Measurement and Control

Control and
Optimize
Nonlinear Systems
Turbocharging a constrained model predictive
controller (CMPC) with fuzzy logic allows
CMPCs to be applied to nonlinear systems.
Andrew C. Burden,
Roberto Z. Tantalean and
Pradeep B. Deshpande,
University of Louisville

T he developments in the field of computing

and communications in recent decades have
made it possible to devise complex control
strategies that would have been impossible to
implement with analog hardware. Much of the
progress, however, has been limited to linear systems.
corporates traditional advanced controls (feedforward
control, dead-time compensation, interaction compen-
sation, constraint handling) and constrained optimiza-
tion into a single software package. With CMPC, both
square (the number of manipulated variables (MVs) is
equal to the number of controlled variables (CVs)) and
Real-world multivariable systems often exhibit signifi- non-square (the number of MVs is not equal to the
cant nonlinearities and their control and optimization number of CVs) systems can be accommodated. When
pose challenges, since general stability criteria for the number of CVs exceeds the number of MVs,
nonlinear systems are not yet available. Self-tuning CMPC permits the regulation of the CVs within user-
and adaptive control concepts based on linear systems specified bounds. When the number of MVs exceeds
theory are often used to account for nonlinearities the number of CVs, the MVs may be allocated on the
when dynamic process models are available. basis of a suitable optimization objective. In such situ-
In this context, the developments in the field of ations, it becomes possible to maximize throughput,
fuzzy logic appear to show promise. The central idea minimize energy consumption, improve quality con-
is to use fuzzy logic to account for system nonlineari- trol, and improve the yield of more valuable products
ties and a linear constrained model predictive con- when more than one product is manufactured.
troller (CMPC) to achieve constrained optimization. In Dynamic process models are the backbone of every
this article we show how fuzzy logic concepts can be CMPC product available in the market. Linear dynam-
combined with CMPC for a class of nonlinear sys- ic models are used most often. While commercial
tems. An experimental application is presented to rein- packages vary in how computations are carried out,
force the concepts. the description of a generic CMPC algorithm can be
given in terms of the following steps:
CMPC basics Step 1. Measure the CVs and sample them at the
CMPC is one of the most significant control tech- current sampling instant. The sampling frequency is
nologies to arrive in the marketplace since PID-type selected so that the slowest dynamics in the multivari-
control was introduced over five decades ago (1–4). able system are accurately represented.
CMPC is a digital computer-control algorithm that in- Step 2. Using the same inputs (MVs, measured dis-

CEP February 2003 www.cepmagazine.org 63

Measurement and Control

■ Figure 1. Membership Function, µ ■ Figure 2.

Membership Function, µ
A fuzzy transi- 1.0 1.0 Plot of the
tion in which the membership
state of the pro- µA functions.
cess contains
the properties of
both states A µB
and B in a cer-
tain proportion
0.0 0.0
A B 50 53 60

