Journal of Process Control 122 (2023) 113–133

Contents lists available at ScienceDirect

Journal of Process Control

journal homepage: www.elsevier.com/locate/jprocont

Review

Transformed inputs for linearization, decoupling and feedforward

control
∗
Sigurd Skogestad a , , Cristina Zotică a , Nicholas Alsop b
a
Department of Chemical Engineering, Norwegian University of Science and Technology (NTNU), 7491, Trondheim, Norway
b
Senior Process Control Engineer, Borealis AB, Stenungsund, Sweden

article info a b s t r a c t

Article history: This paper introduces powerful static input transformations which transform the original system
Received 21 June 2022 (process) into a transformed system which is easier to control. The transformed inputs (controller
Received in revised form 17 November 2022 outputs) may be implemented in many ways and under many names, for example, as ratio, feedforward
Accepted 20 December 2022
and decoupling control, or more generally as nonlinear computation blocks. These methods are
Available online xxxx
frequently used in industry, but are often introduced in an ad-hoc fashion. The present paper provides
Keywords: a systematic method for deriving such control strategies from a nonlinear process model. For a static
Nonlinear process control model, the ideal transformed input is simply the right-hand-side of the model equations. The resulting
Feedforward control transformed system is linear, decoupled and independent of disturbances. In some cases, use of extra
Model-based control measurements simplify the input transformation by replacing model equations. It is also possible to
Cascade control
derive ideal transformed inputs from a dynamic model, which turns out to be a special case of a
PID control
nonlinear control approach called feedback linearization. However, except for achieving linearization
also dynamically, the benefits of using feedback linearization are small compared to using transformed
inputs based on a static model. For implementation we need to invert the input transformation, and
for this we may use an exact model-based inverse or an approximate feedback-based inverse. The
latter leads to the use of cascade control.
© 2022 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).

Contents

1. Introduction....................................................................................................................................................................................................................... 114
2. Motivating case study: Decoupling of mixing process ................................................................................................................................................ 115
2.1. Example 1: Mixing process with flowrates as physical inputs...................................................................................................................... 116
2.2. Example 2: Mixing process with valve positions as physical inputs ............................................................................................................ 116
3. Implementation of transformed inputs ......................................................................................................................................................................... 118
3.1. Alternative A: Model-based inversion (Figs. 1 and 8(a)) ................................................................................................................................ 118
3.2. Alternative B: Feedback inversion with slave v -controller (cascade control) (Fig. 8(b))............................................................................ 118
3.3. Alternative C: Combined feedback- and model-based inversion with slave w -controller (Fig. 8(c)) ....................................................... 118
4. Derivation of ideal transformed inputs ......................................................................................................................................................................... 118
4.1. Obtaining an ideal transformed system from a static process model .......................................................................................................... 119
4.1.1. Choice of the tuning parameter B0 .................................................................................................................................................... 120
4.2. Obtaining ideal transformed input from a dynamic process model ............................................................................................................. 120
4.2.1. Choice of tuning parameter B............................................................................................................................................................. 121
4.2.2. Choice of tuning parameter A............................................................................................................................................................. 121
4.3. Transformed input in terms of measured state variables w .......................................................................................................................... 121
4.3.1. Dynamics of transformed system with measured state (w ) variables .......................................................................................... 122
4.3.2. Possible unstable zero dynamics and internal instability ............................................................................................................... 122
4.4. Ideal static transformed variable v0 applied to a dynamic system............................................................................................................... 122
5. Examples ........................................................................................................................................................................................................................... 122
5.1. Example 3. Ideal transformed inputs for mixing process (Motivating Example 1, continued) ................................................................. 122
5.1.1. Ideal transformed input v0 from static model ................................................................................................................................. 123

∗ Corresponding author.
E-mail address: Sigurd.Skogestad@ntnu.no (S. Skogestad).

https://doi.org/10.1016/j.jprocont.2022.12.012
0959-1524/© 2022 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

