Você está na página 1de 7

Control Engineering Practice 9 (2001) 11771183

PID-deadtime control of distributed processes


F.G. Shinskey*
Process Control Consultant, 260 Whiteface Rd., North Sandwich, NH 03259, USA Received 6 April 2001; accepted 6 April 2001

Abstract While model-based controllers have been used successfully to control paper machines and other processes dominated by time delay, matching the model to the process gives poor load regulation over lag-dominant processes. An important class of lagdominant processes including heat exchangers and distillation columns consists of distributed lags. A PID controller having timedelay compensation, while functionally similar to a model-based controller, is a much better load regulator, and twice as eective as a conventional PID controller on these processes. It is applied in this paper to regulate steam superheat temperature. r 2001 Elsevier Science Ltd. All rights reserved.
Keywords: Adaptive control; Distillation columns; Distributed-parameter systems; Heat exchangers; Load regulation; Model-based control; PID control; Steam plant; Temperature control; Time delay

1. The PID-deadtime controller The insertion of a time delay into the integral feedback circuit of a PID controller produces a transfer function similar to a Smith predictor or an Internal Model Controller. Its internal conguration is shown in Fig. 1. Integration is accomplished by positive feedback of controller output m through time delay td and integral lag I. The controlled variable c passes through a ltered derivative block of time constant D and lter aD; but set point r does not. An additional lter tf is located so as to be in both the process and integral feedback loops. The transfer function of m responding to c is ms Kc 1 Ds1 Is ; cs 1 aDs1 Is1 tf s etd s 1

against integral windup and prepared to resume control when the feedback connection is restored. An internal model controller (IMC) is shown in Fig. 2 for comparison. Its transfer function is ms 1 ; cs Km gm * =gf gm

where Km is the steady-state gain and gm is the dynamicgain vector of both the process and its model; gm* is the invertible part of gm, and gf is a lter. For the simplest case where the process is a rst-order lag with deadtime, the IMC function becomes ms 1 t1 s ; cs Km 1 tf s etd s

where s is the Laplace operator and Kc is the proportional gain of the controller. This PID form is known as the interacting or series controller; time-delay compensation can also be added to the noninteracting or parallel controller. The connection from the controller output to the time-delay compensator is shown dashed because it can be broken, allowing another signal to be substituted for the output signal. In this way, the controller can be protected
*Tel/fax: +1-603-284-6404. E-mail address: shinskey@msn.com (F.G. Shinskey).

where t1 is the process lag, tf the time constant of a rstorder lter, and td the process deadtime. If the integral time of the PID controller with deadtime were set to zero as well as its derivative lter, the transfer function of Eq. (1) would reduce to that of Eq. (3). The PID controller with deadtime is then seen to be similar to a second-order IMC controller. However, this paper is intended to illustrate their dierences more than their similarities.

0967-0661/01/$ - see front matter r 2001 Elsevier Science Ltd. All rights reserved. PII: S 0 9 6 7 - 0 6 6 1 ( 0 1 ) 0 0 0 6 3 - 6

1178

F.G. Shinskey / Control Engineering Practice 9 (2001) 11771183

Fig. 1. Time-delay compensation is added to the integral feedback loop of the PID controller.

Fig. 3. For lag-dominant processes, IMC gives unacceptable load response.

Fig. 2. Internal model control conguration.

2. Limitations of IMC While model-based controllers are intended to match the process parameters as closely as possible, so as to produce a particular set-point response, the author (Shinskey, 1996) has shown that this results in poor load regulation for processes that are lag-dominant. The problem is illustrated in reference to the block diagram of IMC in Fig. 2. In parallel with the process is its model, also identied as Kmgm, assuming a perfect match. This produces an estimated value of c designated as ce, which is subtracted from its real value. Their dierenceFgiven a matched modelFis the process load q after passing through its steady-state and dynamic gains Kqgq. The IMC is designed to produce a desired set-point response: cs g s gf s m rs gm * s 4