Process State Process State

turbances) that produced the outputs in step 1, predict the of the constraints on the MVs and the CVs. The parameter
CVs for P sampling intervals into the future. The parame- M is the control horizon.
ter P is the prediction horizon. Step 5. Apply the first of these M moves in each of the
Step 3. Correct the vector of predicted CVs to account MVs to the process and repeat steps 1–4 at the next sam-
for the presence of modeling errors and unmeasured dis- pling interval.
turbances. In this context, the differences between the The parameters N (longest open-loop settling time), M
measured CVs and the predicted CVs serve as indications and P have a bearing on controller responsiveness and ro-
of modeling errors and unmeasured disturbances. bustness (the ability to maintain stability in the presence
Step 4. Solve a constrained optimization problem to of a plant/model mismatch). With proper choices of these
compute a set of M controller movements, such that a parameters, perfect control (minimum variable control)
user-selected optimization index is satisfied, subject to all can be specified. However, in this instance, excessive
movements of the MVs result and the system can become
Nomenclature unstable in the presence of modeling errors. If P is set
equal to N plus M, the computations are simplified and a
a0 = linear membership function constant for process state A high-performance controller with desirable robustness
b0 = linear membership function slope for process state A
properties results.
c0 = linear membership function constant for process state B
CV = controlled variable (process output) CMPC contains a number of parameters for specifying
d0 = linear membership function slope for process state B operational objectives. These include:
E = error (R–CV) • The upper and lower limits of the CVs specify the tar-
Kc = proportional controller gain gets. A unique setpoint is specified by setting the upper
M = control horizon limit equal to the lower limit. Different values specify the
MV = manipulated variable (process input) bounds within which the outputs are to be contained. The
N = longest open-loop settling time controller first tries to regulate the outputs within their re-
NL = negative large, fuzzy linguistic variable
spective bounds. If it can do so, then it focuses on achieving
NM = negative medium, fuzzy linguistic variable
NS = negative small, fuzzy linguistic variable the economic objectives specified with the cost coefficients.
P = prediction horizon • Weights associated with the CVs are used to prioritize
PL = positive large, fuzzy linguistic variable the relative importance of the outputs. It is possible to spec-
PM = positive medium, fuzzy linguistic variable ify tighter control of some outputs relative to the others.
PS = positive small, fuzzy linguistic variable • The upper and lower limits on the MVs specify their
R = PID controller setpoint bounds, which the CMPC will not violate.
Rule = fuzzy rule • The cost coefficients associated with the MVs allow
T = sampling period
for their allocation on the basis of economic criteria speci-
Z = zero, fuzzy linguistic variable
∆E = change in error fied in the objective function. Move suppression coeffi-
∆M = manipulated variable movement cients allow for smooth MV movements.

Greek symbols Fuzzy logic concepts

µ = membership function When the available process information is imprecise,
τI = i ntegral time fuzzy logic concepts may prove useful (5–8). Fuzzy
logic essentially involves interpolation between two or
Subscripts
more defined states. When the process drifts into an un-
i = defined process state
i,j = row i and column j certain region, fuzzy logic concepts may be used to esti-
n = sampling interval mate the state of the process based upon the location rel-
ative to the surrounding defined process states. As an ex-

64 www.cepmagazine.org February 2003 CEP

ample, consider the transition of a process from a de-
E
fined state A to a defined state B. As the transition pro-
ceeds, the state of the process becomes uncertain (fuzzy) NL NS Z PS PL
and contains the properties of both states A and B in a
certain proportion (Figure 1). PL Z PS PL PL PL
The extent to which the uncertain state contains the PS NS Z PS PL PL
properties of states A and B depends upon membership
functions, µι. The membership functions measure the frac- ∆E Z NL NS Z PS PL
tion of properties belonging to particular defined states. NS NL NL NS Z PS
Thus, the membership functions must sum to one.
NL NL NL NL NS Z
µA + µB = 1 (1)
■ Figure 3a. Organizing the fuzzy rule base.
The choice of the membership function is left to the de-
signer, although linear membership functions are often 1.0
used. The property of interest at an uncertain process state µ1
C is thus given by
µ2
0.0
Property at state C = (2) –5 –1 0 1 5
(Property at state A × µ A ) + (Property at state B × µ B )
0.5 Z PS PL PL PL
µA + µB
0.1 NS Z PS PL PL

∆E = -0.35
Let us take a numerical example to illustrate the con- 0 NL NS Z PS PL
cepts and to show how the membership functions are deter-
mined. Assume at process state A, the CV has a value of 50 –0.1 NL NL NS Z PS
and its value at process state B is 60. The property of inter- –0.5 NL NL NL NS Z
est is the proportional controller gain, Kc. Process testing
1.0
µ3
µ4
0.0

suggests a proportional controller gain of 2.0 at A and 3.0 at E=2

B. The proportional gain at a process state C, where the CV
= 53, is to be estimated using a linear membership function. ■ Figure 3b. An illustrative example of the fuzzy rule base.
Figure 2 plots the membership functions. At process
state A, µA = 1.0 and µB = 0.0 and at B, µB = 1.0 and µA =
0.0. Process state C belongs partially to A and partially to NL NS Z PS PL
B. Assuming a linear form for the membership functions: –10 –1 0 1 10