5.1.2. Comparison with engineering-based variables from Motivating Example 1................................................................................ 123
5.1.3. Applying the ideal static transformed input v0 to the dynamic system....................................................................................... 123
5.1.4. Ideal transformed input vA from dynamic model ............................................................................................................................ 123
5.2. Example 4. Heated tank ...................................................................................................................................................................................... 124
5.3. Example 5: Simple nonlinear level process...................................................................................................................................................... 124
5.4. Example 6: Heat exchanger................................................................................................................................................................................ 125
5.4.1. Ideal transformed input v0 based on full static model ................................................................................................................... 126
5.4.2. Transformed input v0,w based on parts of static model and measured state w = T2 ................................................................ 126
5.4.3. Implementation .................................................................................................................................................................................... 126
5.4.4. Simulations ........................................................................................................................................................................................... 127
6. Discussion.......................................................................................................................................................................................................................... 127
6.1. Outer controller C for case with ideal transformed inputs............................................................................................................................ 127
6.2. Time scale separation for feedback implementation (alternative B)............................................................................................................. 127
6.3. Decoupling and disturbance rejection when ideal static transformed input v0 is applied to the dynamic system in (18) ................ 128
6.4. Chain of input transformations .......................................................................................................................................................................... 129
6.5. Comparison with feedback linearization .......................................................................................................................................................... 129
6.6. Comparison with nonlinear MPC ....................................................................................................................................................................... 130
6.7. Transformed outputs ........................................................................................................................................................................................... 130
6.8. Internal instability with exact inversion........................................................................................................................................................... 130
6.8.1. Example: Simple linear system with unstable zeros....................................................................................................................... 131
6.8.2. Linear stability analysis of transformed system ............................................................................................................................... 131
6.8.3. Relationship between internal stability and stable zero dynamics for output y (for systematic transformed inputs v0 and
vA ) .......................................................................................................................................................................................................... 132
7. Summary and conclusion ................................................................................................................................................................................................ 132
Declaration of competing interest.................................................................................................................................................................................. 133
Data availability ................................................................................................................................................................................................................ 133
Acknowledgments ............................................................................................................................................................................................................ 133
Appendix. Tuning parameter A for ideal dynamic transformed inputs vA ................................................................................................................ 133
References ......................................................................................................................................................................................................................... 133

Shinskey [1] (on page 119) writes in relation to selecting input

1. Introduction and output variables for the controller:

Industry frequently makes use of nonlinear static model-based ‘‘There is no need to be limited to single measurable or ma-
‘‘calculation blocks’’, ‘‘function blocks’’, or ‘‘ratio elements’’ to nipulable variables. If a more meaningful variable happens to
provide feedforward action, decoupling or linearization (adaptive be a mathematical combination of two or more measurable
gain), and Shinskey [1] provides many examples of this. The main or manipulable variables, there is no reason why it cannot be
motivation for this work is to provide a better theoretical basis for used’’.
these model-based nonlinear control elements, which we study in Some simple examples of transformed inputs (controller out-
this paper in the context of nonlinear input transformations. puts) are
Let u denote the original (physical) input and let v denote
the transformed input (controller output) which depends on u v =u+d (2a)
and other variables. The main idea is that the controller C (or in u
some cases the operator) sets the value of the transformed input
v= (2b)
d
v rather than the physical input u, see Fig. 1. In this paper, we v = u1 − u2 (2c)
define the transformed input v as a nonlinear static function g of
u1
the physical input u and other variables v= (2d)
u2
v = g(u, w, y, d) (1) v=w (2e)
Note that the specific function g is a design choice for the control Such transformed inputs are often introduced by engineers on
engineer. The variables are defined as follows: simple physical grounds. The transformed input v = u + d in (2a)
v= transformed inputs (controller outputs) may provide feedforward action from a measured disturbance d.
u= original (physical) process inputs It may be used, for example, for a case where u and d represent
d= measured disturbances two feedrates and we want to control the combined flowrate
y= controlled process outputs (measured) u + d. The ratio v = ud in (2b) may provide feedforward action
and linearization. It is typically used when u and d represent two
w= other measured dependent process variables (states).
feedrates and we want to control the quality (e.g. composition) of
In this paper, we do not include dynamic elements in the the combined feed. A transformed variable with two inputs may
definition of the transformed input v , although dynamic elements provide decoupling, for example the difference v = u1 − u2 in
(2c) and the ratio v = u1 in (2d).
u
are frequently used in industrial practice (e.g., [2]).
2
Nevertheless, even without dynamics, (1) provides a very The transformed input v = w in (2e) with w = F (measured
generic definition so let us state more clearly the objective of flowrate) is probably the most common of all in process control.
introducing the transformed input. Its implementation results in a slave flow controller, where the
The transformed input v replaces the physical input u as the physical input u is the valve position and the transformed input
manipulated variable for control of the output y, with the aim of sim- v is the flowrate F (or more precisely, v = Fs , the setpoint for
plifying the control task by including elements such as decoupling, the flowrate). This particular transformed input (v = F ) and
linearization and feedforward action. the use of slave flow control is so common that in many cases
114
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

