You are on page 1of 10

Demystifying PID Controller Tuning

for Desired Performance and Robust Stability


Jeffrey E. Arbogast
American Air Liquide, Delaware Research and Technology Center, Newark, DE USA
jeffrey.arbogast@airliquide.com

KEYWORDS
Robustness Plot, Robust Stability, PID, Controller Tuning, IMC, SIMC

ABSTRACT
While the Proportional-Integral-Derivative (PID) control algorithm is fundamentally very simple,
tuning PID controllers can be an intimidating and confusing challenge. Manufacturers use different
forms of the PID algorithm with different terminology. These different forms are fundamentally the
same. However, the intuitive feel for tuning one form of the PID algorithm may not be applicable for
other forms.

While there are many methods for tuning PID controllers, this work focuses upon the Internal Model
Control (IMC) tuning correlation because it allows users to specify both the desired performance and
the necessary level of robust stability by adjusting a single parameter, the closed-loop time constant.
Specify a larger closed-loop time constant for a more robustly stable controller that produces a slower,
more conservative response with a longer rise time. Specify a smaller closed-loop time constant for a
less robustly stable controller that produces a faster, more aggressive response with a shorter rise time.
Thus, performance and robust stability are related.

Performance characteristics (e.g. rise time, settling time, peak-overshoot ratio) are relatively well
understood. Robust stability is less well understood. The PID controller is tuned based upon a model
of the plant. The inaccuracy of this model in describing the real plant is referred to as plant-model
mismatch. This work uses both visual robustness plots and a novel quantitative robust stability factor
metric (RSF) to provide a clear, intuitive indication of robust stability. Both RSF and the related
robustness plot indicate the amount of plant-model mismatch that may be tolerated without the system
becoming unstable in closed loop. This understanding of robust stability is reinforced through visual
examples of how closed-loop performance changes with various levels of plant-model mismatch.

One critique of the IMC tuning correlation is its sluggish performance when the process time constant
is much greater than the process dead time. The Skogestad IMC (SIMC) tuning correlation addresses
this concern and gives the user greater ability to specify acceptable levels of both performance and
robust stability. This work extends the application of the RSF metric to the SIMC tuning correlation
and for non-self-regulating (integrating) processes.

Industrial examples and simulation studies are used to provide visual examples of the effect of various
values of the closed-loop time constant on performance. In some applications, a quick return to set

Distributed with permission of authors by ISA 2010.


Presented at ISA Applications in Automation Conference 2010, Wilmington, DE; http://www.isa.org
point is desirable and overshoot is acceptable. In other applications, overshoot is not acceptable and a
slower response is preferred. These examples demonstrate how the closed-loop time constant
facilitates the application of the IMC and SIMC tuning correlations in both of these scenarios.

INTRODUCTION
This paper provides three examples of industrial application of the author’s academic research work.
While this paper focuses upon PI control, [1] provides information about the differences and
similarities between the various PID with Filter controllers offered by manufacturers. [2] details a
method (applied in Example 3) for graphically fitting a model to data from an integrating (or non-self
regulating) process. [7, 8] detail the development of the robust stability factor (RSF) metric as applied
in Examples 1-3.

In this paper, the PI controller is assumed to have the following form:


⎡ 1 ⎤
GC ( s) = KC ⎢1 + ⎥. (1)
⎣ τIs ⎦
In Eq. 1, KC is the controller gain and τI is the reset time.

This paper presents methods for tuning control loops based upon models fit to dynamic data collected
from the process. For self regulating processes, the following First Order plus Dead Time (FOPDT)
model is used:
K P e −θ P s
GP ( s ) = . (2)
τ Ps +1
In Eq. 2, KP is the process gain, τP is the process time constant, and θP is the process dead time.

The IMC-PI tuning correlations for the FOPDT model are as follows [3-5]:
τP
KC = ; τI =τ P. (3)
K P (τ C + θ P )
The closed-loop time constant, τC, is the single adjustable IMC tuning parameter that allows the
desired levels of performance and robust stability to be specified. Increase τC for a more robustly
stable controller or decrease τC for a less robustly stable controller.