Given only rst-order dynamics, c will follow an exponential trajectory until the deadtime in gm elapses, then returning to set point on another exponential trajectory, both governed by gq. The time lag in the load path for the curve in Fig. 3 is only twice the deadtime. Also shown is a recovery trajectory which returns to set point at the end of the deadtime, a response identied as the best achievable. It is obtained by causing m to overshoot the step in q, with the overshoot sustained for one deadtime. The ratio of the integrated error IE for the two responses is IEbest 1 etd =tq ; IEIMC 1 tf =td 6

which includes only the process deadtime and the lter. The load response, however, includes the load dynamics as a dominant term:   cs gm s Kq gq sgf s 1 : 5 qs gm * s Following a step change in load, after the deadtime in the load path elapses the controlled variable will begin to deviate from set point. At that time, controller output m will step an amount equal to the step in q multiplied by Kq/Km. Fig. 3 illustrates the case where Kq=Km.

where tf is the time constant of a rst-order lter in the IMC controller. Even without a lter, the curves in Fig. 3 show an integrated error for IMC that is 2.5 times the best. For most uid processes where temperature and composition are controlled, the ratio of time constant in the path of the load input to the deadtime is 510, giving integrated-error ratios of 5.510.5. The advantage of the PIDtd controller lies in that it is tunable, much like a conventional PID controller. And when tuned to minimize integrated absolute error (IAE) or some similar objective function, it outperforms both PID and model-based controllers, approaching the best load response without compromising set-point response. In this paper, the PID controller with timedelay compensation is evaluated as a regulator for distributed-lag processes.

3. Distributed-lag processes Multiple-stage chemical processes consist of a series of connected rst-order lags of similar time constant. If

F.G. Shinskey / Control Engineering Practice 9 (2001) 11771183

1179

information can ow in both directions through such a process, these lags are considered to be interacting. In a distillation column, for example, a step change in reux ow entering the top will aects the composition of the overhead vapor almost immediately. But it also changes the ow and composition of the liquid leaving the top tray, which subsequently aects the composition of the vapor leaving the next tray. This new vapor composition then shifts the equilibrium on the top tray, causing a secondary change to the composition of the overhead vapor. Each change in either the ow or composition of either the vapor or liquid thus propagates both upward and downward, ultimately aecting the composition of both overhead and bottom products in a series of steps. If the stages are few in number, the individual steps can sometimes be seen in the response of product compositions to step disturbances. But most columns contain 20 or more trays, whose capacity smoothens the composition response curves into long exponential lags. An electrical analog of such a multistage process is a series of identical resistors with identical capacitors connected across the line at each junction, forming a ladder network. This network has also been used to simulate the response of electrical and pneumatic transmission lines. 3.1. Open-loop response Fig. 4 shows a step response of a network of 20 equal interacting lags. Note that it is dominated by the exponential response of an equivalent rst-order lag, but also with some initial time delay. The time scale is normalized so that 63.2% response occurs at a normalized value of 1.00. The time at this point is St ti n2 n=2; 7

sum of the numbers from 1 to n. This squared relationship can give a 100-tray distillation column a four-times slower response than a 50-tray column. Interestingly, there is almost no dierence in the shape of the response curve when n=100Fit does not change signicantly for values of n>20. As a consequence, the 20-lag model is satisfactory for simulating much higherorder systems. The highest-order interacting system is the distributed lag, consisting of an innite number of innitesimal lags. Distillation columns containing packing rather than discrete trays behave as distributed lags. Perhaps the most common distributed process is the heat exchanger: heat is transferred across a broad area having a distributed temperature gradient and distributed heat capacity as well. 3.2. Variable parameters Distillation columns tend to operate with internal circulation rates of liquid and vapor which are much higher than their throughput, especially when there are many trays. As a result, their dynamics tend to be unaected by production rate. Another example of a distributed lag with constant dynamics is a stirred tank, providing that its internal circulation rate is reasonably constant and greater than throughput. By contrast, heat exchangers generally operate on a once-through basis, neither uid being recirculated. Given that an exchanger has a xed heat capacity, the variable ow rate of the uids causes residence time St to vary inversely with ow. The shape of the response curve does not change, but both its equivalent time delay and rst-order lag change with ow, while keeping the same ratio between them. The steady-state response of extrinsic variables such as temperature and composition also depends on ow rates. Consider the following steady-state model of heat transfer between steam owing at rate W and having a latent heat of H, and a liquid owing at rate F and having a specic heat of C: Q WH FCT2 T1 ; 8

where ti is the value of the individual time constant and n is their number. The factor (n2+n)/2 is equal to the

where Q is the rate of heat ow and T1 and T2 are the respective inlet and outlet temperatures of the liquid. The steady-state gain of exit temperature in response to steam ow is seen to vary inversely with liquid ow: Kp dT2 H : dW FC 9

Fig. 4. The step response of 20 interacting lags; the shape does not change signicantly above 20 lags.

