0 evaluări0% au considerat acest document util (0 voturi)
471 vizualizări11 pagini
The document summarizes a 1967 paper that developed tuning rules for minimizing error integrals like IAE (Integral of Absolute Error) and ITAE (Integral of Absolute Error multiplied by Time). The rules were aimed at optimizing disturbance response for first-order plus dead-time processes using PI and PID controllers. While testing showed the ITAE rules worked best, all the rules pushed loops close to instability due to low stability margins, requiring detuning for real-world use.
The document summarizes a 1967 paper that developed tuning rules for minimizing error integrals like IAE (Integral of Absolute Error) and ITAE (Integral of Absolute Error multiplied by Time). The rules were aimed at optimizing disturbance response for first-order plus dead-time processes using PI and PID controllers. While testing showed the ITAE rules worked best, all the rules pushed loops close to instability due to low stability margins, requiring detuning for real-world use.
The document summarizes a 1967 paper that developed tuning rules for minimizing error integrals like IAE (Integral of Absolute Error) and ITAE (Integral of Absolute Error multiplied by Time). The rules were aimed at optimizing disturbance response for first-order plus dead-time processes using PI and PID controllers. While testing showed the ITAE rules worked best, all the rules pushed loops close to instability due to low stability margins, requiring detuning for real-world use.
I came across the Minimum IAE and other error-integral tuning rules very early in my career, but until recently I did not have the original paper describing the development of these rules. A few weeks ago I contacted Dr. C.L. Smith to get a copy of the paper he coauthored in 1967: Tuning Controllers with Error-Integral Criteria (Reference at end of this article). The error-integral tuning rules described in this paper minimize the area that develops over time if a process variable deviates from its setpoint, shown as shaded areas in Figure 1.
Figure 1. Shaded areas indicate the integral of absolute error after a disturbance to a process. Explain, please! You may wonder exactly what you are looking at in Figure 1, so let me explain. Imagine we have a running pump for pumping chemicals into a reactor, and a second pump on standby (Figure 2). If the operator starts the second pump, the flow rate (process variable) will increase. As a result, the flow controller will close the control valve a little to get the flow back to setpoint. The control action is not shown in Figure 1, but it is so strong that it over-corrects, causing the flow rate to undershoot the setpoint and oscillate a few times before finally settling out. The shaded areas between the process variable and setpoint are the integral of the error.
Figure 2. The total flow rate will be disturbed when pump 2 starts up. Tuning Objectives So, lets get back to the technical paper. The paper describes tuning rules for minimizing several error-related integrals: 1. Integral of the error squared (ISE) 2. Integral of the absolute error (IAE) 3. Integral of the absolute error multiplied by time (ITAE) The authors recognized that a controllers integral and derivative times should be based not only on the length of the process time constant (or dead time as in the case of Ziegler & Nichols), but also on the ratio of dead time to time constant (td / tau). Optimized for Disturbance Response The tuning rules were developed for optimizing a control loops disturbance response. Tuning for a setpoint change will require different controller settings. However, Smith and Murrill did develop tuning rules for minimizing IAE and ITAE on setpoint changes, but these are outside the scope of this article. Target Process The authors also stated that the rules were developed for a 1 st -order plus dead-time process. One shortfall is that their tuning rules were designed only for processes with time-constants equal to or longer than dead times (tau >= td). This is not a major restriction since most processes do fall in this category. But it excludes these tuning rules from being used on dead-time dominant processes. Target Controller and Tuning Rules The authors developed tuning rules for P, PI and PID controllers for a non-interactingcontroller algorithm with controller gain, integral time, and derivative time. Since P-only control is hardly ever used, I list only their PI and PID tuning rules in Table 1 below. The process characteristics are denoted in Table 1 by: - Process gain = gp - Dead time = td - Time Constant = Greek letter tau
Table 1. Minimum ISE, IAE, and ITAE Tuning Rules. Click to enlarge. Performance Assessment So, how well do these tuning rules perform? I tested them on simulations of flow and temperature control loops. The PI control rules left the temperature loop with a very oscillatory response, but the PID tuning rules worked a bit better. As expected, the PID rules did not work well at all on the flow loop. Their Minimum ITAE tuning rules seemed to work the best in my opinion because they had the fewest oscillations (Figure 3).