Table I. Effect of τC Selection upon Control Loop Performance


*applies generally for IMC-tuned PI control of self regulating processes where τP ≤ 8 θP

Performance Settling % Decay


τC RSF2D
Benchmark Time Overshoot Ratio
Very Aggressive 0.22 θP 1.4 10.4 θP 34% 0.12
Aggressive 0.72 θP 1.6 8.8 θP 11% 0.01
Moderate Aggressive 2.40 θP 2.3 11.3 θP none n/a
Moderate 8.00 θP 3.8 28 θP none n/a
Moderate Conservative 26.6 θP 6.6 85 θP none n/a
Conservative 89.0 θP 11.9 270 θP none n/a
Very Conservative 297 θP 23.5 890 θP none n/a

Distributed with permission of authors by ISA 2010.


Presented at ISA Applications in Automation Conference 2010, Wilmington, DE; http://www.isa.org
Table I [6] outlines the relationship between the τC value and the values of four statistics: robust
stability factor (RSF2D) [6-8], settling time, % overshoot, and decay ratio. The decay ratio indicates the
amount of oscillation – in terms of the ratio of the height of the second peak over the new SP to the
height of the first peak over the new SP. Examples 1 and 2 demonstrate the application of the IMC-PI
tuning correlation described by Eq. 3.

One critique of the IMC tuning correlation is that it produces sluggish performance when the process
time constant is much greater than the process dead time. The following Skogestad IMC (SIMC)
tuning correlation [9] addresses this concern by adjusting τI according to the selection of τC:
τP
KC = ; τ I = min ⎡⎣τ P ,4 (τ C + θ P ) ⎤⎦ . (4)
K P (τ C + θ P )
Similar to IMC tuning, the closed-loop time constant, τC, is the single adjustable SIMC tuning
parameter that allows the desired levels of performance and robust stability to be specified. Increase τC
for a more robustly stable controller or decrease τC for a less robustly stable controller. Example 2
demonstrates the application of SIMC tuning.

For non-self-regulating (or integrating) processes (like the one presented in Example 3), the First Order
plus Dead Time Integrating (FOPDT Integrating) model is used:
K * e −θ P s
GP ( s ) = P . (5)
s
In Eq. 5, KP* is the integrator gain and θP is the process dead time.

The IMC-PI tuning correlations for the FOPDT Integrating model are as follows [3-5]:
2τ + θ P
KC = * C ; τ I = 2τ C + θ P . (6)
K P (τ C + θ P )

EXAMPLE 1 – TUNING FOR NONLINEAR PROCESS BEHAVIOR

While processes are nonlinear (to different levels of significance), PID controllers are linear.
Therefore, the performance of a PID controller will vary greatly over the operating range of a highly
nonlinear process [7, 8]. The following discussion of the tuning of an industrial pressure control loop
illustrates the importance of considering this nonlinearity in tuning controllers.

Before departing the plant for the evening, we conducted bump tests in manual mode – the controller
output (OUT) was stepped several times between 34.5% and 35.5% with the pressure – the process
variable (PV) – around 205 psig. Tuning software [10] produced a high-quality model fit to the data
(as shown in Fig. 1) and recommended new tuning parameters (as shown in Fig. 2) based upon the
IMC tuning correlation. Using the simulation shown by the green plot in Fig. 2, the proposed tuning is
designed to respond as fast as possible without overshooting the set point or oscillating.

Distributed with permission of authors by ISA 2010.


Presented at ISA Applications in Automation Conference 2010, Wilmington, DE; http://www.isa.org
Figure 1. Model Fit to data collected in manual Figure 2. Analysis of IMC tuning
mode (around OUT = 35%, PV = 205 psig). recommended based upon model fit in Fig. 1.
KP = 25.9 psig/%OUT = 8.6 %PV/%OUT, τP = 0.13 min, θP = 0.058 min Recommended Tuning (Green): KC = 0.10 %OUT/%PV, τI = 0.13 min
Top Plot: PV data (white) and PV model fit (yellow). Original Tuning (Purple): KC = 0.2 %OUT/%PV, τI = 0.8 min
Bottom Plot: OUT data

On the next morning, the tuning was implemented and tested with another bump test (this time in
automatic mode) around 18.25 %OUT and 205 psig. While the simulation (in Fig. 2) predicted neither
overshoot nor oscillation, the actual performance exhibited both overshoot and oscillation (as shown in
Fig. 3). This was due to process conditions changing overnight. The PV remained around 205 psig.
However, the controller output changed significantly, from 35 %OUT in the evening and 18.25 %OUT
in the morning. Clearly, the model fit from Fig. 1 no longer applies at the new operating conditions.