As liquid ow increases, the exit temperature responds to changes in steam ow faster and with less gain, making it easier to control. However, control-loop stability must also be provided at low ow rates, where the response is slower and the gain is higher.

1180

F.G. Shinskey / Control Engineering Practice 9 (2001) 11771183 Table 1 Controller settings and integrated error for optimum load regulation of distributed processes Controller PI PID PIDtd P/Kp 20 15 6.6 I=St 0.54 0.25 0.067 D=St F 0.10 0.16 td =St F F 0.16 IE/Dm 0.11 (KpSt) 0.038 (KpSt) 0.015 (KpSt)

3.3. Closed-loop response Modeling a distributed-lag process as rst-order plus time delay is not very satisfactory in predicting its closed-loop response. The distributed lag is much more responsive to derivative action in the controller. For example, the step load response of a rst-order process with deadtime can be reduced in integrated error by a factor of 2.25 by changing from PI to interacting PID control. But for a distributed process, the reduction achieved by the same change in controllers is 2.89. The optimum controller tuning is somewhat dierent as well. Proceeding from a PID controller to one with time-delay compensation is even more striking: the integrated error is reduced by a factor of 1.5 for the rst-order-plus-delay process and 2.5 for the distributed process. As a consequence, the PIDtd controller seems particularly well-suited to regulating distributed processesFover seven times as eective as PI control on the basis of integrated error. Fig. 5 compares the step load response of a distributed process (simulated by 20 interacting lags) using three dierent controllers: PI, interacting PID, and interacting PIDtd : All controllers were tuned to minimize integrated absolute error (IAE), Table 1 lists their mode settings in terms of the process parameters Kp and St: Proportional band P is expressed in percent. (The proportional gain Kc=100/P.) The last column in the table is the integrated error (IE) per unit change in controller output m required to meet the load change, calculated as IE P I td : Dm 100 10

to avoid lifting safety valves, and steam superheat, where extreme temperatures shorten the life of metal heat-transfer tubes. The size of the peak deviation following a load change is directly proportional to the proportional-band setting. So the progression of proportional-band settings in the table relates to the peak deviations shown in Fig. 5. 3.4. Tuning vs. matching Model-based control was developed primarily for processes having a pronounced time delay, the intent being to match the process delay with one in the control system. Improved set-point response was promised, along with the elimination of tuning, except for that of the lter. In the distributed process, there is no true time delay, yet a controller having a time delay is more eective regulating it than those that do not. Furthermore, if the PIDtd controller were replaced with a model-based controller matching the process faithfully, the load response would be much poorer, due to the dominant lag in the load path. So the concept of matching the controller to the process must be rejected on two counts: a higher performance can be achieved not only by tuning the controller instead of matching its parameters to the process, but by using a controller whose structure does not match the process, either. In other words, the high performance of the PIDtd controller on the distributed process refutes the notion that model-based control is superior to other methods. What is proved instead is that the PIDtd controller has a high performance capable of maximizing load response on a variety of processesFits structure does not depend on the nature of the process being controlled, although its tuning does. While elimination of tuning may be desirable, this exercise shows that it is still required if load rejection is to be optimized, especially for a lagdominant process. (Although not a part of this discussion, the set-point response of this controller is also excellent with the tuning that is optimal for load response.)

Integrated error can be related to operating cost in most product-quality loops, where it represents valuable ingredients given away or excessive energy consumption used to assure that specications are met. In some applications, however, the peak deviation may be as important. Examples are the control of boiler pressure,

4. Control-loop robustness
Fig. 5. Step load responses for a distributed lag; all three controllers were tuned to minimize IAE.

