sensitivity analysis. The changes in Infer predictions, on a run-to-run basis, were compared quantitatively with plant input data variations. Tests were conducted over a period of "steady" unit operation to ensure that normal variations did not lead to unacceptable variations in Infer predictions.
Directional response - Ideally Infer responses were to be tested during changes in yield, crude feed, and fractionation.
Stability - This test was carried out at the same time as the previous one. All Infer predictions were compared with plant and laboratory data over a long time period ( 4 weeks). Any unexplained changes in Infer predictions were investigated further.
Accuracy - The absolute accuracy of Infer predictions was tested against plant and laboratory data. Assessment of the results took into account ASTM test accuracies and repeatabilities derived from studies undertaken regularly by BP.

Table 1 summarizes the maximum allowable standard deviation (SD) of the difference between an Infer-predicted value and the laboratory analysis for any given quality over all valid sets of results.

CLOSED-LOOP CONTROL

The inferential model produced three predictions that were used by control schemes:

Kerosine flash point
Kerosine freeze point
Diesel ASTM D86 90% point.

The suitability of Infer predictions for use in closed-loop control was viewed as critical. Even if all the acceptance tests were passed, however, the problem of how to include a discrete prediction (based inherently on steady-state assumptions) in a control loop remained. SRC and BP decided to employ dynamic reconciliation. This scheme allows the Infer predictions to effectively update the much-simplified models used to control the process.

Safety and robustness considerations led to the development of a number of DCS level checks to ensure that both system failures and "bad" runs from Infer were handled correctly, without any need for control engineers to be called in. Tuning of the quality control loops was carried out using a combination of plant tests, analyzer data, lab data, and operating experience.

During plant tests, it became clear that, rather than control stripping steam in ratio to kerosine draw, it was more sensible to maximize the stripping steam flow, up to a constraint. Repeated plant tests led to the development of a target-type steam controller that automatically maximized stripping steam ratio for a given crude, while staying within equipment constraints.

FEED PATE MAXIMIZATION

During the early stages of the CDU-2 work it was recognized that, given the operating economics of the region, it would be desirable to maximize crude throughput at all times.

Because of the financial, time, and resource constraints on the project, SRC decided that any closed-loop control of feed rate would have to be based on TDC-3000 algorithms and not on proprietary multivariable control software.

After a few months of stable, closed-loop quality control, efforts were made to manually increase feed rates and observe process constraints for a variety of crude feeds. In this way, a multiple constraint "pusher" was developed.

ECONOMIC MONITORING

The objective of the controls-including feed rate maximization-was to improve the gross contribution to the refinery in terms of $/day, from CDU-2.

Because of concerns over yield losses at higher throughputs, a series of plant tests was conducted to ascertain the relationship between throughput, crude type, and distillate yield (at a fixed diesel 90% point).

This work, however, only considered loss in separation at the bottom of the tower. Loss of separation at the top (leading to high overheads and low kerosine yields for a given kerosine flash point) was also important.

Regular laboratory tests were carried out to determine overheads 95% point, final boiling point, and kerosine 5% point. Over a long period of time, these data could be used to examine the effects of feed rate on tower top separation.

Once these two effects had been quantified, a series of off-line sensitivity runs was carried out to estimate the feed rate beyond which unit contribution to refinery profit would begin to fall, for varying feed and product price sets and feed types.

The results from these tests were used as the basis for tuning the model. Part 2 will describe that process and the benefits achieved by project implementation.