Figure 3. Minimum ITAE tuning on a simulated temperature control loop responding to a disturbance. Low Stability Margin My biggest concern is that all of the tuning rules pushed the control loop very close to instability. To analyze stability, we can calculate a loops stability margin. The stability margin tells us how much the process gain can increase before a loop will become unstable with its current tuning settings. The stability margin on the temperature loop under Minimum ITAEs PI control was only 0.7. The PID loop was marginally better at 1.1. Their other tuning rules fared even worse. The control characteristics of industrial processes can change substantially based on valve position, process throughput, head pressure, pH, etc. Using these tuning rules on real processes will very likely cause stability problems because of their small stability margins. In all fairness this would be similar to using the unmodified Ziegler- Nichols, Cohen-Coon, and many other high-performance tuning rules. The solution would be to detune the controller (use a lower controller gain) to increase the stability margin and tolerate these changes. I normally use a stability margin of 2 to 3 on loops I tune, sometimes even more. But this would contradict the authors original tuning objective of super-fast response. This is yet another example of the inevitable tradeoff between a loops speed of response and stability. Reference to Minimum IAE Paper Tuning Controllers with Error-Integral Criteria, A.M. Lopez, J.A. Miller, C.L. Smith, and P.W. Murrill, Instrumentation Technology, November 1976, pp. 57 62. More Info Read more about tuning methods and loop stability in Process Control for Practitioners. Stay tuned! Jacques Smuts OptiControls Inc. Posted in 4. Controller Tuning 3 Responses to Minimum IAE Tuning Rules Ron Compton: February 17, 2013 at 2:32 pm Is loop tuning software recommended rather then manually calculating the tuning parameters? Jacques: February 20, 2013 at 8:13 pm Ron, I definitely recommend using a good software program for tuning. It makes the model identification, tuning, and analysis so much easier. However, a tuning tool is no replacement for understanding your process, PID controllers, and the tuning process. Note that there are more than a dozen commercially available tuning software packages. Make sure you try out a few and pick one that is easy to use and meets your criteria. - Jacques Peter Nachtwey: April 3, 2013 at 12:04 pm I agree that minimum IAE, ITAE or SSE will result in barely stable systems. One gets more robust results if you add extra terms to the cost function that is being minimized. For instance one can add a cost for the magnitude of the control output or changes in the control output. However, if one does this then it isnt simply a minimum IAE, ITAE or SSE any more and is closer to linear quadratic control. My other problem with minimum IAE etc, is that it minimizes what case? Minimizing the the response in a step from 0 to 1 will be different from minimizing the step between 0 and 2. Also, minimum IAE etc methods dont take into account that the control output is limited. I calculated the minimum IAE, ITAE and SSE coefficients using Mathcad and Scilab. Scilab is free.
Measures of controlled system performance Before thinking about how we tune a controller (i.e. adjust the controller parameters to produce an optimal controlled response) we need to decide what constitutes a good response. This turns out not to be an easy question to answer! There are, in fact, many different measures which can be used to compare the quality of controlled responses. Academic measures The term 'academic' is not intended to be pejorative (especially since I am one!). The control measures describe in this section are very precise and give exact comparisons between different control schemes, or different sets of tuning parameters, and are widely used in academic journal papers and simulation studies. They are also completely useless for measuring the performance of real control systems. The three commonly used measures are Integral Squared Error (ISE), Integral Absolute Error (IAE) and Integral Time-weighted Absolute Error (ITAE), and are defined as:
All the measures require a fixed experiment to be performed on the system (i.e. a fixed setpoint or disturbance change) and the integrals are evaluated over a fixed time period (in theory to infinity, but usually until a time long enough for the responses to settle). ISE integrates the square of the error over time. ISE will penalise large errors more than smaller ones (since the square of a large error will be much bigger). Control systems specified to minimise ISE will tend to eliminate large errors quickly, but will tolerate small errors persisting for a long period of time. Often this leads to fast responses, but with considerable, low amplitude, oscillation.
IAE integrates the absolute error over time. It doesn't add weight to any of the errors in a systems response. It tends to produce slower response than ISE optimal systems, but usually with less sustained oscillation. ITAE integrates the absolute error multiplied by the time over time. What this does is to weight errors which exist after a long time much more heavily than those at the start of the response. ITAE tuning produces systems which settle much more quickly than the other two tuning methods. The downside of this is that ITAE tuning also produces systems with sluggish initial response (necessary to avoid sustained oscillation). A VisSim model with a PI controlled tuned to optimise each of the three measures can be found here. The reason that the three measures can't be used in practical system comparisons is that they require a carefully controlled experiment, where only a particular parameter is changed in a predefined way. Although it may sometimes be possible to perform experiments on real plant, it is impossible to stop random disturbances affecting the process during an experiment. The random disturbances will affect the process output and hence alter the integrated error measures in an unrepeatable way. Practical Measures The performance measures described in this section are not nearly so precise as IAE, ISE or ITAE, but have the advantage that they are easy to apply to real systems. Some measures can only be applied to controlled responses to setpoint changes (in real systems, setpoint changes are much easier to make than disturbance changes - by their nature disturbances can't usually be manipulated).