Robustness has been dened as the minimum change in given process parameters which brings the loop to the

F.G. Shinskey / Control Engineering Practice 9 (2001) 11771183

1181

limit of stability, as identied by an undamped oscillation. As a general rule, the higher the controller performance, achieved either through tighter tuning or the addition of more modes, the lower the robustness of the loop. (The converse is not necessarily true, in that it is possible for a controller to exhibit low performance and low robustness at the same time, which has been the experience of some engineers attempting to control pH or steam temperature using a Smith predictor.) Adjustments to proportional, integral, and derivative mode settings provide tighter control when moved in one direction (higher gain, faster integration, longer derivative time), and more robustness when moved in the other (lower gain, etc.). Time-delay compensation, however, can cause instability when pushed too far in either direction. Therefore, the increase in performance which it brings, comes with the requirement of very precise tuning. 4.1. Robustness considerations with heat exchangers There are only two parameters required to specify the distributed lagFKp and StFand as indicated in Table 1, to tune the controller. However, in the case of a heat exchanger, they do not vary independently, but together, inversely with process ow. As a consequence, robustness properties for the three loops described in Table 1 and Fig. 3 need to be evaluated as these two process parameters change together with ow. In simulations of the PI loop, the limit of stability was reached when the ow fell to 63% of the value where the controller was optimally tuned. For the PID loop, instability was reached at 69% of the original ow. With the PIDtd controller in the loop, instability was reached when the ow fell to 79% of the original, or rose to 112% of the originalFthe last being the narrowest of the margins observed. For this application, then, the robustness of the PIDtd controller is only 12%, compared to 31% for the PID controller and 37% for the PI controller.

Fig. 6. Estimating the time delay and time constant using an openloop step response.

5.1. Evaluating the reaction curve Ziegler and Nichols apparently did not want to wait to achieve complete response of lag-dominant processes, and so terminated their test after observing the slope of the reaction curve pass its steepest point. This has merit, in that the dominant lag of a distillation column could be more than an hour, and ve time constants are required for 99% complete response. In the meantime, other disturbances could arise. Their method, therefore, used only two features of the curve upon which to base their controller settings: the estimated time delay and the steepest slope of the curve. A self-regulating process really requires three parameters for identication: time delay, time constant, and steady-state gain. However, an estimate of the gain requires a return to the steady state, which may require too much time, and may not be reached at all. Reasoning that the steepest slope was a function of both the gain and the time constant, they replaced these two parameters with one. Fig. 6 shows that the estimated time delay tde lies between the initiation of the step at time zero, and the intersection of the steepest slope with the baseline. In an alternative to calculating the slope, Fig. 6 shows how to estimate the dominant time constant t1e ; as if the steadystate gain were unity. The size of the step disturbance Dm is marked against the steepest slope, projected if necessary, and the time along that slope is the estimated time constant.

5. Process identication If the PID controller with time-delay compensation is to be used to control a heat exchanger, the dynamics of the process need careful identication across the entire operating range. The earliest and easiest identication method to administer is the step test used by Ziegler and Nichols (1942). A step change is manually introduced into the controller output from a steady state, and the resulting response in the controlled variable recordedFthe reaction curve. The step-response curve for the distributed lag from Fig. 4 is analyzed in Fig. 6.

5.2. Application to the distributed lag One of the concerns when using the Ziegler and Nichols tuning rules is their limited scope. They work best for step load changes applied to lag-dominant processes. In that the distributed lags are in fact lag dominant, having a consistently low ratio of time delay

1182

F.G. Shinskey / Control Engineering Practice 9 (2001) 11771183

to time constant, their method is reasonably eective here. To make use of the tuning rules given in Table 1 requires conversion from the parameters of tde and t1e to Kp and St: Given that all distributed processes have the same shape for their reaction curve, a direct correlation is possible: Kp 7:5tde =tle St 7:0tde : 11

factors are also adjusted, including the tilt of the burners and the recirculation of ue gas. Therefore, the process parameters cannot be expected to vary with steam ow following a simple formula such as Eq. (9), but must be determined on-line by testing at several load levels. 6.1. Plant identication The PIDtd controller was applied to the superheaters of a 500 MW power boiler in Ontario, Canada. Steam temperature was to be controlled at 538751C. Because of the robustness limitations of the controller, the process parameters had to be determined precisely across the full load range, and the mode settings would need to be gain-scheduled as a function of steam ow. This means adapting all the mode settingsFnot only the gainFbased on the observed response of the process, as it changes with load. Open-loop step tests were, therefore, conducted at load levels from 100 through 530 MW, with the time delay and time constant estimated as in Fig. 6. The results of those tests appear in Table 2. Over the range tested, generated load varied by 5.3 : 1. However steam ow is the variable used to index the controller settings, and that changed over a 4.5 : 1 range. Based on Eq. (9), Kp would be expected to vary as much, but its actual variation was only 2.3 : 1, half as much. There is even a slight reversal in the estimate of Kp, probably caused by a change in another operating variable such as ue-gas recirculation. The dynamic parameter St followed steam ow more closely, showing a variation of 3.3 : 1. Yet even these moderated variations are huge compared to the narrow robustness limits demonstrated by the PIDtd controller in simulated closed-loop testing, where a 12% ow increase moved the loop from optimum performance to undamped cycling. The controller would have to remain tuned very precisely to deliver the expected performance improvement over PID control. 6.2. Implementation The temperature controller was implemented by inserting a time-delay compensator in the integral feedback loop of an interacting PID controller, just as shown in Fig. 1. After obtaining the estimated process