keeping only parts of the benefit of the input transformation,

for example decoupling, and leave the disturbance rejection to
the outer feedback controller C . The other approach is to use
a soft sensor, estimator or observer (e.g., [6,7]) to estimate the
non-measured variables d and w .
Surprisingly, there does not seem to be any academic liter-
ature on the common industrial approach of Shinskey [1,3] of
using static models to derive transformed inputs. On the other
hand, for dynamic models, there is a large body of mathematical
Fig. 1. Use of transformed inputs v . For example, the transformed input could
be the ratio v = g(u, d) = ud , and the ‘‘inverse input transformation’’ block which theory on variable transformations to transform nonlinear differ-
inverts this relationship would then be u = g −1 (v, d) = v d. The outer feedback ential equations into linear differential equations, which has been
controller C has the aim of correcting for uncertainty, including for model error applied in the control field. The most well-known approach is
and unknown disturbances. feedback linearization based on mathematical concepts from Lie
algebra (e.g., [5,6,8,9]). This theory is closely related to the input
transformations for dynamic systems studied in this paper. How-
people consider the flowrate F to be the (physical) input u, and to ever, the theory of feedback linearization, although extensively
simplify the expressions we will sometimes do this in this paper. taught in nonlinear control classes, is rarely used in industrial
However, it is not enough to define the transformed input practice, at least within the field of process control. There are
v , as in (1) and (2). For implementation we need to invert the several reasons for this. One is that the mathematics are seem-
transformation, that is, to generate from a given value of the ingly complicated and there are restrictions in terms of the class
controller output v = g(u, w, y, d), the corresponding physical of systems and the assumption of full state feedback. Another
input u. Shinskey [1,3] calls this ‘‘reversing the process model’’. reason is that, mainly for reasons of mathematical generality and
There are two main ways of generating this inverse: simplicity, Isidori [5] selects the transformed inputs such that the
resulting transformed linear system is integrating, dt = v . This
dy
A. Model-based inverse1 using the ‘‘exact inverse input trans-
means that the transformed system is at the limit of instability,
formation’’ u = g −1 (v, w, y, d); see Fig. 1.
so the transformed inputs v cannot be kept constant. For example,
B. Feedback-based inverse using a cascade implementation
with a fixed v , any unmeasured disturbance will result in an
with a slave controller for v or w .
integrating output y. Therefore, Isidori [5] introduces an outer
The model-based approach may be used for (2a)–(2d) above. state feedback controller as part of the solution. However, in
For example, for v = g(u, d) = u+d in (2a) the exact model-based many cases it is strongly desirable to be able to fix v , at least on an
inverse becomes intermediate time scale, and actually the transformation into an
integrating system is not necessary. For example, [6,7], who study
u = g −1 (v, d) = v − d (3) process control applications, use a formulation that gives a stable
linear transformed system on the form dt = Ay + Bv (where the
dy
However, a model-based inverse is sometimes not possible,
because g(u, w, y, d) may not depend explicitly on u. For ex- matrices A and B are tuning parameters), and this is the approach
ample, for the transformed input v = F (flowrate), we have taken in this paper. In a personal communication, Isidori [10]
g(u, w, y, d) = w = F which is independent of the physical input emphasizes that the decision to make the transformed system
u. In this case, the inverse must be generated by feedback using integrating was just chosen as an example, but this message has
a slave flow controller. not made its way to many of the potential users of feedback lin-
In other cases, a model-based inverse may exist, but neverthe- earization. For our purposes, the advantage with the large body of
less we may use a feedback controller as a ‘‘trick’’ to generate an literature on feedback linearization, is that this literature provides
approximate inverse. In the control literature, this trick is often a mathematical basis for issues related to the invertibility and
referred to as ‘‘dynamic inversion’’ [4]. stability of the proposed transformations.
In most cases, the selection of transformed inputs v is based The paper starts in Section 2 with a motivating mixing exam-
on simple static models, for example, from material or energy ple on the use of input transformations for decoupling. Next, in
balances [1,3]. However, the treatment of Shinskey is case-study Section 3 we discuss in more detail the two main approaches for
based and in this paper, we aim to show how to derive the implementing the transformed inputs, which are exact model in-
transformed variables in a systematic manner. Ideally, assuming version and inversion by feedback (cascade control). In Section 4,
no model error and that we measure all disturbances, these ‘‘ideal we show how to derive ‘‘ideal’’ model-based transformed inputs
transformed variables’’ result in a transformed system from v that provide linearization, decoupling and feedforward control.
to y (see Fig. 1) that is linear, decoupled and independent of The model can either be a static model or a low-order dynamic
disturbances. model. In the dynamic case, the theory is closely related to the
We also show how this approach can be extended to dynamic theory of feedback linearization. In Section 5 we present several
models, and this case is closely related to the theory of feedback case studies. In Section 6, we discuss the results and study the po-
linearization [5], which has a strong theoretical basis. tential problem of internal instability. We also discuss the use of
Throughout the paper we assume that we measure all the transformed outputs. In Section 7 we make our final conclusions
parameters that enter the transformations, such as disturbances and remarks.
d and internal variables (states) w . This is often not true, so We have attempted to keep the mathematical treatment at a
in practice there are two alternatives. The most common is to quite low level, so that the paper will be readable also for an
simplify the expression for the ideal transformed variable, by industrial audience. One reason is that we strongly believe that
the results in this paper can be very useful in industrial practice.
1 In this paper, the notation g −1 means that we invert or reverse the static
2. Motivating case study: Decoupling of mixing process
function g between independent and dependent variables. For example, if the
original function is v = g(u, d), where u is the independent variable and v the
dependent variable, then the solution that results from solving v = g(u, d) with The main reason for introducing transformed inputs v is to
respect to u for a given v is written as u = g −1 (v, d). simplify the control of the outputs y. In Section 4, we introduce
115
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