µA = a0 + b0(CV) (3a)
■ Figure 3c. Fuzzy rule base definitions and controller actions.
Since µA = 1.0 when the CV = 50.0, and µA = 0.0 when
the CV = 60.0, the constants in Eq. 3a are a0 = 6.0 and ( K c ( A ) × µ A ) + ( K c ( B) × µ B )
b0 = –0.1. Therefore: Kc (C ) = = 2.3 (5)
µA + µB
µA = 6.0 – 0.1 (CV) (3b)
Fuzzy-PID
Similarly, PID-type control is commonly used — meeting over
90% of the industrial control requirements — so advanced
µB = c0 + d0 (CV) (4a) control and optimization strategies that work in conjunc-
tion with this type of control are likely to be accepted by
Here, µB = 1.0 when the CV = 60.0, and µB = 0.0 when operators. This article will focus on the popular propor-
the CV = 50.0, so c0 = –5.0 and d0 = 0.1. Therefore: tional plus integral (PI) control law, which can be readily
extended to include the derivative mode. Essentially,
µB = –5.0 + 0.1 (CV) (4b) there are two ways of configuring a fuzzy PI controller.
Approach I. The fuzzy PI controller is designed to
For the process state C, where the CV = 53, Kc may be mimic traditional PI control action, adjusting the MV di-
computed according to: rectly. To construct this type of controller, we begin with

CEP February 2003 www.cepmagazine.org 65

Measurement and Control

NL NM NS Zero PS PM PL
∆E
PL NL NM NS Zero PS PM PL
PM NL NM NS Zero PS PM PL
PS NM NS NS Zero PS PS PM
Zero NM NS Zero Zero Zero PS PM
NS NM NS NS Zero PS PS PM
NM NL NM NS Zero PS PM PL
NL NL NM NS Zero PS PM PL

■ Figure 4. Control surface produced by a 2-D fuzzy PI controller. ■ Figure 5. Fuzzy rule base for tuning parameters for determination.

the velocity form of the PI control law, which states that Step 2. Compute µ1 and µ2 by referring to Eq. 3a and
the manipulated variable movement, ∆Mn, is a function of an illustrative fuzzy plot for En (Figure 3b).
the error and the change in error: ∆Mn = f(En, ∆En): Step 3. Compute µ3 and µ4 by referring to Eq. 3a and
an illustrative fuzzy plot for ∆En (Figure 3b).
T Step 4. Using the fuzzy rule base in Figure 3a, use the
∆Mn = Kc ( ∆En + En ) (6) nearest four rules that encompass the location of the pro-
τI
cess state to compute ∆Mn.
Therefore, a 2-D fuzzy PI controller may be construct- Denoting the four rules as Rulei,j, where i (i =1, 2) rep-
ed to imitate traditional PI control action at every control resents a row index and j (j = 1, 2) represents a column
interval using the known values of En and ∆En. To do so, a index, ∆Mn may be computed using the method of weight-
set of rules is needed to determine what control action to ed averages according to:
take for various combinations of En and ∆En. For example,
a rule might state that if En is positive small (PS) and ∆En 2 2
is positive large (PL), then the control action should be ∑ ∑ Rulei, j µ i µ j +2
i =1 j =1
PL. The linguistic variables (PS, NL, Z, etc.) correspond to ∆Mn = 2 2 (7)
numerical values selected by the designer. The linguistic ∑ ∑ µ i µ j +2
i =1 j =1
control actions suggested by the rules in Figure 3a corre-
spond to illustrative values of ∆Mn shown in Figure 3b.
The stepwise procedure for constructing a 2-D fuzzy PI Example 1. Suppose En = 2 and ∆En = –0.35. Then,
controller is as follows: from Figure 3b µ1 = 0.375, µ2 = 0.0625, µ3 = 0.75 and µ4
Step 1. Compute En and ∆En. = 0.25. Comparing the highlighted area in Figure 3b and
the associated rules together with
their numerical values in Figure
3c, gives: Rule1,1 = Z = 0.0, Rule1,2
CV CV = PS = 1.0, Rule2,1 = NS = –1.0,
PL
and Rule2,2 = Z = 0.0. Therefore,
Defined State

PM
PS
CVi+2 substitution in Eq. 7 results in ∆Mn
= –0.375.
CVi CVi+1 These computations are re-
NS peated at every control interval.
NM CV i
NL Note the diagonal symmetry of the
fuzzy rule base of Figure 3a. This
}