Figure 3. Model Fit to SP step down Figure 4. Analysis of IMC tuning


(around OUT = 18.25%, PV = 205 psig). recommended based upon model fit in Fig. 3.
KP = 39.8 psig/%OUT = 13.3 %PV/%OUT, τP = 0.45 min, θP = 0.033 min Recommended Tuning (Green): KC = 0.32 %OUT/%PV, τI = 0.45 min
Top Plot: PV data (white), SP data (blue) and PV model fit (yellow). Existing Tuning (Purple): KC = 0.1 %OUT/%PV, τI = 0.13 min
Bottom Plot: OUT data

In Fig. 3 and 5, the controller oscillates significantly more in response to the SP step up (200 to 210
psig) than it does in response to the SP step down (210 to 200 psig). This indicates that the control
loop has significantly nonlinear behavior. To better understand this behavior, we fit one linear model
to the SP step down from 210 psig to 200 psig and another linear model to the SP step back up to 210
psig.

Distributed with permission of authors by ISA 2010.


Presented at ISA Applications in Automation Conference 2010, Wilmington, DE; http://www.isa.org
Figure 5. Model Fit to SP step up Figure 6. Analysis of IMC tuning
(around OUT = 18.25%, PV = 205 psig). recommended based upon model fit in Fig. 5.
KP = 30.6 psig/%OUT = 10.2 %PV/%OUT, τP = 0.28 min, θP = 0.16 min Recommended Tuning (Green): KC = 0.074 %OUT/%PV, τI = 0.28 min
Top Plot: PV data (white), SP data (blue) and PV model fit (yellow). Existing Tuning (Purple): KC = 0.1 %OUT/%PV, τI = 0.13 min
Bottom Plot: OUT data

The robustness plot [11-17] (the right-hand plot in Fig. 6) graphically indicates how much plant-model
mismatch may be tolerated before the control loop becomes unstable. It plots plant-model mismatch in
the process gain (KP,Actual/KP,Model) versus plant-model mismatch in the process dead time
(θP,Actual/θP,Model). As with the RSF2D, it is assumed there is no plant-model mismatch in the process
time constant (τP,Actual = τP,Model).

The RSF2D indicates the factor by which the process gain, KP, and process dead time, θP, may both
change from their model values (fit in Fig. 5 in this case) – assuming that the process time constant, τP,
remains at its model value. A RSF2D of 1.0 would indicate a critically stable process with oscillations
that neither grow nor decay – the brink of instability. The level of robust stability increases with
RSF2D.

The existing tuning has a RSF2D of 1.2. Therefore, the loop would become unstable if KP and θP both
increase by 20% to 36.7 psig/%OUT and 0.19 min respectively while τP remains at its model value
(0.281 min).

Therefore, the existing tuning does not provide a high degree of robustness – in terms of either stability
or performance. Recall from Fig. 2 that this tuning was designed at one operating level to respond as
fast as possible without overshooting the SP or oscillating. At another operating level, Fig. 3 shows
that this tuning produced a response overshooting the SP with significant oscillation. In the three
model fits shown by Fig. 1, 3, and 5, the KP varies from 25.9 to 39.8 psig/%OUT, the τP varies from
0.13 to 0.45 min, and the θP varies from 0.033 to 0.16 min. This variation needs to be accommodated
with robust controller tuning.

In contrast to the existing tuning, the proposed tuning is significantly more robust as indicated by a
RSF2D of 1.9. Therefore, the loop would only become unstable if two of the model parameters (KP and
θP) both increase by 90% to 58 psig/%OUT and 0.30 min respectively while τP remains at its model
value (0.281 min). A RSF3D value of 1.7 (not shown in Fig. 6) indicates that loop will remain stable if
all three model parameters (KP, τP, and θP) each change by a factor of 1.7 or less (remaining between
60% and 170% of their model values). The difference between the RSF2D (1.9) and RSF3D (1.72)
values indicates that plant-model mismatch in τP has an impact on the robust stability.

Distributed with permission of authors by ISA 2010.