Both ratios vr and v2 give decoupling and they are equivalent

in the sense that fixing one keeps the other constant (since
v2 = vrv+r 1 ). However, the normalized ratio v2 in (4b) has some
properties that makes it better for implementation. First, v2 is
always in the range 0 to 1, whereas vr may vary between 0
and ∞. Second, as shown later in Example 3, the ratio v2 is a
special case of the ideal static transformed input (v0 ) and provides
linearization. Third, the expression for the exact inverse in the
decoupling block becomes very simple with v2 , see (5). Fourth, v2
avoids division by zero when F2 = 0, which may be an advantage
Fig. 2. Flowsheet of mixing process. if we obtain the inverse by feedback control (Alternative B in
Fig. 6).
To find the exact inverse transformation (decoupling block),
systematic methods for selecting ‘‘ideal’’ transformed input v that we solve the expression v = g(u) for the transformed input in
provide linearization, decoupling and disturbance rejection. How- (4) with respect to u for a given v , to derive u = g −1 (v ) where in
ever, in many cases, engineers use simpler transformed inputs [1] this case
that do not provide all these features. In this section, we consider u1 = F1 = v1 v2 (5a)
a simple motivating case study.
u2 = F2 = v1 − v1 v2 (5b)
−1
2.1. Example 1: Mixing process with flowrates as physical inputs The nonlinear decoupling g in (5) may be implemented in
the block ‘‘inverse input transformation’’ in Fig. 1 to provide a
The mixing process in Fig. 2 has two inlet streams, and initially decoupled response from v to y for the traditional two-handle
we consider for simplicity the two flowrates F1 and F2 [kg/s] design.
as the physical inputs, rather than the valve positions. The inlet Note that the inverse transformation contains no divisions,
flows are mixed to get a given total flow (F ) and quality (T ), which which is important to avoid division by zero. Furthermore, in
we want to control. Depending on the application, T could rep- this simple case, the equations can easily be implemented using
resent temperature or composition. We will consider the mixing standard multiplication and subtraction elements, for example, as
of hot (F1 ) and cold (F2 ) water where T is temperature. The main shown in Fig. 4.
disturbances are the two inlet temperatures. Thus, we have: In Section 5.1 we will use a systematic procedure to derive a
[ ] [ ] [ ] generalized (ideal) version of the transformed inputs in (4), where
F1 F T1 disturbance rejection and linearization are also included.
u= ; y= ; d=
F2 T T2
2.2. Example 2: Mixing process with valve positions as physical
A real design of this process using a traditional faucet with two
inputs
separate handles (valves) is shown in Fig. 3(a). We know that
this process is quite interactive. For example, to increase the
We have so far assumed that the flow rates F1 and F2 are the
temperature y2 = T while keeping a constant total flow y1 = F , physical inputs, but in practice it is more likely that the valve
we need to increase the input u1 = F1 (hot water) while reducing positions z1 and z2 are the physical inputs and that the flowrates
u2 = F2 (cold water) by the same amount. are extra measured variables w :
[ ] [ ] [ ] [ ]
z F T F1
Input transformation for decoupling u= 1 ; y= ; d= 1 ; w=
z2 T T2 F2
To eliminate the interactions and make the process decoupled,
we may use the alternative one-handle faucet in Fig. 3(b). Here, We will now consider three ways of implementing this in
one direction of the handle (usually up–down) is used for adjust- order to retain a decoupled transformed system from the trans-
ing the total flow (F = F1 + F2 ), and the other direction (usually formed input v in (4) to the output y.
left–right) is used for adjusting the temperature by changing the Alternative A (purely model-based inversion). The first option
F
ratio F1 of hot and cold water. This corresponds to using the does not make use of any measured flows (w ). It is based on
2
inverting the entire model, including the valve model. A general
transformed inputs v1 = F1 + F2 and vr =
F1
. The fact that these
F2 block diagram for this option is shown in Fig. 1, and the corre-
transformed inputs give decoupling is probably clear on physical sponding flowsheet for this particular example is shown in Fig. 5.
grounds, and it can easily be proven by making use of the mass A typical valve equation is
and energy balances as shown later in Example 3 (Section 5.1). √
For the modern one-handle design in Fig. 3(b), the transformed F = kv fv (z) ∆P (6)
variables v1 and vr are implemented physically (mechanically). Here, F is the flow, z is the valve position, kv is the valve constant
However, to implement decoupling using the traditional two- and ∆P is pressure drop over the valve which is assumed to be
handle design in Fig. 3(a), we need to add in the control scheme a measured disturbance. The valve characteristic fv (z) is assumed
a decoupling block to compute the physical inputs (u) from the to be known. For a linear valve, we have fv (z) = z. Inverting the
transformed inputs (v ). With this digital rather than physical valve Eq. (6) gives
decoupler, we may use the opportunity to replace the flow ratio ( )
F
vr = FF1 by the normalized ratio v2 = F F+1F . The transformed z = fv−1 √ (7)
2 1 2
kv ∆P
inputs v = g(u) then become
For a linear valve, we get z = k √F ∆p . The resulting model-based
v1 = F1 + F2 = u1 + u2 (4a) v
inversion may be implemented as shown in Fig. 5, where the
F1 u1 exact inverse block u = g −1 (v, d) computes the valve positions
v2 = = (4b)
F1 + F2 u1 + u2 u = z (physical input) by combining the inversions in (5) and (7).
116
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

