Optimize Your Process Using Normal Operation Data Michel Ruel, PE Top Control, Inc. 49, rue du Bel-Air, bur.103, Lévis, QC G6V 6K9, Canada Phone +1.418.834.2242, michel.ruel@topcontrol.com Henri (Hank) A. Brittain, PhD Top Control, Inc. 2221 South Webster Ave., # 141, Green Bay WI 54301, USA Phone +1.513.543.6494, hank.brittain@topcontrol.com Optimization by tuning regulatory (PID) controllers remains one of the simplest and most cost effective means of managing and reducing variability in nearly every process today. Yet few sites have anything resembling a systematic maintenance-like approach to this, despite methodology that has been known for decades. The reason is that this decades-old methodology is time consuming, requires expert knowledge, and most often entails risky open-loop testing that few are willing to undertake. This paper will show that Tai Ji PID, software migrated from the Tai Ji model identification methodology used for MPC applications, can perform closed loop analysis, often with normal operating data. For the first time, it is now possible to optimally tune every PID loop in a plant with minimal time and risk. 1 Introduction Most plants have well thought out maintenance plans for nearly all of the assets on site, except for the soft part of PID loops. Regular maintenance is scheduled for rotating equipment such as pumps, turbines, and centrifuges. Processing equipment such as heat exchangers and columns is also included. Vibration analysis is done where appropriate. Control valves are often monitored through travel and reversals. Smart transmitters can provide information on pending failures and inconsistent operation. In spite of this, little is done to ensure that the PID brain behind the control system, that which keeps the process variables close to target, is maintained at optimal performance. In part, this speaks to the robustness of PID control; PID works well in many instances, even when tuned less than optimally. Given this, and the difficulty and time required to tune these loops, most plants neglect the need to tune their loops and blindly suffer the costs in reduced quality, increased energy use, increased raw material use, lower production rates, process trips, and a multitude of other inefficiencies. The maintenance involved with PID loops, other than the aspects mentioned above for the measurement and final control elements, involves optimal tuning of the elements in the PID algorithm (proportional, integral, and derivative terms) to meet the process needs. This is the soft part of the loop. One may argue that unlike the mechanical components mentioned above, software does not break. However, processes are subject to changes, and when processes change, tuning that once worked will provide less than optimal results, and may not work at all. Furthermore, most loops are never analyzed in any fashion to calculate tuning to begin with. Many of these have the default values installed by the DCS or Page 1
PLC integrator at the time of commissioning. The need for thorough analysis and optimal tuning to improve and maintain PID loop performance, and ultimately, process performance, is very real. 2 Traditional Methods of PID Tuning Other than trial-and-error tuning, most approaches to tuning PID loops involve moving the process in some fashion, obtaining a model based on the move and the process response, and using tuning rules to determine optimal tuning from the model or frequency response of the process. Some common methods that accomplish this are discussed below. 2.1 Open-Loop PID Tuning The first systematic approach to tuning PID controllers emerged more than 5 decades ago with the tuning rules developed by John Ziegler and Nathaniel Nichols for a pneumatic controller, employing the ultimate gain and period method. Most methods now rely on open-loop tests, whereby the loop is placed in manual and the controller output is moved. The vast majority of these tests involve moving the output in a step wise fashion. For a self regulating process (usually temperature, flow, and liquid pressure), one waits for the process variable to settle. From this response a process model is obtained, usually a first-order plus dead time (FOPDT) model, consisting of the process gain, dead time, and time constant. For a non-self regulating process (usually an integrating process; for example, level in a tank), a change in the slope, or rate of change of the process variable, provides the data for a model, consisting of a process gain, and dead time. These types of tests are well-documented and the models can be calculated by hand. Different tuning rules are available to develop tuning that meets the needs of the process. The issues with the step test are many. The primary issue is that the test will disturb the process. For fast loops such as flow or liquid pressure, this may not be a major problem but for slow processes, such as temperature, a seemingly small output move may result in a large process change that will require the test to be abandoned before completion. These loops can also take a long time for the process variable to settle; by this time disturbances may have entered the process that will render the test useless. An additional problem with these tests is the manpower required; someone must devote the time to monitor the process while the loop remains in manual mode, then calculate the model, and