}
}

MV ∆M 5 ∆M 5 ∆M 5
MV
symmetry mimics a PI controller
in the velocity form. The symme-
try may be manipulated by adjust-
■ Figure 6a (left). The manipulated variable movements of six step-tests at a defined controlled-variable value. ing the rules and/or the associat-
■ Figure 6b (right). Multiple step tests conducted within the positive medium (PM) class. ed values of controller actions

66 www.cepmagazine.org February 2003 CEP

 E CV

PS PM PL

NL NM NS Zero PS PM PL
∆E
PL NL NM NS Zero PS PM PL
PM NL NM NS Zero PS PM PL
PS NM NS NS Zero PS PS PM
Zero NM NS Zero Zero Zero PS PM NS NM NL
NS NM NS NS Zero PS PS PM
NM NL NM NS Zero PS PM PL
NL NL NM NS Zero PS PM PL

Tuning Parameter

■ Figure 7. The flow of information for a 3-D fuzzy logic controller.

PS Kc1 Kc2 PM Kc7 Kc8 Step 1. Compute En and ∆En.

τl1 τl2 τl7 τl8 Step 2. Compute µ1 and µ2 by referring
1.0 1.0 to a fuzzy plot for En (Figure 5).
µ6 µ12 Step 3. Compute µ3 and µ4 by referring
to a fuzzy plot for ∆En (Figure 5).
µ

µ5 µ11 The values of En and ∆En point to four

0.0 0.0 adjacent rules in the rule base. In Figure 5,
50 55 58 50 55 58
Rule1,1 = PS, Rule1,2 = PS, Rule2,1 = PS,
Controlled Variable Controlled Variable Rule2,2 = PM. These rules lead to multiple
sets of PI controller tuning parameters,
■ Figure 8. Example of positive small (PS) and positive medium (PM) tuning parameter fuzzy sets. which additionally depend upon the value
of the CV. To obtain these tuning parame-
in Figure 3c to produce nonlinear controller action. Figure ters, a number of suitable open-loop experiments (e.g., step
4 illustrates a classic control surface produced by a 2-D tests) must be conducted. Referring to Figure 5, it would be
fuzzy PI controller. necessary to conduct step tests for each ∆M size (NS, NM,
Approach II. The fuzzy PI controller of approach I NL, PS, PM and PL) and to determine the appropriate val-
will produce the same control action for specific values ues of the PI tuning parameters across the CV range of in-
of En and ∆En, regardless of the location (or value) of the terest. Figure 6a depicts the ∆M of six step tests at a defined
CV. This shortcoming may be eliminated by a fuzzy value of CV. Note that only gain relationships are shown.
logic controller designed to take into account three fac- To proceed further, multiple step tests within each class
tors, En, ∆En and the CVn. We refer to this type of con- (for example, let PM correspond to ∆M = 5) must be con-
troller as a 3-D fuzzy PI controller. This controller can ducted at various starting values of CV and proper corre-
be configured to provide controller tuning parameters or sponding values of Kc and τI determined (Figure 6b). The
∆Mn. The approach outlined here provides optimal con- results of this exercise are grouped leading, to Figure 7.
troller parameters, Kc and τI, as the output. The advan- Step 4. Using the fuzzy rule base in Figure 5, identify
tage of this approach is that the existing PI controller the nearest four rules that encompass the location (defined
configuration may be retained in the distributed control by En and ∆En) of the process state.
system (DCS) or the supervisory control and data acqui- Step 5. Use CVn to determine the membership functions
sition (SCADA) system. for the particular tuning parameter fuzzy set indicated by
The procedure to compute Kc and τI is as follows: each rule.

CEP February 2003 www.cepmagazine.org 67

 Measurement and Control

CV=40% CV=30% CV=20%

Proportional Band

Proportional Band

Proportional Band
∆E ∆E ∆E
E E E

■ Figure 9. Multiple control surfaces produced by a 3-D fuzzy PI controller.