Fig. 3. Valve (faucet) designs for mixing hot and cold water.

Fig. 4. Implementation of inverse transformation in (5) using standard multipli-

cation and subtraction elements. The two transformed inputs are the sum (v1 )
and the normalized ratio (v2 ) of the two inputs (u1 and u2 ).

Fig. 6. Feedback inversion (Alternative B) for implementation of transformed

inputs v1 = F1 + F2 and v2 = F +1F for the mixing process when the physical
F
1 2
inputs u are the valve positions z. This solution makes use of slave v -controllers
(VC1 and VC2).

This alternative avoids the inverse block g −1 , so instead the

decoupling from v to y is taken care of by the two slave
v -controllers (VC1 and VC2). With integral action in the two
v -controllers, decoupling will be exact at steady state.
However, decoupling will not be perfect dynamically, both
Fig. 5. Model-based inversion (Alternative A) for implementation of transformed because the slave v -controllers are not infinitely fast and because
inputs v1 = F1 + F2 and v2 = F +1F for the mixing process when the physical
F
1 2 of interactions between the loops. To reduce interactions, we
inputs u are the valve positions z. The inversion block u = g −1 (v, d) is a need to choose good variable pairings for VC1 and VC2. In Fig. 6,
combination of the inversions in (5) and (7).
we have paired the total flow (v1 = F1 + F2 ) with z1 (which affects
F1 ), and the ratio (v2 ) with z2 (which affects F2 ). This can be shown
to be the preferred pairing if F1 the larger flow.
Specifically, (5) is used to calculate the flowrates F1 and F2 , and Since interactions result in dynamic couplings from v to y with
(7) is used to calculate the corresponding valve positions z1 and the purely feedback-based implementation (Alternative B), one
z2 . may question if there is any benefit compared to the simplest
However, in practice, this model-based implementation may scheme, which is to have no input transformation (not shown in
not work well, mainly because of uncertainty (error) in the valve any figure), that is, letting the outer controller C (which are the
characteristic fv (z) and the valve constant kv , but also because of two controllers TC and FC in this case) manipulate directly the
incorrect measurements of the pressure drop disturbance (∆P). flows (F1 and F2 ) using two slave flow controllers.
Therefore, if the inlet flows (F1 , F2 ) can be measured with reason- The answer is that there can indeed be a significant decoupling
able accuracy, it is better to use a feedback-based implementation benefit when we use input transformations with feedback control
where the feedback corrects for model uncertainty. Two such of v (Alternative B) if there are effective delays associated with
implementations are discussed next. the control of y (e.g., measurement delays for F or T in our
Alternative B (purely feedback-based inversion). The first option case), such that two slave v -controllers (VC1 and VC2) can be
significantly faster than outer controllers in C (TC and FC).
(Fig. 6) uses only feedback. The measured w -variables (flows F1
Alternative C (combined model- and feedback-based inversion).
and F2 ) are used to compute the actual value of the transformed
The best option for this example is most likely to use the exact
input v (v1 = F1 + F2 and v2 = F +1F ), and two slave v -controllers
F
1 2 model-based inverse in (5) to compute the desired flowrates
(VC1 and VC2) manipulate the physical inputs u (valve positions) (F1s and F2s ) and combine this with two slave flow (w = F )
to make v equal to the desired setpoints v s which are computed controllers (FC1 and FC2) that generate by feedback the corre-
by the two master controllers (FC and TC). sponding valve positions (u1 = z1 and u2 = z2 ). This alternative
117
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

