MULTIVARIABLE CONTROL SYSTEM INSTALLED AT ARCO WEST TEXAS GAS PLANT

Kelley Chou
ARCO Oil & Gas Co.,
Midland, Tex.

Roger M. Clay
ARCO Exploration & Production Technology Inc.
Plano, Tex.

Jorge P. Gamez
Gas Research Institute
Chicago

P. N. Berkowitz, M. N. Papadopoulos
Continental Controls Inc.
Houston

A PC-based, multivariable process control (MVC) system was installed last year at an ARCO Oil & Gas Co. gas plant in West Texas. This gas-processing application was developed under sponsorship of the Gas Research Institute. The system was installed, tuned, and on-line within 2 weeks and fully verified in closed loop service by operations in 8 weeks. Four more gas processing installations are currently under way.

The general and main objective of the MVC control system is to achieve continuous optimum operation of a process unit through on-line prediction and control of setpoints for the key process variables in the unit.

Specifically, the objective is to achieve this operation, especially under constantly changing conditions, with reliable solutions requiring minimal operator intervention, customization, or up-date effort upon each plant change. MVC was developed by Continental Controls Inc. (CCI), Houston.

ON-LINE CONTROL

MVC is a multivariable advanced control system designed for on-line optimum control of continuous processes. Alumina refining, fractionation, cryogenic expansion, and gas processing and treating are example applications.

Optimum operation is achieved through a reduction in plant operating variability, tighter specification targets, and improved variable cost control.

The MVC system is PC-based, has software drivers connecting it to the plant's regulatory control system whether that system is based on pneumatics, electronics, programmable logic controllers (PLCs), or a distributed control system (DCS). It uses modules of rigorous process simulations tuned to the actual plant performance.

In operation, the MVC system continuously polls the many field variables then uses the inputs to feed-forward predict the optimum setpoints for control of manipulated process variables (MVs).

Before a feedforward-based setpoint is dispatched to the regulatory control system, a feedback trim based on actual vs. predicted process reactions is computed.

(Throughout this discussion, "feedforward" refers to the process in which the system calculates the effect of a process setpoint before it happens; "feedback" refers to the correcting data received from the unit once it has received the feedforward setpoint.)

Its predictions are subject to the plant's economic, contractual, and equipment constraints. The variables are optimized approximately every 30 sec.

With UNIX as its operating platform, the PC-based MVC system compares favorably with conventional large minicomputer-based systems in terms of variables' setpoints processing capacity.

The system is also adaptable. Process lags, critical equipment constraints, and performance and instrument drift are updated on-line through self-tuning routines. Before each calculated, economics and feedforward-based optimum setpoint is dispatched, a feedback trim based on actual-vs.-predicted process reactions is computed.

OPERATION

The system is based on rigorous process simulations of actual plant performance.

In operation, the MVC system continuously polls the many field variables, then uses the inputs to feedforward predict the optimum setpoints for control of manipulated process variables (MVs).

As stated previously, before a feedforward-based setpoint is dispatched to the regulatory control system, a feedback trim based on actual-vs.-predicted process reactions is computed.

The trim is essentially a "sanity check" of the predicted plant output vs. the observed output which may change from nonlinearities between unit loading and separation efficiency. The observed output itself is filtered through drift factors which effectively provide on-line calibration of key measuring instruments.

Such filtered, feedback-corrected optimum MV setpoint computation and dispatch to the process occur on every control cycle, approximately every 30 sec.

Drifts in process performance are updated on-line periodically through self-tuning routines. These are computed as calibration factors for the predictor equations' consistent deviations from observations.

These "equation tuning factors" are automatically applied at chosen time intervals, daily, for example. Their effect is to minimize the size of the feedback corrections which are computed on every control cycle, as stated.

Changes in process equipment (and concomitant changes in process performance) such as addition of exchangers, piping, etc., are also updated through similar self-tuning routines.

In this case, the predictor equations are "relinearized" through recomputation of the predictor equations' coefficients. The purpose of changing coefficients rather than the model is that small coefficient corrections give greater tuning alignment without invalidating the structure of the model.

Thus, the predictor equations or algorithms remain similar for given process unit operations and thus minimize the need to modify the model as long as the basic function of the process unit remains the same.

Although MVC has been designed to cover the entire range of normal plant operations (usually greater than 20% of the average throughput, compositions, state variables, etc.), regulatory control takes over beyond this range during start ups.