2.5 where Kc (i = 1 through 8) is a controller

gain developed from the step test con-
ducted at CVi.
2 Step 7. Compute the final PI con-
troller gain according to Eq. 7.
Example 2. Suppose that based
Controlled Variable

upon the current values of En and ∆En,

1.5 the four highlighted rules in Figure 5
surround the process state in which, µ1
= 0.3, µ2 = 0.7, µ3 = 0.4 and µ4 = 0.6.
1 Therefore, Rule1,1 = PS, Rule1,2 = PS,
Rule2,1 = PS, Rule2,2 = PM. From this
information, the four rules indicate that
0.5
two tuning parameter fuzzy sets (PS
and PM) should be used, as seen in
Figure 7.
Now, let the current value of CVn be
0
0 5 10 15 20 25 30 35 40 45 50 55 60 55%, on a scale of 0–100%, and assume
that the tuning parameter fuzzy sets
Time, s shown in Figure 8 are available from
step tests. Furthermore, say at an initial
■ Figure 10. Open-loop responses of the single-input single-output system. CV value of 50%, PS and PM step tests
were performed, producing the tuning
parameter sets [Kc = 1.5, τI= 5.0s] and
Step 6. Compute the tuning parameters for each rule [Kc = 1.3, τI = 4.7s], respectively. Likewise, PS and PM
according to: step tests with CV initially at 58% produced [Kc = 3.8, τI =
2.5s] and [Kc = 4.5, τI = 2.0s], respectively. Then, from
Kc, Rule1,1 = µ 5 Kc1 + µ 6 Kc2 (8a) Figure 8, for PS, µ5 = µ7 = µ9 = 0.375 and µ6 = µ8 = µ10 =
0.625, and for PM, µ11 = 0.375 and µ12 = 0.625.
Kc, Rule1,2 = µ 7 Kc3 + µ8 Kc 4 (8b) Note that in this example µ5 = µ11 and µ6 = µ12. This is
because the initial values of CV for the various step tests
Kc, Rule2 ,1 = µ 9 Kc5 + µ10 Kc6 (8c) of PS and PM have been selected to be the same. Using
Eqs. 8a–d in conjunction with the membership functions
Kc, Rule2 ,2 = µ11 Kc 7 + µ12 Kc8 (8d) µ5 through µ8, the controller tuning parameters may be
calculated to

68 www.cepmagazine.org February 2003 CEP

 1.6 dependent of the value of the CV,
which leads to a single control surface.
1.4 In contrast, the 3-D fuzzy PI controller
action, being additionally dependent
1.2
upon the value of the CV, generates a
Controlled Variable

1 control volume.
Commercial processes generally ex-
0.8 hibit two types of nonlinearities. In one
case, the process gain varies across the
0.6 operating range of the CV. This is re-
0.4
ferred to as gain nonlinearity. In the
other, parameters involving time dy-
0.2 namics may also vary across the operat-
ing range. In both cases, the control ob-
0 jective is to use appropriate tuning pa-
0 5 10 15 20 25 30 35 40 45 50 55 60
rameters determined via the fuzzy logic
Time, s concepts described to prevent perfor-
mance degradation, regardless of where
■ Figure 11. Normalized closed-loop responses of the single-input single-output system using a the CV may be located. The next exam-
fixed-term PI controller. ple, built around a first-order system
containing a gain nonlinearity, offers a
1.4
comparative assessment of fixed-term
PI control vs. 3-D fuzzy PI control.
1.2 Example 3. The dual process simulator
Controlled Variable