Some practical performance measures 1. Steady-state offset - this is very important for controlled systems without integral action. It was discussed in the section on proportional control. 2. Percentages overshoot. This is a quantity which is only meaningful for set point response. It can be estimated from the process response curve as:
The percentage overshoot is, in linear systems, independent of the size of the setpoint change (for small changes it is almost independent in non-linear systems). Large overshoots are undesirable and percentage overshoots exceeding 100% probably indicate instability. 3. Maximum absolute overshoot. This applies to disturbance and setpoint responses and is the value 'A' on the diagram. The size of the absolute overshoot is dependant on the size of the setpoint or disturbance change causing it, but is a useful measure when the system is to be tested by known setpoint or disturbance changes. 4. Rise time. The rise time only has meaning for setpoint changes. It is defined at the time taken for the process response to first reach the new setpoint. It is marked as 'TR' on the diagram. Short rise times are good, but a fast response here usually comes at the cost of increased overshoot and oscillation. The rise time for linear systems is independent of the size of the setpoint change. 5. Period of oscillation. The period of oscillation is shown as 'To' on both diagrams and is the length of time for the oscillation to complete one cycle. Provided the oscillation is adequately damped, short oscillation periods are preferred since this will settle the system more quickly. However, very short oscillation periods generally cause large overshoots and poor damping. 6. Settling time. This is the time taken for a response to settle within some band around the setpoint. It can be defined for both setpoint and disturbance changes. Shorter settling times are best, although they can lead to long rise times and generally sluggish response. 7. Decay ratio. This is defined for both setpoint and disturbance responses. The decay ratio is the ratio of the height of successive peaks of the process response. The diagrams show the heights of the first two peaks as 'A' and 'B' and the decay ratio is given by B/A. Although the first two peaks are normally used, in most system the decay ratio stays approximately constant throughout the response. The most widely used measure of practical control performance is the decay ratio as it gives a good indication of the stability of the controlled response. An often used rule for controller tuning is 'one-quarter decay ratio', where the controller is tuned to produce a decay ratio of 1/4 (the second peak is a quarter the height of the first).
Learning PID Tuning III: Performance Index Optimization Most PID tuning rules are based on first-order plus time delay assumption of the plant hence cannot ensure the best control performance. Using mordern optimization techniques, it is possible to tune a PID controller based on the actual transfer function of the plant to optimize the closed- loop performance. Contents Performance Indices Initial Controller Parameters Example 1: Example 2: Performance Indices The performance indices used here are: 1. The Integral of Squared Error (ISE)
2. The Integral of Absolute Error (IAE)
3. The Integral of Time Multiply Squared Error (ITSE)
4. The Integral of Time multiply Absolute Error (ITAE)
Initial Controller Parameters Like mose optimization problems, the control performance optimization is non-convext, hence a trap of local minimum is inevitable. To counteract this, the initial controller parameters is set to be those determined by one of existing tuning rul es. In this way, the controller derived is at least better than that determined by the tuning method. The stability margin based Ziegler- Nichols is used for initial controller parameters and for performance comparison. Example 1: Consider a 4th-order system:
Compare closed-loop performance of PID controllers designed with different performance indices. G=zpk([],[-3 -2 -1 0],1); % The plant C1=optimPID(G,3,1); % PID-Control, ISE index C2=optimPID(G,3,2); % PID-Control, IAE index C3=optimPID(G,3,3); % PID-Control, ITSE index C4=optimPID(G,3,4); % PID-Control, ITAE index K=znpidtuning(G,3); % Ziegler-Nichols stability margin tuning t=0:0.1:30; y1=step(feedback(C1*G,1),t); %Closed-loop step response of C1 y2=step(feedback(C2*G,1),t); %Closed-loop step response of C2 y3=step(feedback(C3*G,1),t); %Closed-loop step response of C3 y4=step(feedback(C4*G,1),t); %Closed-loop step response of C4 %Closed-loop step response of K y=step(feedback(G*(K.kc*(1+tf(1,[K.ti 0])+tf([K.td 0],1))),1),t); plot(t,y1,t,y2,t,y3,t,y4,t,y,'--','Linewidth',2) legend('ISE','IAE','ITSE','ITAE','Z-N','Location','Best') grid
% The comparison shows that the ITSE index leads to the best PID % controller.
Example 2: A 4th-order system with a repeated pole.
Compare closed-loop performance of PI controllers. G=tf(1,[1 4 6 4 1]); % The plant C1=optimPID(G,2,1); % PID-Control, ISE index C2=optimPID(G,2,2); % PID-Control, IAE index C3=optimPID(G,2,3); % PID-Control, ITSE index C4=optimPID(G,2,4); % PID-Control, ITAE index K=znpidtuning(G,2); % Ziegler-Nichols stability margin tuning t=0:0.1:40; y1=step(feedback(C1*G,1),t); %Closed-loop step response of C1 y2=step(feedback(C2*G,1),t); %Closed-loop step response of C2 y3=step(feedback(C3*G,1),t); %Closed-loop step response of C3 y4=step(feedback(C4*G,1),t); %Closed-loop step response of C4 %Closed-loop step response of K y=step(feedback(G*(K.kc*(1+tf(1,[K.ti 0]))),1),t); plot(t,y1,t,y2,t,y3,t,y4,t,y,'--','Linewidth',2) legend('ISE','IAE','ITSE','ITAE','Z-N','Location','Best') grid