The MVC algorithms are linear forms of polynomials rather than nonlinear mechanistic equations. Otherwise, input data reconciliation is extensive because bad data are catastrophic in mechanistic solutions.

Essentially, it is repeatability of instrument performance, rather than instrument absolute accuracy, which is important in data filtering and reconciliation.

Filtering is employed for both live measurements and manual data entry because single-point variations can significantly affect model output predictions. Linear solutions can be forced to track the process.

Nonlinear mechanistic solutions may never converge, resulting in an indeterminate value or a clamped output to the process.

In operation, MVC relieves the operator of the need constantly to readjust setpoints and of the burden of overusing such resources as utilities to remain safely within specification.

BASED ON 386-PC

The general architectural concept of MVC is illustrated in Fig. 1.

The PC-based system connects to the process units through a standard manufacturer's DCS, remote or local data acquisition equipment, or a PLC.

The system resides wholly within a PC-386 containing a data-acquisition program, the real-time data base (RTDB) for the actual process plant, and the control modules (CM) which are standardized software packages for on-line control of the particular process units.

A CM is defined as a relatively self-contained process segment control package including the "dead-time" (time of analysis) process variables. Fig. 2 shows a screen display of a CM flow diagram for the refrigerated lean-oil absorption segment of a typical gas-processing plant.

CMs have been developed for cryogenic expansion, refrigerated absorption and fractionation, amine treating, sulfur recovery, and a number of the most commonly encountered distillation process segments in gas processing and treating.

The MVC scheme uses rigorous off-line plant simulations to characterize particular unit operations. An onsite survey provides the data for such simulations which are carried over ranges of expected variables of plant operations.

These rigorous simulations yield the coefficients for the linearized equations that form the MVC algorithms which represent unit response to the multivariable process inputs.

Equipment constraints and contractual and economic factors are incorporated into the algorithms. The on-line control programs use the MVC algorithms to set the process conditions in each section of the plant. Thus, only modest customization is required in applying the MVC modules from plant to plant.

The MVC uses on-line and historical data to calculate equipment performance such as fouling factors of key heat exchangers and fractionator tray efficiencies. Plant changes, such as additions of compressor and heat-exchanger capacity, require nominal changes of the equation coefficients.

Significant process flow-scheme changes, such as addition of recycle streams, rearranging process sequence, etc., may require more extensive re-normalization of coefficients. Generally, such "recalibration" engineering efforts are short because the nature of modules is fairly standard in the industry.

ARCO INSTALLATION

CCI installed the MVC at an ARCO West Texas gas plant. The project was sponsored by GRI which contracted with CCI to support the development and commercialization of several MVC gas processing and treating modules.

The ARCO facility processes gas produced in CO2 miscible flooding operations. The NGLs are extracted from the inlet gas together with H2S by refrigerated lean-oil absorption, and the CO2-rich residue gas is returned to the field for reinjection.

The MVC system was installed with the two modules shown in Fig. 2, the absorber and fractionator.

The MVC objective for this specific application was to provide economically optimal recovery of C2+ (with minimal loss of CO2 in the sour NGL) against varying costs of heat and power consumption, while satisfying H2S specification in the residue gas.

In operation, the system continuously polls the field measured variables, selected feedback variables, and laboratory data obtained in the plant. The polling of variables is through a CCI software driver interfacing MVC with the plant DCS.

Polling frequency varies from a few to several seconds, depending on the nature of the variable polled. An on-line gas chromatograph is also polled by MVC through the DCS and driver.

MVC uses its algorithms and control programs to compute and dispatch the optimum setpoints for control of MVs. As shown in the display in Fig. 3 for the absorber module, there are three such optimized manipulated process variables: the reflux rate, lean-oil flow, and reboiler duty.

The frequency of optimization varies from every 30 sec to a few minutes, depending on the process dynamics.

The computation of the setpoints for the manipulated variables is feedforward with feedback trim. Process lags and equipment constraints were established during the plant survey conducted initially for the configuration of the MVC data base.

An example of the control logic is shown in the screen display in Fig. 4 for the lean-oil flow manipulated process variables. As shown in the upper left of Fig. 4, a tentative set point (S.P.) change for the lean-oil flow controller is computed based on all feedforward (FF) information, specifications, and constraints, including a "tuning factor" from the previous computation cycle.

Before this feedforward-derived set point change is sent to the controller, a feedback correction (FBC) is computed as follows (lower right of Fig. 4):

The key variable affected by lean flow, C2 recovery, is read as the ethane content of residue from the on-line gas chromatograph and is corrected by the MVC-based "drift factor" program which keeps a running comparison of on-line chromatograph results with laboratory calibrations.

