OPTIMAL TUNING OF PI CONTROLLERS FOR FIRST

ORDER PLUS DEAD TIME/LONG DEAD TIME MODELS

USING DIMENSIONAL ANALYSIS
Saeed Tavakoli, Peter Fleming

Department of Automatic Control and Systems Engineering

The University of Sheffield, Mappin Street, Sheffield, S1 3JD, UK
E-mail: [email protected]
Fax: +44-114-222-5138

Keywords: PI controller, FOPDT model, dimensional the 1940s, many of them reported in [1]. These methods are
analysis, optimisation, robustness. usually different in complexity, flexibility and in the amount
of process knowledge used. Nevertheless, there is no
Abstract generally accepted design method for this controller [2].
Therefore, the design of PI controllers still remains a
For first order plus dead time models, an optimal method for challenge before researchers and engineers.
tuning PI controllers is presented using dimensional analysis Traditionally, PI controllers have been tuned empirically, e.g.,
and numerical optimisation techniques. Considering a step by the first method of Ziegler and Nichols described in [18].
change in setpoint, optimal equations for determining PI This method, called the “continuous cycling method” has
parameters are obtained through minimising the integral of been widely known as a fairly accurate heuristic method to
absolute error (IAE). The optimisation process is constrained determine good settings of PI and PID controllers for a wide
to guarantee a minimum Gain margin (G.M.) of 2 and a range of common industrial processes [9]. It also has the
minimum Phase Margin (P.M.) of 60°. The proposed advantage of requiring very little information about the
formulas can also be used for first order systems with long process, however, it requires knowledge about the ultimate
dead time. Simulation results show that the proposed method data which are obtained by destabilising the system under
has a considerable superiority over conventional techniques. proportional feedback. Moreover, the method inherently leads
In addition, the closed loop system shows a robust to an oscillatory response in the face of a change in the
performance in the face of model parameters uncertainties. setpoint [7, 14].
A large number of industrial plants can approximately be
modelled by a first order plus dead time transfer function as
1 Introduction follows:
In the earlier paper [16], the use of dimensional analysis in
Ke -τ d s
tuning of PID controllers for first order plus dead time G (s) = . (1)
(FOPDT) models was proposed. As an extension of this Ts + 1
study, the current paper considers tuning of PI controllers for
first order systems with dead time/long dead time. In addition, In order to design PI controllers for this important category of
because gain and phase margins are often used as a measure industrial plants, various methods have been suggested during
of robustness [10], the optimal formulas are determined so the past sixty years. The second method of Ziegler and
that the predefined amounts for G.M. and P.M. are Nichols known as the “process reaction curve” method [19]
guaranteed. and that of Cohen and Coon [4] are the most prominent
Despite the continual advances in control theory, the PI methods mentioned in most control textbooks. Similar to the
controller is still the most commonly used controller in the Ziegler and Nichols methods, Cohen and Coon technique
process control industry [2]. This is mainly due to its sometimes brings about oscillatory responses, because it was
noticeable effectiveness and its simple structure which is designed to provide closed loop responses with a damping
conceptually easy to understand. According to the reports, ratio of 25% [16]. However, Ziegler and Nichols methods are
more han 90% of the industrial controllers are PID, mostly PI, still widely used, either in their original form or with some
controllers [5,11,12]. In [3] a typical paper mill was reported modifications [1].
with more than 2000 control loops while 97% of these loops
used PI controllers. As a result, any improvement in the PI 2 Proposed method
controller tuning methods is priceless because of its broad
range of applications. A number of analytical and numerical An efficient design method should cope with a wide range of
methods have been proposed for tuning this controller since systems. It should satisfy the design specifications and be
robust in the face of model uncertainties. The aim of this determining the second and third dimensionless numbers
paper is to propose a set of formulas for tuning PI controllers from the first one, as shown below:
for FOPDT models. Therefore, as shown in equation (2), the
PI parameters should be defined based on the model τd
parameters: KK c = g 1 ( ).
T
(5)
Ti τ
K c = f 1 ( K , τ d , T ). = g 2 ( d ).
(2) τd T
Ti = f 2 ( K , τ d , T ).
Obviously, the functions in equation (5) can be determined
The problem is that it is very difficult to determine these much more easily than those in equation (2).
functions, because each parameter of the controller is a Considering a step change in the setpoint, this paper aims to
function of three parameters of the model. Therefore, we determine g1 and g 2 so that the IAE is minimised. In order
propose to use dimensional analysis to reduce the number of to ensure that the resulting system has enough robustness, two
parameters involved. Dimensional analysis is a mathematical constraints are used to guarantee a minimum G.M. of 2 and a
tool often applied in physics and engineering to simplify a minimum P.M. of 60°. In addition, as the phase margin is
problem by reducing the number of variables to the smallest known to be related to the damping of the system, it also
number of essential ones [20]. In other words, dimensional serves as a measure of performance [10].
analysis is a process for eliminating extraneous information First, the best values of controller parameters are determined
from a relation between quantities [17]. τ
for any given d using genetic algorithms [6]. In order that
T
Definition 1:
the resulting formulas can also be applied to first order plus
A dimensionless number is a pure number without any
τ
physical unit. Such a number is typically defined as a product long dead time systems, d is changed from 0.1 to 10. Then
or ratio of quantities that have units, in such a way that all T
units can be cancelled. τ
the optimal values of KK c versus d are drawn. Finally,
T
Buckingham’s pi-theorem: g1 is determined using curve-fitting techniques. The situation
Any equation such as
Ti T
is same for g 2 if KK c is replaced with or i .
f ( x1 , x 2 ,..., x n ) = 0. (3) τd T
Equations (6,7) represent the proposed formulas for tuning PI
with nonzero x1 , x 2 ,..., x n , is equivalent to an equation of the controllers:
form
T
KK c = 0.4849 + 0.3047. (6)
g (π 1 , π 2 ,..., π k ) = 0. (4) τd