used in this investigation is a device con-

1 taining electronic circuits capable of re-
producing the behavior of several simple
0.8 dynamic processes. The knob settings on
the simulator allow for the selection of
0.6 system parameters (linear and nonlinear
gain, dead-time, time parameter(s), and
0.4 damping ratios). The simulator can ac-
commodate up to two inputs and two out-
0.2 puts. The inputs and outputs of the simu-
lator are in the standard 4–20-mA range.
0 This example is based on a single-input
0 5 10 15 20 25 30 35 40 45 50 55 60 single-output (SISO) system with gain
nonlinearity.
Time, s The outputs (CVs) from the simulator
■ Figure 12. Normalized closed-loop responses of the single-input single-output system using a are connected to the analog-to-digital
combination of fuzzy logic and PI control. converter (A/D) inputs of a process ter-
mination panel by suitable wiring. The
A/D signals are available to a PC-based control system
Kc, Rule1,1 = Kc, Rule1,2 = Kc, Rule2 ,1 = 2.938, Kc, Rule2 ,2 = 3.3; and upon demand at every sampling interval. The SCADA
system used is FIX32 from GE Fanuc Intellution. The PI
τ I , Rule1,1 = 3.438s and τ I , Rule2 ,2 = 3.013s.
control law is executed using programming tools provided
by FIX32. In the “local” mode, the user may enter the tun-
The final tuning parameters are ing parameters of the controller for fixed-term operation.
In the “remote” mode, they may be adjusted by a fuzzy
Kc n = 3.09 and τ I n = 3.26s. logic program that is dynamic-data-exchange (DDE) com-
pliant. In either case, the digital PI controller outputs are
converted into continuous signals with digital-to-analog
Figure 9 show the multiple control surfaces produced converter (D/A) modules in the process termination panel.
by a 3-D fuzzy PI controller. The 2-D fuzzy PI controller Further wiring connects the continuous controller outputs
previously discussed produces a single control action in- to the inputs of the dual process simulator.

CEP February 2003 www.cepmagazine.org 69

Measurement and Control

The open-loop responses of the

100
Linear Region 3 SISO system are shown in Figure 10.
The normalized closed-loop responses
80 of this system containing a fixed-term
PI controller, tuned at the nominal
Controlled Variable, %

steady-state operating point, to step

60 Nonlinear Region 2 changes in setpoints of various magni-
Linear Region 2 tudes are shown in Figure 11. The
degradation in performance due to the
40 nonlinearity is evident.
The fuzzy logic concepts are used
Linear to determine the state of the process
Region 1
20 at every sampling interval. The in-
Nonlinear Region 1 ferred state allows for the determina-
tion of appropriate tuning parameters,
0 as previously illustrated. The use of
0 20 40 60 80 100
these tuning parameters results in
Manipulated Variable A, %
vastly improved performance, as
shown in Figure 12.
■ Figure 13. Essentially, what has been
Fuzzy achieved is that although the open-loop
Manipulated variable A
exhibits a pH-type gain system is nonlinear, the closed-loop
-
+ MVA nonlinearity. system is approximately linear. This
PI
Nonlinear opens up the prospect of adjusting the
MVB Process CV setpoint of the closed-loop linear sys-
ONLINE ■ Figure 14. Control
R strategy of the 1 × 2 tem with a linear CMPC to achieve
process. constrained optimization.

NL NM NS Z PS PM PL Combining fuzzy logic

1 with CMPC
0.8 Linear CMPC has made a significant
impact in the process industries, since it
0.6 can achieve throughput maximization,
µE

0.4 energy minimization and improved

quality control. However, its applica-
0.2
tions are limited to linear systems.
0 Fuzzy logic concepts can linearize non-
-20 -15 -10 -5 0 5 10 15 20
linear systems when gain nonlinearity is
Error present or prevent performance degrada-
■ Figure 15a. Schematic of error fuzzy set. tion when dynamic system parameters
are nonlinear. Consequently, a frame-
work may be visualized to extend the
NL NM NS Z PS PM PL
1 applicability of CMPC to nonlinear sys-
tems by combining fuzzy logic concepts
0.8 with linear CMPC, as shown in the fol-
lowing example. This example can be
0.6 applied to pH systems, fermentation sys-
µ∆E

tems, chemical reactors, steam boilers

0.4
and dryers.
0.2 Example 4. The dual process simu-
lator previously described was used.
0 The multivariable system selected is a
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 one-output by two-input nonlinear sys-
Change in Error tem. Manipulated variable A (MVA) has
■ Figure 15b. Schematic of change-in-error fuzzy set. a pH-type gain nonlinearity shown in

70 www.cepmagazine.org February 2003 CEP

 3 D Fuzzy PI Controller Fixed-term PI Controller
50 50
40 40
Variable, %
Controlled

30 30
20 20
10 10
0 0
0 20 40 60 80 100 0 25 50 75 100 125
Proportional Band

120 0.2 30 0.25

100 25 0.2

Reset (1/s)
0.19
80 20 0.15
60 0.18 15
0.1
40 10
0.17 0.05
20 5
0 0.16 0 0
0 20 40 60 80 100 0 25 50 75 100 125