determine tuning based on an appropriate tuning rule. Given these issues, many resort to trial-and-error tuning, a process that in reality takes much longer than the scientific approach offered by the step test. Trial and error rarely, if ever, yields optimal tuning, meaning that more manpower has been used, with little to no benefit. Other open-loop tests can be used, which when employed in conjunction with the appropriate software tools, may shorten the test time and will lead to models of the process. The software does this either by curve fitting a time response model, or by fitting a model in the frequency domain. These tests include so called pulse and double-pulse tests. A double pulse test, when executed properly, can help move a process variable back toward setpoint and shorten the test time. A single pulse test does not have this advantage, but shares the benefit of directing the process variable back to the original value. Page 2
The issues with pulse tests, single or multiple, are fewer than with step tests but still exist. Again, the primary issue is that the test will disturb the process. Manpower is also an issue, as the loop must be monitored while in manual mode. The complexity of the double pulse test requires additional knowledge. 2.2 Closed-Loop PID Tuning Closed loop tests require software tools, and usually involve stepping the setpoint, or making single or multiple pulses of the setpoint. The key advantage over open loop tests is that the value at which the process value settles can be set. A disadvantage is that if the loop is so poorly tuned that it cycles before settling, a greater disturbance may result than in an open loop test. 3 PID Tuning Using Normal Operating Data or Low-impact Tests The methods described above are designed to establish a cause-and-effect between the controller output and the process variable. This is essential to any effort that seeks to establish a model for tuning purposes; one simply must know how the controller output affects the process in order to develop models and utilize tuning rules and develop tuning values for the PID controller. Data that meets this criterion include: 1) setpoint changes that are a normal part of the operation, 2) loops that are cycling because of poor tuning, 3) secondary loops in a cascade control strategy with setpoint activity generated by the primary loop, and 4) low-impact tests that utilize Generalized Binary Noise (GBN) type tests. Note that data from a controller response to a disturbance does not meet the criterion. The cause and effect that is sought is due to the controller moving the process variable, not the process variable responding to an external disturbance. The advantage to using this data or these tests is reduced disturbance to the process and much less operator and engineer intervention. 4 Examples of PID Tuning Using Normal Operating Data or Low-impact Tests Below are some examples where normal operating data was used to develop new PID tuning. In some cases, the data involved setpoint changes that were a normal part of the operation, or part of a start-up, in the case of a batch system. In another case, a loop was cycling due to poor tuning. Since the excitation came from within the loop, this data could also be used to develop a model and new PID tuning. A cascade example is also shown, where the secondary (inner) loop is driven by the primary loop. Lastly, a low-impact Tai Ji test is shown where the setpoint is moved in a designed fashion during the course of normal operation. 4.1 Setpoint changes that are part of the normal operation Depending upon the process, setpoint changes are often a part of normal operation. In other cases, operators may occasionally adjust a setpoint. Figures 1 through 3 below show the tuning results from setpoint changes as part of the normal operation. During fuelling of a rocket, the temperature of heated nitrogen blown onto the rocket and gantry is increased. Nitrogen is used to eliminate one-quarter of the combustion tetrahedron. As the rocket is fuelled with liquid oxygen and hydrogen, the setpoint is stepped up several times to prevent ice from forming on the exterior of the rocket. Ice adds to the vehicle mass, and may also damage the rocket. Page 3
Figures 4 through 6 show the data from the start-up of a batch reactor. At start-up, the reactor setpoint is stepped from 80 to 120 o F. This data is sufficient for tuning analysis. Figure 1. The temperature of heated nitrogen on a rocket gantry is shown in this plot. During fuelling of the rocket, the temperature setpoint is stepped from 60 to 90 to 110 to 130 o F, as the cryogenic fuels are loaded. The response is becomes more aggressive as the setpoint is increased, indicating the process is nonlinear. Tuning values at lower temperatures do not work well at higher temperatures. Page 4
Figure 2. By slicing out the first two setpoint changes in the data contained in Figure 1, a model is obtained from the remaining data contained in the last response. New tuning is calculated for this operating point, and is implemented as adaptive tuning in the figure below (e.g. the tuning is changed when the higher setpoint change is made). Page 5
Figure 3. A lower controller gain is implemented when the setpoint is stepped from 110 to 130 o F, resulting in a satisfactory response. Page 6