where π 1 , π 2 ,..., π k are independent dimensionless numbers. Ti τ

= 0.4262 d + 0.9581. (7)
Here k = n − m where m is the number of fundamental units T T
used. In other words, m is the minimum number of
x1 , x 2 ,..., x n , which include all the units in equation (3). 3 Simulation results
In equation (1), the unit of τ d and T is time. The unit of K is
In this section, the performance of the proposed method is
unit of output compared with that of other techniques. For simplicity, the
equal to which is different from a plant to
unit of input Cohen-Coon, first Ziegler-Nichols, second Ziegler-Nichols,
another one. K and τ d (or T) are fundamental units and refined Ziegler-Nichols, Lee-Edgar, Hagglund and proposed
include all the units in equation (1). As a result, there is only methods are abbreviated to C.C., Z.N.1, Z.N.2, R.Z.N., L.E.,
Hag. and Pro., respectively.
τ
one dimensionless number in the model, namely d . Other For the first example, the performance of the proposed
T method is compared with that of the C.C., Z.N.1, Z.N.2,
dimensionless numbers for the model and the controller are: R.Z.N. [9] and L.E. [13] techniques.

Ti T Example 1:
KK c and either of or i .
τd T
e -0.5 s
G1 ( s ) = . (8)
Based on Buckingham’s pi-theorem, the PI parameters are s +1
obtained from the parameters of the model through
The results of the comparison are shown in Table 1. The In order to use the C.C., Z.N.2 and proposed techniques to
worst response is given by the C.C. method. Since the obtain the PI parameters, G 2 ( s ) is approximately modelled
proposed method has the least percentage of overshoot and with a FOPDT model, using half rule [15].
the highest phase margin and acceptable values of settling
time and gain margin, it gives the best performance. Half rule:
Let the original model be:
C.C. Z.N.1 Z.N.2 R.Z.N. L.E. Pro.
Kc 1.88 1.8 1.71 0.73 1.45 1.27 m

Ti 0.83 1.5 1.45 0.69 1.16 1.17 ∏ (−α

j =1
j s + 1)

P.O. 55.3 21.28 18.36 13.03 15.68 7.54 Q( s ) = K n

e −θs . (12)
Ts
G.M.
6.9
1.58
5.24
1.91
3.54
2
4.58
3.73
3.68
2.27
3.81
2.58
∏ (β s + 1)
i =1
i

P.M. 29.08 51.61 53.6 53.19 54.3 59.95

where β 1 ≥ β 2 ... ≥ β n and α 1 , α 2 ,..., α m are positive values.
Table 1: Comparison of the performance of different methods Then the FOPDT is given by:
in controlling G1 ( s ) .
∧ Ke -τ d s
Figure 1 shows the closed loop step responses resulted from Q( s ) = . (13)
Ts + 1
the R.Z.N., L.E. and proposed methods.

β2 n m
β2
where τ d = θ +
2
+ ∑
i =3
βi + ∑α j =1
j and T = β 1 +
2
.

Using half rule, the approximate model for G 2 ( s ) is given

by:

∧ e -5s
G 2 (s) = . (14)
7s + 1

Table 2 shows the results of the comparison.

C.C. Z.N.1 Z.N.2 R.Z.N. Pro.