Presented at ISA Applications in Automation Conference 2010, Wilmington, DE; http://www.isa.org
EXAMPLE 2 – TIME CONSTANT DOMINANT PROCESS
The tuning software [10] fit a good model to the data collected during a SP step on a pressure control
loop. This model is the basis for the following analysis of the control loop tuning. Note that the time
constant (τP = 4.03 min = 242 sec) is about 32 times larger than the dead time (θP = 0.126 min = 7.6
sec). Therefore, this definitely qualifies as a time constant dominant process.

Figure 7. Model Fit to SP step data. Figure 8. Analysis of IMC tuning recommended
KP = 23.3 psig/%OUT = 23.3 %PV/%OUT, τP = 4.03 min, θP = 0.126 min based upon model fit in Fig. 1.
Top Plot: PV data (white), SP data (blue) and PV model fit (yellow). Bottom
Recommended Tuning (Green): KC = 0.36 %OUT/%PV, τI = 242 sec
Plot: OUT data
Original Tuning (Purple): KC = 0.01 %OUT/%PV, τI = 3 min

The predicted performance of the original tuning (shown by the purple plot in Fig. 8) is very similar to
the actual performance shown in Fig. 7. This confirms the validity of the model for use in controller
tuning. Application of the IMC tuning correlation with an aggressive choice for the closed-loop time
constant (τC = 0.36 min) produces the following controller tuning parameters: KC = 0.355, τI = 242 sec.
This controller tuning will provide faster SP tracking performance without the oscillatory behavior
caused by the original tuning (shown by the green plot in Fig. 8).

However, the disturbance rejection performance of the IMC tuning parameters must also be
considered. The IMC tuning correlation sets τI equal to τP, regardless of the τC value (the user’s
selection of aggressive, moderate, or conservative control). The large τP-to-θP ratio of 32 indicates that
the τI of 242 sec is much larger than necessary. As shown in column (e) of Table II, this produces slow
disturbance rejection performance – taking 15 minutes to gradually return to the SP value without
overshoot or oscillation.

While the IMC tuning correlation recommended an increase in the τI from 3 sec to 242 sec (80 times
higher), the τI was instead increased to 30 sec (10 times higher). The KC was only increased from 0.01
to 0.1 (10 times higher). The predicted performance of this tuning is shown in column (b) of Table II.
In response to the disturbance, the pressure returns to its SP value in only 5 minutes but overshoots and
oscillates around the SP before settling after about 12 minutes.

Based upon a tuning software simulation, we recommended changing the controller tuning to (KC =
0.2, τI = 60 sec). The proposed tuning improved both SP tracking and disturbance rejection
performance (compare (b) and (c) in Table II). In response to the disturbance, the pressure continues
to return to its SP value in only 5 minutes. However, the maximum deviation from the SP is reduced
by a quarter and there is less oscillation around the SP.

Distributed with permission of authors by ISA 2010.


Presented at ISA Applications in Automation Conference 2010, Wilmington, DE; http://www.isa.org
Table II. SP Tracking and Disturbance Rejection Performance for Various Tuning
(a) (b) (c) (d) (e)
Original Intermediate Final SIMC IMC
Tuning Tuning Tuning Tuning Tuning
KC 0.01 0.1 0.2 0.355 0.355
τI (sec) 3 30 60 118 242

SP
Tracking

Disturb.
Rejection

RSF2D 1.51 3.21 3.00 2.44 2.49


RSF3D 1.51 2.50 2.16 1.83 1.85

While the custom simulation capability is a useful expert tool, an applicable tuning correlation would
be more useful (especially for those with less expertise). The SIMC tuning correlations are
fundamentally based upon the IMC tuning correlation. However, the SIMC tuning correlations allow
the τI to decrease with the closed-loop time constant for time constant dominant processes (like this).

For an aggressive τC of 0.36 min, the SIMC tuning correlation recommends a τI of only 118 sec (as
opposed to the 242 sec recommended by the IMC correlation). As shown in columns (d) and (e) in
Table II, the SIMC and IMC tuning perform similarly in SP tracking. The SIMC tuning produces a
slight (10%) overshoot of the SP and a correspondingly longer settling time.