3.1. Alternative A: Model-based inversion (Figs. 1 and 8(a))

The first approach is to invert the input transformation

v = g(u, w, y, d) in (8), by analytically or numerically finding the
input u that corresponds to given values of v, w, y and d. We can
formally write the solution as

u = g −1 (v, w, y, d) (9)
This gives the exact inverse g (v, w, y, d) if the inverse exists,
−1

if there is no model uncertainty and if all variables w, y and d are

measured perfectly.

3.2. Alternative B: Feedback inversion with slave v -controller (cas-

cade control) (Fig. 8(b))

A dynamic approximation of the inverse input transforma-

tion may be generated using an inner (slave) feedback controller
Fig. 7. Combined model and feedback inversion (Alternative C) for implemen- Cv as shown in Fig. 8(b). Here, we compute the actual value
tation of transformed inputs v1 = F1 + F2 and v2 = F +1F for the mixing process
F
1 2 v = g(u, w, y, d) from measurements of u, w, y and d, and
when the physical inputs u are the valve positions z. This solution combines
use the inner controller Cv to dynamically generate the input
the exact inverse g (v ) in (5) for computing the flow setpoints F1s and F2s with
−1

two slave flow(w )-controllers (FC1 and FC2) that generate the corresponding u that makes v approach the desired value v s . The inner con-
physical inputs u1 = z1 and u2 = z2 . troller Cv may be tuned based on an experimental response from
u to v . Usually, a PI-controller, or even a pure I-controller, is
sufficient. For the n × n multivariable case, one usually designs
is shown in the flowsheet in Fig. 7. It combines the decoupling n single-loop linear controllers for Cv , although it is possible
of the model-based inversion in (5) with the linearization and to use multivariable control, for example, to reduce dynamic
disturbance rejection of the flow controllers. The two slave flow interactions.
controllers, through the action of fast feedback, indirectly gener-
ate the inverse in (7), and makes the process (almost) linear and 3.3. Alternative C: Combined feedback- and model-based inversion
independent of disturbances in ∆p1 and ∆p2 , at least as seen from with slave w -controller (Fig. 8(c))
slower time scale of the outer controllers (TC and FC).
This combined implementation with an inner w -controller is
The next section generalizes the three alternative ways (A, B
of particular interest for the case when v = g(w, y, d) does
and C) of generating the inverse.
not depend explicitly on u and there is only one measured w -
variable associated with each input u. In addition to the inner
3. Implementation of transformed inputs controller Cw for w , we also need a block that inverts the transfor-
mation g with respect to w , that is, which computes the setpoint
The transformed input v is defined as a nonlinear static func- ws = g −1 (v, y, d).
tion g that depends on the original (physical) input u and other Feedback implementation C (Fig. 8(c)) has the advantage that
measured variables we can include some model-based inversion, which may con-
tribute to linearization, feedforward and decoupling. It also has
v = g(u, w, y, d) (8) the advantage that control of w (Alt. C) is usually less interactive
All variables may be vectors. For the multivariable case, we will than control of v (Alt. B), which is a significant advantage for
assume that we have an equal number (n) of inputs u, outputs y faster convergence with single-loop control. A feedback imple-
and transformed inputs v . Often the function g is independent of mentation (alternatives B or C) is required if v does not explicitly
y and in many cases we do not have extra measurements w . Note depend on u. In other cases, a feedback implementation may
that g may not depend explicitly on u, but it should then depend be used as a ‘‘trick’’ for numerically generating an approximate
indirectly on u through the measured variables w . inverse. In the control literature, this trick is often referred to as
‘‘dynamic inversion’’ [4]. The reason for using this trick could be
As mentioned in the introduction, the idea is that the outer
to avoid the complexity of deriving the inverse in (9) (see heat
controller C or the operator sets the desired value (setpoint) v s
exchanger example) or to avoid problems with singularities [4].
of the transformed input v . However, to implement v on the real
process, we need to generate the corresponding physical input
4. Derivation of ideal transformed inputs
u. There are two main approaches for implementing the physical
input u:
Input transformations are in common use and as illustrated in
A. Model-based implementation, see Figs. 1 and 8(a). With the motivating mixing example they may be very useful. How-
an exact inverse g −1 , this gives exactly v = v s (assuming ever, the main question we want to answer in this paper is:
all variables are measured perfectly and there is no model
error). How do we derive good transformed inputs in a systematic
B. Feedback-based implementation, see Fig. 8(b). With inte- manner?
gral action in the slave controller Cv this gives v = v s after
a dynamic transient. Starting from a static or dynamic process model, we show in
this section how to derive ideal transformed inputs (denoted v0
We also discuss a third implementation (Fig. 8(c)) which is a and vA , respectively) which ideally achieve linearization, decou-
combination of the two. The three alternatives are the same as pling and disturbance rejection, see Fig. 9. We assume that we
the ones presented in Figs. 5–7 for Example 2, respectively. have an n × n control problem with n inputs u and n outputs y,
118
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