Kc 1.34 1.59 1.26 0.71 0.98
Ti 6.9 14.47 15 6.6 8.84
P.O. 46.08 20.44 3.17 11.19 10.93
Figure 1. Closed loop step response resulted from applying Ts 73.4 54.7 61.8 33.1 40.1
the R.Z.N., L.E. and proposed methods to G1 ( s ) . G.M. 1.72 1.89 2.4 3.18 2.65
P.M. 31.15 48.83 67.19 56.35 58.32
Example 2:
Table 2: Comparison of the performance of different methods
-s in controlling G 2 ( s ) .
(− s + 1)e
G 2 (s) = . (9)
(6s + 1)(2s + 1) 2
The response of the C.C. method is again the worst, while the
performance of the Z.N.2, R.Z.N. and proposed techniques
In order to determine the PI parameters suggested by the are the best. Figure 2 shows the closed loop step responses
Z.N.1 and R.Z.N. methods, first K u and ω are determined resulted from the Z.N.2, R.Z.N and proposed methods.
from the following equations: Although the response given by the Z.N.2 method shows the
least overshoot, it has the biggest undershoot and longest
settling time. The percentage of overshoot and phase margin
36ω 2 + 1(4ω 2 + 1)
Ku = . (10) of the proposed method are slightly better than those of the
ω 2 +1 R.Z.N. From settling time and gain margin points of view, the
response of the R.Z.N. is better than that of the proposed
tg −1 (ω ) + ω + tg −1 (6ω ) + 2tg −1 (2ω ) = π . (11) method, however, the proposed technique suggests a faster
response with a smaller rise time.
 parameters in G1 ( s ) and G3 ( s ) are deviated as much as 20%
of their nominal values. The worst case is related to an
increase of 20% in K and t d and a decrease of 20% in T [16].
Considering G1 ( s ) , the perturbed model is given by:

∧ 1.2 e -0.6s
G 1 (s) = . (16)
0.83s + 1

Table 4 shows the results of the comparison.

C.C. Z.N.1 Z.N.2 R.Z.N. L.E. Pro.