The significant differences between the RSF2D and RSF3D values listed in Table II confirm that plant-
model mismatch in τP should be considered in evaluating robustness for time constant dominant
processes like this [7, 8]. For a τC of 0.36 min, the IMC and SIMC tuning have similar robust stability.
The RSF3D of 1.83 indicates that the loop will remain stable as long as KP and θP both increase by less
than a factor of 1.83 (or 83%) and the τP decreases by less than a factor of 1.83 (or 45%). This gives a
different impression of robust stability from the RSF2D value of 2.44 that does not account for variation
in τP, suggesting a significantly larger factor by which KP and θP may both increase.

EXAMPLE 3 – BLOWDOWN DRUM LEVEL CONTROL (INTEGRATING)

Plant operators noticed recurring problems with the performance of a blowdown drum level control
loop. Approximately once every 1.5-2 hours, a disturbance would cause the pressure and level to
spike. To reject this disturbance, the level controller would over-react producing a slowly-dampening
oscillation around the set point before settling in about 75 minutes (as shown in Fig. 11).

Distributed with permission of authors by ISA 2010.


Presented at ISA Applications in Automation Conference 2010, Wilmington, DE; http://www.isa.org
The first step in tuning the control loop was to collect dynamic data from the process – in this case the
response to a step in the controller output (in manual mode). As shown by Fig. 9, the tuning software
[10] fit a FOPDT Integrating model to this data.

Figure 9. FOPDT Integrating Model Fit. Figure 10. Analysis of IMC tuning
KP* = -0.326 %level/%OUT-min, θP = 0.215 min
P

recommended based upon model fit in Fig. 9.


Top Plot: PV data (white) and PV model fit (yellow).
Recommended Tuning (Green): KC = 1.84 %OUT/%PV, τI = 373 min
Bottom Plot: OUT data
Original Tuning (Purple): KC = 0.5 %OUT/%PV, τI = 10 min

Based upon this model fit, the tuning software recommended IMC tuning parameters as listed in the
caption for Fig. 10. The conservative τC of 3.0 min was selected to produce a control loop tuning
responding as fast as the original control loop tuning, but with reduced overshoot and oscillation.
While a smaller, more aggressive τC of 0.68 min produced a faster response, the larger KC caused
occasional noise in the level measurement to be amplified in the controller output, increasing valve
wear.

As shown by the lack of a purple region in Fig. 10, the tuning software finds that the original tuning is
unstable for this FOPDT Integrating model fit (and therefore the RSF2D is undefined). The original
tuning is, in fact, marginally stable as shown by the response in Fig. 12 with a decay ratio of 75% (a
decay ratio over 100% would indicate instability). Therefore, the original tuning is clearly not very
robust. The new tuning is much more robust, as indicated by its RSF2D of 3.4.

As shown in Fig. 11, application of the recommended IMC tuning improved the disturbance rejection
performance considerably. Oscillations around the set point were eliminated and the time required to
settle back to the original set point was approximately reduced from 70 to 20 minutes. Compare the
initial peaks and note that there is no sacrifice in the initial disturbance rejection with the new tuning.
Note that the new tuning significantly reduces the movement in the controller output, reducing valve
wear.

Distributed with permission of authors by ISA 2010.


Presented at ISA Applications in Automation Conference 2010, Wilmington, DE; http://www.isa.org
Figure 11. Disturbance Rejection Performance Figure 12. Set Point Tracking Performance

As shown in Fig. 12, application of the recommended IMC tuning also improved the set point tracking
performance considerably. Tuning reduced the percent overshoot from 97% to 12%, eliminated
oscillations around the set point (decay ratio reduced from 75% to not applicable (N/A)), and reduced
the settling time by a factor of 5 from 75 to 15 minutes.

CONCLUSIONS
This paper demonstrates the application of the IMC-PI tuning correlation for two self regulating
pressure control loops (Examples 1 and 2) and a non-self-regulating (or integrating) blowdown drum
level control loop (Example 3). In Example 2, the Skogestad IMC (SIMC) tuning correlation is
applied to improve disturbance rejection performance for a time constant dominant process while
maintaining robust stability. The adjustable closed-loop time constant (τC) parameter used in the IMC
and SIMC tuning correlations is a useful tool for specifying desired levels of performance and robust
stability.