The distributed-lag process has, as seen in the second correlation above, an equivalent rst-order time constant of 6 times its delay. In this context, St represents the sum of the equivalent rst-order lag and delay. The rst correlation above has no other function than to identify the steady-state gain. Earlier, it was mentioned that a rst-order lag with deadtime is not a particularly useful simulation of a distributed process. To demonstrate, if the deadtime and time constant parameters estimated from the stepresponse curve are used to calculate the best integrated error for a unit load change, as shown for a rst-order process in Fig. 3, its value comes out to be 0.0219 Kp St: Then estimating the integrated error produced under PIDtd control by inserting its settings from Table 1 into Eq. (10) gives a result of 0.0150 Kp St: This would indicate a controller performance of 0.0219/ 0.0150 or 146%, a value that is clearly unrealistic. The best integrated error estimate really only applies to processes having deadtime, which the distributed lag does not. Controller performance on this process can only be estimated relative to other controllers. Therefore, the use of the parameters estimated from the reaction curve is limited to determining optimum controller settings.

6. Controlling a superheater One of the most dicult loops in a modern power plant is that of controlling superheated steam temperature. The eciency of the generating station depends strongly on maximizing this temperature, but within very narrow constraints. Allowable temperature is limited by the ability of the steel-alloy heat-transfer tubing to retain its strength. Excessive temperatures, and especially temperature variations, cause stress and distortion which can signicantly shorten the life of a superheater. The steam temperature is measured after leaving the last of a series of heat-transfer sections, but is controlled by spraying attemperating water between upstream sections. While the process is truly a distributed lag, and its parameters generally vary inversely with ow, the relationships in the context of the boiler are somewhat more complex. As steam ow is increased to satisfy increasing demand for electrical power, other

Table 2 Parameters estimated from step tests Load (MW) Steam ow (%) tde (min) t1e (min) Kp 100 200 300 400 530 20 35 50 66 89 3.5 2.5 2.3 1.35 1.05 5.0 4.8 4.3 3.5 3.5 St (min)

5.25 24.5 3.90 17.5 3.98 16.1 2.89 9.5 2.25 7.4

F.G. Shinskey / Control Engineering Practice 9 (2001) 11771183

1183

parameters in Table 2, the controller was then tuned following the rules given in Table 1. After achieving stable, responsive control at one load level with these settings, the controller was retuned at each of the other levels, with similar results. Remarkably, no ne tuning was required at any loadFthe rules which had been developed by simulation of a process consisting of 20 interacting lags proved to be entirely satisfactory. This is an important consideration, in that the PIDtd controller is not easy to tune on-line. It is common experience to nd even the derivative setting of PID controllers left at zero by operators and technicians because it complicates the tuning procedure. Adding a fourth parameterF deadtimeFcomplicates the procedure much more. In addition, cycling can develop if the deadtime setting is either too long or too short, and it is not always clear in which direction to adjust it when attempting to stabilize a cycling loop. Finally, the optimum settings found for each of the load levels were programmed into a gain-scheduler, which interpolated between the values obtained by

testing. In this way, the controller would remain optimally tuned at all loads. Control of steam temperature proved to be tighter than had ever been achieved for that loop previously. At the time of writing this paper, the controller has been providing satisfactory performance for over a year. Acknowledgements The author gratefully acknowledges the work of * Sigifredo Nino of Foxboro Canada, who conducted the plant tests and implemented the control system so successfully.

References
Shinskey, F. G. (1996). Process control systems (4th ed.) (pp. 130132). New York: McGraw-Hill. Ziegler, J. G., & Nichols, N. B. (1942). Optimum settings for automatic controllers. Transactions of the ASME, 759768.

Você também pode gostar