Figure 4. During start-up the temperature setpoint of a batch reactor is stepped from 80 to 120 o F. The time scale from setpoint change to the end of the plot is 5 minutes (given for reference with Figure 6). Figure 5. The early part of the data from above yielded a suitable model quality, and new tuning was determined from this. Page 7
Figure 6. Batch reactor, after tuning. The time scale from setpoint change to the end of the plot is about 5 minutes (given for reference to figure 4). Page 8
4.2 Loops that are oscillating due to poor tuning In some cases, loops are cycling due to poor tuning. This situation is not unlike the ultimate gain and ultimate period method proposed by Ziegler and Nichols. If the response of the process variable is due solely to the movement of the controller output, this data is also suitable for tuning analysis. Figure 7. The data shown is for a cascade level-flow loop in a degasifier. The tuning on the level loop is driving the oscillation. This in turn drives the secondary loop, the flow. Page 9
Figure 8. Analysis of the oscillating level-flow cascade loops results in good model fits. Note the disturbance at the end of the data set resulting from switchover to another flow loop. The process employed multiple secondary loops feeding de-ionizers that were routinely switched over to allow regeneration. Page 10
Figure 9. Tuning rules are based on IMC closed loop time constants. From the oscillating data, new tuning was determined and employed. See Figure 10 for results. Page 11
Figure 10. Data taken from a screenshot of the plant historian. The loop ceases cycling when new tuning is installed (at about midway in the plot). Page 12
4.3 Secondary cascade loops As disturbances drives the primary controlled variable from setpoint, the secondary loop will follow a setpoint as directed by the primary loop. Note in this case that the primary loop process variable is driven by the disturbance, so the primary process variable movement is due to external loop forces, and is thus not suitable for tuning. But the movement of the secondary loop is due to the loop itself, so should lend it self to tuning analysis. In the example below, continuing disturbances drives the primary loop, the main steam temperature, from setpoint. The secondary loop follows the setpoint generated by the primary loop. Note that neither loop was settled in the data below. Being settled is often a requirement of many a requirement of many tuning methods and software tools. Figure 11. Disturbances to the primary loop, TIC2117, cause the primary loop to move the setpoint to the secondary loop, TIC2118. The model quality for the secondary loop is A, suitable for tuning. The model quality for the primary loop is D ; unsuitable for tuning. This is expected as the primary process variable movement is caused by a disturbance external to the loop. 4.1 A low-impact Tai Ji test Tai Ji uses a Generalized Binary Noise (GBN) approach that is more or less a low frequency version of a Pseudo Random Binary Step (PRBS) test. A high frequency dither is added on top of the low-mid frequency pattern to improve dead time estimation (which is a high frequency component). Page 13
A key advantage to this test is that it can run during normal operation, even over night. Figure 12. A closed loop Tai Ji test for a cascade temperature loop entails specification of the amplitude of the setpoint moves. Tai Ji moves the setpoint +/- this amount in such a way as to ensure the mean SP is stationary, unless the operator changes the setpoint. Page 14
Figure 13. The completed Tai Ji test is shown for a cascade temperature control loop. One temperature excursion beyond the normal movement of the process variable is noted (at about 13:27). Page 15
Figure 14. The model quality for both the primary and secondary loops in the cascade temperature control loop are an A. 5 Conclusions and Discussion It is now possible to tune every PID loop in a process safely and quickly, in closed loop mode, often with normal operating data. The criterion for this is that the data contain sufficient excitation of the process variable and that the excitation comes from the controller output. This paper has shown that movement of the process variable caused by setpoint changes during normal operation and start-ups, from poor tuning that causes oscillations, from primary controllers in a cascade strategy, and from low impact tests that move the setpoint are sufficient for tuning PID controllers. Since the data may already exist, no manpower is required to sit with and monitor the process as is the case with risky open-loop tests. Little specialized knowledge is needed. If the data does not already exist, or insufficient movement of the process variable (by the controller output) occurs during normal operation, a Tai Ji test can be conducted. This involves specifying the amplitude of the desired setpoint moves during the test. Alternatively, the operator can be requested to make a few setpoint changes during the shift. The test can be done overnight, and the results evaluated the following day. Improved results with reduced manpower will follow. References Blevins, T.L., McMillan, G.K., Wojsznis, W.K., Brown, M.W. (2003). Advanced Control Unleashed. ISA Press, Research Triangle Park, NC. Page 16
Zhu, Y.C. (1998). Multivariable process identification for MPC: the asymptotic method and its applications. Journal of Process Control, Vol. 8, No. 2, pp. 101-115. Page 17