In Example 1, the control loop exhibits significantly nonlinear behavior – described by three very
different linear models shown in Fig. 1, 3, and 5. The robust stability factor metrics (RSF2D and
RSF3D) provide a methodology for ensuring robustness (in terms of stability and performance) despite
this nonlinearity. In this example, each of the three model parameters (KP, τP, and θP) varies
significantly. This causes the choice of the model to use for tuning to be a very challenging decision.

In Example 2, τP,Model is 32 times as large as θP,Model – qualifying the pressure control loop as time
constant dominant. Therefore, plant-model mismatch in τP has a more significant effect in Example 2
than if the θP,Model–to-τP,Model ratio was greater (as is the case in Example 1). The significance of this
effect is evident from the difference between the RSF2D and RSF3D values listed in Table II. Despite
assertions to the contrary [14, 17], this confirms [7, 8] that plant-model mismatch can have a
significant effect upon robust stability.

Distributed with permission of authors by ISA 2010.


Presented at ISA Applications in Automation Conference 2010, Wilmington, DE; http://www.isa.org
In Example 3, tuning improved the disturbance rejection performance considerably. Oscillations
around the set point were eliminated and the time required to settle back to the original set point was
approximately reduced from 70 to 20 minutes.

REFERENCES
[1] J.E. Arbogast and D.J. Cooper, “Extension of IMC Tuning Correlations for Non-Self Regulating (Integrating)
Processes.” ISA Transactions 46 (2007) 303-311.
[2] J.E. Arbogast, R.C. Rice and D.J. Cooper, “Graphical Technique for Modeling Integrating (Non-Self Regulating)
Processes without Steady-State Process Data.” Chem. Eng. Comm. 194 (2007) 1566-1578.
[3] D.E. Rivera, M. Morari, and S. Skogestad, “Internal Model Control. 4. PID Controller Design,” Industrial &
Engineering Chemistry Process Design & Development, vol. 25, pp. 252-265, 1986.
[4] I.-L. Chien, P.S. Fruehauf, “Consider IMC Tuning to Improve Controller Performance,” Chemical Engineering
Progress, vol. 86, pp. 33-41, 1990.
[5] C.A. Smith and A.B. Corripio, Principles and Practice of Automatic Process Control, 3rd ed. New York: Wiley,
2006.
[6] J.E. Arbogast, “Industrially Relevant Modeling, Analysis and Tuning for Process Control.” Storrs, CT: University
of Connecticut, Ph.D. Dissertation (2007).
[7] J.E. Arbogast, B.M. Beauregard and D.J. Cooper, “Controller Tuning for Performance and Robust Stability.”
2007 ISA Expo: Houston, TX.
[8] J.E. Arbogast, B.M. Beauregard and D.J. Cooper, “Intuitive Robust Stability Metric for PID Control of Self-
Regulating Processes.” ISA Transactions 47 (2008) 420-428.
[9] S. Skogestad, “Simple analytic rules for model reduction and PID controller tuning.” Journal of Process Control
13 (2003) 291-309.
[10] Loop-Pro Developer 5.1 software. Tolland, CT: Control Station, Inc., 2009.
[11] J. P. Gerry and P. D. Hansen, "Choosing the Right Controller," Chemical Engineering (New York), vol. 94, pp. 65-
68, 1987.
[12] J. P. Gerry, "Find out how good that PID tuning really is," Control engineering, vol. 35, pp. 69-71, 1988.
[13] P. S. Fruehauf, I.-L. Chien, and M. D. Lauritsen, "Simplified IMC-PID tuning rules," ISA Transactions, vol. 33,
pp. 43-59, 1994.
[14] J. Syder, T. Heeg, and A. O'Dwyer, "Dead-time compensators: performance and robustness issues," presented at
Process Control and Instrumentation 2000, Glasgow, Scotland, 2000.
[15] T. Heeg and A. O'Dwyer, "Compensation of Processes with Time Delay," presented at International Postgraduate
Student Research Conference, Dublin Institute of Technology, Dublin, Ireland, 1998.
[16] F. G. Shinskey, "Putting Controllers to the Test," Chemical Engineering (New York), vol. 97, pp. 96-106, 1990.
[17] F. G. Shinskey, Feedback Controllers for the Process Industries. New York: McGraw-Hill, 1994.

Distributed with permission of authors by ISA 2010.


Presented at ISA Applications in Automation Conference 2010, Wilmington, DE; http://www.isa.org

You might also like