Fig. 8. Alternative implementations (red blocks) for inverting the input transformation v = g(u, w, y, d), that is, of generating the physical input (u) from a given
value of the transformed input (v or v s ). Controllers C , Cv and Cw are usually single-loop PID controllers.

and we want to use the model equations to find n transformed

inputs v . The case with a static model is discussed in Section 4.1
and a dynamic model in Section 4.2. Note that we may combine
static and dynamic models as shown in the mixing example
in Section 5.1. In Section 4.3, we discuss that it is simpler in
many cases to write the model in terms of extra measured state
variables w , because we can then replace model equations by
measured variables.

4.1. Obtaining an ideal transformed system from a static process

Fig. 9. With ideal transformed inputs the transformed system (with no uncer-
model tainty) is linear and independent of disturbances. For a static process model, the
transformed input v = v0 = g(u, w, d) is given in (13) or (27). For a dynamic
In the industrial literature, Shinskey [1] shows by examples process model, the transformed input v = vA = g(u, w, y, d) is given in (19)
or (30). In most cases, the tuning parameters (B0 , A, B) are chosen such that
how to use static process models to derive nonlinear feedforward
the ideal transformed system is decoupled and gives y = v at steady state. The
and decoupling blocks which are similar to the input trans- inversion g −1 may alternatively be generated by feedback, see Fig. 8(b), but in
formations derived below. However, he provides no systematic this case the ideal properties of the transformed system are achieved only after
approach, and very surprisingly, for the simple and important a dynamic transient.
case of static systems, there seems to be no academic literature
on how to do derive static feedforward and decoupling blocks
in a systematic manner. Possibly this is because the derivation Consider a static process model with n independent equations
is almost trivial, as shown in the next few lines. We will first written in the following general form
consider the case with no measured states w , and we will then
generalize to include w -variables. 0 = f (u, y, d) (10)
119
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

or even more generally as n + nx equations in the form 4.2. Obtaining ideal transformed input from a dynamic process
model
0 = fx (u, x, y, d) (11)
where x represents additional internal variables (states). In the We next examine the case where we have a dynamic process
more general case in (11), we assume that we can use the nx extra model as given in (18). The derivation of the resulting ideal
equations to eliminate the internal variables x to get a model (at transformed input vA is closely related to the theory of feedback
least formally) as given in (10). linearization for the special case of a low-order model (with a
Since the model equations in (10) are assumed to be indepen- relative order of 1).
dent, they may be solved with respect to y (at least formally) to Ideal transformed input based on dynamic model. Consider
get the static model on the form y = f0 (u, d). We then have the a nonlinear dynamic model in the form
following general result.
Ideal transformed input based on static model. Consider a dy
= f (u, y, d) (18)
static nonlinear model in the form dt
y = f0 (u, d) (12) For the model in (18), the ideal transformed input is

Define from this the ideal static transformed input vA = B−1 (f (u, y, d) − Ay) (19)
  
g(u,y,d)
v0 = B0 f0 (u, d)
−1
(13)
Here, the matrices A and B are tuning parameters. Assume that vA
  