100 100
Set Point, %

80 80
Controller

60 60
40 40
20 20
0 0
0 20 40 60 80 100 0 25 50 75 100 125

50 50
Variable B, %
Manipulated

40 40
30 30
Maximize
20 20 Maximize
10 10
0 0
0 20 40 60 80 100 0 25 50 75 100 125

60 80
Variable A, %

50
Manipulated

60
40
30 40
20
20
10
0 0
0 20 40 60 80 100 0 25 50 75 100 125

Time, s

■ Figure 16. Optimization results for a setpoint change from 20% to 40%.

Figure 13, while the effect of the manipulated variable B from Six Sigma and Advanced Controls, Inc. In this in-
(MVB) on the CV is linear. stance, a dynamic data exchange facility is required to per-
Figure 14 is a schematic of the 1 × 2 process and the form two functions: (1) to update the tuning parameters of
control strategy. The CMPC software used is ONLINE the 3-D fuzzy PI controller and (2) to fetch the CV from

CEP February 2003 www.cepmagazine.org 71

Measurement and Control

3D Fuzzy PI Controller Fixed-term PI Controller

30 30
25 25
Variable, %
Controlled

20 20
15 15
10 10
5 5
0 0
0 25 50 75 100 125 0 25 50 75 100 120
Proportional Band

120 0.25 30 0.25

100 0.2 25 0.2

Reset (1/s)
80 0.15 20 0.15
60 15
0.1 0.1
40 10
20 0.05 5 0.05
0 0 0 0
0 25 50 75 100 125 0 25 50 75 100 125

100 100
Set Point, %

80 80
Controller

60 60
40 40
20 20
0 0
0 25 50 75 100 125 0 25 50 75 100 125

50 50
Variable B, %
Manipulated

40 40
30 30
Minimize
20 20 Minimize
10 10
0 0
0 25 50 75 100 125 0 25 50 75 100 125

50 50
Variable A, %

40 40
Manipulated

30 30
20 20
10 10
0 0
0 25 50 75 100 125 0 25 50 75 100 125

Time, s

■ Figure 17. Optimization results for a –10% load disturbance.

the FIX32 database, making it available to ONLINE, and In this application, the 3-D fuzzy PI controller uses the
to write outputs from ONLINE back to the FIX32 rule base of Figure 5 in conjunction with the error and
database. The ONLINE outputs are the 3-D fuzzy PI con- change-in-error fuzzy sets of Figures 15a and 15b.
troller setpoint and the MVB. To account for the nonlinear effect of the MVA upon

72 www.cepmagazine.org February 2003 CEP

the CV, the experimental 3-D fuzzy Table. ONLINE manipulated variable constraints
logic knowledge base adjusts the tun- and optimization parameters.
ing parameters of the PI algorithm
within FIX32. ONLINE uses an aver- Minimum Maximum Move Maximum
aged step-response model derived Penalty Move
from Figure 12 and a linear step-re- Controller Setpoint 10 90 20 10
sponse model relating the MVB to the Manipulated Variable B 10 40 4.5 None
CV for optimization of the nonlinear1
× 2 process.
To demonstrate constrained opti-
mization of the 1 × 2 nonlinear system, an optimization ob- controller tuned for operation in linear region 1 are
jective and the process constraints are defined for two tests, shown on the right side of Figure 16. The performance
one for servo control and the other for regulatory control. improvements when ONLINE is used with 3-D fuzzy PI
The constraints and certain ONLINE tuning parameters are controller are evident.
listed in the table. The second test is for a step-type load disturbance of
For the first test, the ONLINE setpoint is changed –10% introduced into the CV. The initial steady-state
from 20% to 40%, while maximizing the MVB. The ini- value of CV is 25% (near the nonlinear region 1). The op-
tial steady state values of the CV, the MVA, and the timization objective in ONLINE is specified to minimize
MVB are 20%, 30%, and 10%, respectively. Figure 13 the use of the MVB with an initial steady state value of
shows that CV = 20% lies in linear region 1. The set- 20%. The initial value of the MVA value is 23%. The con-
point change causes the CV to move across nonlinear re- straints and tuning parameters used are the same as those
gion 1 into linear region 2. The results of this test are used in the first test.
shown on the left side of Figure 16. The results with Figure 17 presents the results of this test. Due to the
ONLINE used in conjunction with the fixed-term PI load disturbance, both control strategies take corrective
action with overshoot into nonlinear region 1. However,
the performance of ONLINE used in conjunction with 3-D
Literature Cited fuzzy PI controller is considerably better. CEP