Kc 1.88 1.8 1.71 0.73 1.45 1.27
Ti 0.83 1.5 1.45 0.69 1.16 1.17
P.O. Unstable 65.79 60.55 27.07 51.83 37.68
Figure 2. Closed loop step response resulted from applying Ts closed 17.44 13.28 5.05 9.25 6.03
the Z.N.2, R.Z.N. and proposed methods to G 2 ( s ) . G.M. loop 1.19 1.24 2.27 1.41 1.6
system
P.M. 18.71 22.9 44.22 31.27 41.13
Example 3:
Table 4: Comparison of the performance of different methods
∧
e -10 s
G3 ( s ) = . (15) in controlling G 1 ( s ) .
s +1
Facing with model uncertainties, the closed loop system
This example indicates a system with long dead time. Control resulted from the C.C. controller parameters is unstable, while
of systems with long dead time is difficult [1]. For this the Z.N.1 and Z.N.2 methods resulted in highly oscillatory
example, the performance of the proposed method is responses. Among the remaining techniques the performance
compared with that of the C.C., Z.N.1, Z.N.2, R.Z.N. and of the proposed method is acceptable, whilst the best and
Hagglund [8] techniques to show the effectiveness of the worst responses are related to the R.Z.N. and L.E. techniques,
proposed method when facing with such systems. The results respectively.
of the comparison are shown in Table 3.
Considering G3 ( s ) , the perturbed model is given by:
C.C. Z.N.1 Z.N.2 R.Z.N. Hag Pro.
Kc 0.17 0.47 0.09 0.38 0.25 0.35 ∧ 1.2 e -12 s
G 3 (s) = . (17)
Ti 2.87 18.66 30 5.61 3.9 5.22 0.83s + 1
P.O. 3.38 0 0 1.99 1.34 1.23
The results of comparison are shown in Table 5.
Ts 45 182 1362 45.2 25.7 45.7
G.M. 2.83 2.18 11.5 2.13 2.58 2.22 C.C. Z.N.1 Z.N.2 R.Z.N. Hag Pro.
P.M. 61.42 100.01 93.27 66.07 62.83 65.2 Kc 0.17 0.47 0.09 0.38 0.25 0.35
Ti 2.87 18.66 30 5.61 3.9 5.22
Table 3: Comparison of the performance of different methods
to control G3 ( s ) . P.O. 26.08 0 0 38.6 27.29 34.75
Ts 71.7 153 1139 85 77.4 84.2
The responses resulted from Z.N.1 and Z.N.2 methods are G.M. 2 1.77 9.37 1.61 1.87 1.67
very slow, however, all of the remaining methods show good P.M. 47.68 97.42 93.54 49.88 48.34 49.42
responses. The response of the Hag. method has the least
settling time, while the least percentage of overshoot comes Table 5: Comparison of the performance of different methods
from the proposed method. From G.M. and P.M. points of ∧
view, the C.C. and R.Z.N. techniques are the best, in controlling G 3 ( s ) .
respectively.
While the C.C. method shows the highest degree of
4 Robustness studies robustness and gives the best response, Z.N.1 and Z.N.2
methods bring about sluggish responses. The proposed,
R.Z.N., and Hag. methods show relatively robust
In order to investigate the robustness of the mentioned
performances and give acceptable responses.
methods in the face of model uncertainties, the model
5 Conclusions [13] J. Lee, T. F. Edgar. “Improved PI Controller with
Delayed or Filtered Integral Mode”, AIChE Journal, 48
In this paper an optimal technique for tuning PI controller (12), pp. 2844-28450, (2002).
parameters for FOPDT models was proposed. Dimensional [14] G. P. Liu, S. Daley. “Optimal-Tuning PID Control for
analysis and numerical optimisation methods were used to Industrial Systems”, Control Engineering Practice, 9, pp.
simplify the procedure of obtaining optimal formulas for 1185-1194, (2001).
control parameters. Simulation studies for three common [15] S. Skogestad. “Simple Analytic Rules for Model
examples showed that the proposed method could deal with Reduction and PID Controller Tuning”, Journal of
Process Control, 13, pp. 291-309, (2003).
τ
the FOPDT models over a large range of d . In addition, for [16] S. Tavakoli, M. Tavakoli. “Optimal Tuning of PID
T Controllers for First Order plus Time Delay Models
systems of higher orders which are capable of being reduced Using Dimensional Analysis”, Proceeding of Forth
to FOPDT models, the performance of the method is quite International Conference on Control and Automation,
satisfactory. Comparing the proposed method with well- Canada, (2003).
known techniques, suggested that the proposed method was [17] E. S. Taylor. Dimensional Analysis for Engineers,
advantageous to most of them such as Ziegler-Nichols and Clarendon Press, Oxford, (1974).
Cohen-Coon methods. In addition, robustness studies proved [18] J. G. Ziegler, N. B. Nichols. “Optimum Settings for
that the PI controller given by the proposed formulas was Automatic Controllers”, Trans. ASME, 64, pp. 759-768,
satisfactorily robust against model uncertainties. (1942).
[19] J. G. Ziegler, N. B. Nichols. “Process Lags in Automatic
References Control Circuits”, Trans. ASME, 65, pp. 433-444,
(1943).
[1] K. J. Astrom and T. Hagglund. PID Controllers: Theory, [20] M. Zlokarnik. Dimensional Analysis and Scale-up in
Design and Tuning, Instrument Society of America, Chemical Engineering, Springer-Verlag, Berlin, (1991).
(1995).
[2] K. J. Astrom, H. Panagopoulos, T. Hagglund. “Design of
PI Controllers Based on Non-Convex Optimization”,
Automatica, 34 (5), pp. 585-601, (1998).
[3] W. L. Bialkowski. “Dreams versus Reality: A View from
Both Sides of the Gap”, Pulp and Paper Canada, 94 (11),
pp. 19-27, (1993).
[4] G. H. Cohen, G. A. Coon. “Theoretical Consideration of
Related Control”, Trans. ASME, 75, pp. 827-834, (1953).
[5] P. B. Deshpande. Multivariable Process Control, ISA,
RTP, North Carolina, (1989).
[6] P. J. Fleming, R. C. Purshouse. “Evolutionary Algorithms
in Control Systems Engineering: A Survey”, Control
Engineering Practice, 10, pp.1223-1241, (2002).
[7] B. Friedland. Advanced Control System Design, Prentice
Hall, (1996).
[8] T. Hagglund. “An Industrial Dead-Time Compensating PI
Controller”, Control Engineering Practive, 4 (6), pp. 749-
756, (1996).
[9] C. C. Hang, K. J. Astrom, W. K. Ho. “Refinements of the
Ziegler-Nichols Tuning Formula”, IEE Proceeding-D,
138 (2), pp. 111-117, (1991).
[10] W. K. Ho, O. P. Gan, E. B. Tay, E. L. Ang.
“Performance and Gain and Phase Margines of Well-
Known PID Tuning Formulas”, IEEE Transactions on
Control Systems Technology, 4 (4), pp. 473-477, (1996).
[11] W. K. Ho, T. H. Lee, W. Xu, J. R. Zhou, E. B. Tay. “The
Direct Nyquist Array Design of PID Controllers”, IEEE
Transactions on Industrial Electronics, 47 (1), pp. 175-
185, (2000).
[12] H. N. Koivo, J. T. Tanttu. “Tuning of PID Controllers:
Survey of SISO and MIMO Techniques”, Proceedings of
Intelligent Tuning and Adaptive Control, Singapore,
(1991).