This corrected ethane value is compared with an MVC predicted optimum value computed at the appropriate delay time. This FBC difference is applied through an appropriate PID algorithm to the feedforward-computed set point, which is now dispatched to the controller and process.

Thus, MVC is one level of control above the plant control system. In the event of MVC failure or out-of-range operation, the operator resorts to regulatory control. With MVC on-line, the operator simply provides the desired operating targets, product specifications, and equipment selections.

The system optimally sets the operating process conditions to best achieve target requirements. Economic and engineering information can be entered manually as required by the operator or received remotely. The operator has the option of either manually entering the optimum ethane-recovery (RC2) values from the economics screen to the operator screen or having these values entered by MVC automatically.

In the absorber module example of Fig. 3, the operator normally views two screens: An operator screen (Fig. 5) and an economics screen (Fig. 6).

The principal manipulated variables shown in the operator screen are reflux, which is mainly the H2S specification control; reboiler duty, principally for control of CO2 in NGL; and lean-oil flow, the major control for optimum recovery of ethane.

Column pressure is a pre-selected secondary MV, and lean oil composition is an MV set optimally on-line in the rich-oil fractionator.

The optimum RC2 target, at any given set of process conditions, is shown in the economics screen as a function of ethane and CO2 values and of utility costs-fuel gas for reboilers' heat duty and electricity for refrigeration and pumping.

Changes in these economic parameters are keyed in by the operator or other authorized person.

The "drift factor" program screen shown in Fig. 7 enables the operator automatically to correct for drift of key instruments, the on-line gas chromatograph in this example.

The operator simply enters the results of laboratory gas chromatograph analyses of time-stamped process samples, whereupon MVC retrieves the corresponding on-line gas chromatograph values, computes the resulting correction or "drift" factors for each component, and upon approval from the operator installs these factors into the MVC control modules.

On keystroke command of the operator, the equation tuning program screen of Fig. 8 automatically resets the "feedback trim" for each major process parameter set by MVC.

Simply, this program recalls the differences between predicted and observed parameter values over a period of, say, one shift, and computes the appropriate correction factors for each predictive equation.

Thus, any systematic "drift" in the performance of a process unit-whether because of gradual fouling, change in variables unaccounted for, or simply imperfection of the MVC model itself-is automatically compensated for.

The relinearization program screen (Fig. 9) is reserved for recalibration of the process in the case of significant mechanical process unit change such as addition of exchangers, piping, number of trays or packing, etc.

The screen operation is similar to that of the equation tuning screen, except that the program in this case computes new numerical values for each coefficient of every term of an equation, rather than a single new overall correction factor for an equation.

The control screen (Fig. 10) enables the operator to turn on or off each MVC module. In the "on" position, the system operates in closed loop, and MVC resets the MV setpoints directly.

In the "off" position, MVC still receives and processes all information on-line, including the computation of the MV optimum setpoints but stops short of actually dispatching such setpoint commands to the unit.

STREAM CHANGES

Following a one-day factory acceptance test, the MVC system was installed on-site and has been in operation since.

After about 2 weeks of initial open-loop operation, the system was put on closed loop. On several occasions, when for unit operations' reasons MVC was switched to open loop, switching back to closed loop was seamless.

During the several months' period analyzed, inlet-gas flow and composition changed seasonally, diurnally, and many times during a given shift due to changes in the mix of inlet gas streams.

Both minute-by-minute short duration open/closed loop tests, and longer range, month-by-month historical comparisons were made to measure the effects of MVC optimization. Overall, both types of analyses showed substantial benefit from system application.

Fig. 11 shows the effect of closing the loop for a 52-hr test. The plant already had been guided near its optimum during the open-loop portion on the left half of the chart.

The effect of closing the loop on the right half shows an added 10% effect on relative ethane recovery, plus a 20% relative decrease in specific fuel-gas consumption, BTUs/bbl of NGL.

Similar effects are observed in the long-range tests conducted over a 5-month period (Figs. 12 and 13). Both the relative consumption of electricity-a measure of specific refrigerant use-and that of fuel gas showed an optimization effect in the range of 15-20%, with a closed-loop/open-loop improvement of 10-15%.

Adaptive feedforward control from MVC techniques can be designed for interaction between columns as well as between top and bottom column control without oscillation in the control loops.