1. Deshpande, P. B., et al., “Should You Use Constrained Model

Predictive Control?,” Chem. Eng. Prog., 91 (3), pp. 65–72 (Mar.
1995). PRADEEP B. DESHPANDE is professor and a former chair of the Dept. of
2. Garci, C. E., and M. Morari, “Internal Model Control 1. A Unify- Chemical Engineering at the University of Louisville (Speed Scientific
School, Chemical Engineering Dept., Ernst Hall 106, Louisville, KY
ing Review and Some New Results,” Ind. Eng. Chem. Process Des.
40292; E-mail: [email protected]). He is also is
Dev., 21, p. 308 (1982).
president and CEO of Six Sigma and Advanced Controls, Inc., in
3. Cutler, C. R., and B. L. Ramaker, “Dynamic Matrix Control — A Louisville, Kentucky. Deshpande has supervised over 70 master’s and
Computer Control Algorithm,” presented at AIChE National Meet- doctoral students to date. He has authored five textbooks in process
ing, Houston, TX, (Apr. 1979). control and related areas and has published or presented over 100
4. Richalet, J., et al., “Model Predictive Heuristic Control: Applica- papers. He is program leader of the AIChE’s Short Course, No. 611, “Six
tions to Industrial Processes,” Automatica, 14, p. 413 (1978). Sigma for Global Competitiveness.” He is a recipient of several awards
5. Zadeh, L. A., “Fuzzy Sets,” Information and Control, 8, pp. including ISA’s Donald P. Eckman Award in Process Control Education.
338–353 (1965). Deshpande is a member of AIChE and a Fellow of ISA.
6. Rhinehart, R. R., and P. Murugan, “Improve Process Control
ANDREW C. BURDEN is currently a doctoral student in chemical
Using Fuzzy Logic,” Chem. Eng. Prog., 92 (11), pp. 60–65 (Nov.
engineering at the University of Louisville. He obtained his bachelor’s
1996). and master’s degrees in chemical engineering and an MBA from the
7. Chen, G., and T. T. Pham, “Introduction to Fuzzy Sets, Fuzzy University of Louisville. Burden was a member of the research team
Logic, and Fuzzy Control Systems,” CRC Press LLC, Boca Raton, that recently implemented ONLINE, a constrained model predictive
FL (2001). controller software, real-time optimization, and three-dimensional
8. De Carli, A., et al., “A Fuzzy-PI Control Strategy,” Chem. Eng. fuzzy logic on an industrial-scale boiler at the University of Louisville.
Practice, 2 (1), pp. 147–153 (1994).
9. Lilly, J. H., “Stable Fuzzy Regulation of Nonlinear Systems via ROBERTO Z. TANTALEAN is currently a doctoral student in chemical
engineering at the University of Louisville. He obtained his bachelor’s
Polynomial Surface Fitting,” submitted to IEEE Transactions on
degree in chemical engineering from Universidad de Trujillo, a
Fuzzy Systems (Feb. 2001).
master’s degree in computer science from Universidad de Cantabria in
Spain, and a master’s degree in chemical engineering from the
University of Louisville. Tantalean was a member of the research team
Acknowledgement that recently developed and implemented a real-time communication
facility for the constrained model predictive controller software
The review and comments of Hatim Al-Dekhiel, a PhD scholar at the ONLINE, and implemented an expert system for fault monitoring and
University of Louisville sponsored by SABIC, Saudi Arabia, are abnormal situation management on an industrial-scale boiler at the
appreciated. University of Louisville.

CEP February 2003 www.cepmagazine.org 73