g(u,d)
can be exactly implemented by solving (19) with respect to u to get
where the matrix B0 is free to choose. Assume that v0 can be exactly
implemented by solving (13) with respect to u to get the ideal input
u = g −1 (vA , y, d) (20)
u = g −1 (v0 , d) (14) Then assuming no uncertainty (no model error for f (u, y, d) and
perfect measurements of d and y) the transformed system becomes
Then, assuming that the real system is static with model f0 (u, d)
(no model error) and that we have perfect measurement of d, the dy
transformed system becomes
= Ay + BvA (21)
dt
y = B 0 v0 (15) The transformed system in (21) is linear and independent of dis-
turbances, and for the multivariable (n × n) case, it is also decoupled
The transformed system in (15) is linear and independent of distur-
if we select A and B to be diagonal matrices.
bances, and for the multivariable case it is also decoupled if we select
B0 to be a diagonal matrix.
Proof. Substituting the transformed input in (19) into (18) gives
(21). Note that we have assumed that we can generate from
Proof. The proof is trivial. From (13) we get f0 (u, d) = B0 v0
the transformed input vA the exact corresponding physical input
and substituting this into (12) gives y = B0 v0 in (15). The
assumptions related to (14) are necessary to be able to generate u. □
the corresponding ideal input u. □ Note that we use the notation with a subscript A to show
Note that we use the notation with a subscript 0 to show that that vA is an ideal transformed input derived based on a dynamic
v0 is an ideal transformed input derived from a static model. model and with a tuning parameter A.
Importantly, it may not be necessary to explicitly derive an It may seem that (18) represents a large class of dynamic
expression for y = f0 (u, d) in (12). Rather, since the objective is models, but actually it is quite restrictive since we must assume
to find the ideal input u = g −1 (v0 , d) that gives the transformed that the number of differential equations (states) n is equal to
system y = B0 v0 in (15), it is in some cases simpler to stay the number of inputs and outputs in the vectors u and y. In
with the original model equations in (10) or (11), and rather than particular, we assume that the input u directly affects the time
dy
solving them with respect to y (for a given u), solve them with derivative dt of the controlled output y, which means that the
respect to u for a given y = B0 v0 to obtain the inverse trans- relative order from u to y is assumed to be 1. Specifically, for the
formation u = g −1 (v0 , d). The solution to find the inverse g −1 scalar case (n = 1), we assume that we can write the model for y
can be either analytical or numerical. A numerical solution may using only one scalar differential (18). Thus, for the scalar case we
be necessary for complicated models, like for the heat exchanger are restricted to a first-order system. (However, if we allow the
example discussed later and in [11]. function f to depend on additional measured states w , then the
class of systems is significantly larger. This is discussed in more
4.1.1. Choice of the tuning parameter B0
detail below.)
The choice of B0 is not critical, as it can be compensated by
To guarantee invertibility in (20), it is possible to restrict the
changing the gain of the outer controller C . We usually choose
class of models to guarantee that we always have a solution, as
B0 = I (16) is done in the literature on exact linearization. In particular, in
this literature it is assumed that the model is linear in the input
such that the ideal transformed input is simply the right-hand
u, that is, that we can write the right-hand side of (18) as shown
side of the model equation,
in [8] (p. 293).
v0 = f0 (u, d) (17)
f (u, y, d) = f1 (y, d) + f2 (y, d) u (22)
In this case the transformed system simply becomes y = v0 ,
and it may be tempting to think of the transformed input v0 as where the functions f1 and f2 must satisfy certain smoothness
the setpoint for the output y, but this is misleading because we conditions. Interestingly, many process models are linear in the
usually have an outer feedback controller C which has the ‘‘true’’ flows, so if we make use of inner flow controllers then many
setpoint ys as one of its inputs, whereas v0 is the output from C process models satisfy (22). Nevertheless, we do not make this
(see Fig. 1). Thus, it is better to think of v0 as the transformed assumption in this paper, so the invertibility may need to be
process input. studied separately for each application.
120
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

4.2.1. Choice of tuning parameter B simple, we do not recommend the choice A = 0 for most process
To get dynamic decoupling in (21) for the multivariable case, control applications. The main reason is that with A = 0, a
we need to select both matrices B and A to be diagonal. Dynamic system which was originally stable is turned into an integrating
decoupling is desirable because the optimal outer controller C is transformed system. Thus, the transformed system will not go
then diagonal (single-loop controllers). Otherwise, the choice of to steady state without the outer controller C . In particular, any
B is not critical as it may be compensated by changing the gain unmeasured disturbances will cause the output y to change in
in the feedback controller C . a ramplike fashion and drift away from its desired steady state.
One simple choice is B = I, which is often used in feedback This drifting will only stop when the input u reaches its physical
linearization. Alternatively, to keep the initial (high-frequency) maximum or minimum constraint. This is undesirable, because
gain from vi to yi equal to that of the original system (from ui usually one wants to be able to operate the transformed system
to yi ) one may choose B = diag(B̃) = diag(∂ f /∂ u)∗ where the without the outer controller C . Another reason for not selecting
differentiation is performed at the nominal operating point ∗ .
A = 0 is that we generally want to use integral action in the
However, in most of the examples in this paper we select
outer controller C to correct for uncertainty. With A = 0, the
B = −A (23) integrator in the transformed system poses additional perfor-
mance limitations for disturbances at the plant input (e.g., [12]).
because this gives y = vA at steady state (where = f (u, y, d) =
dy
dt This performance limitation is not considered in the feedback
0).2 With the choice B = −A, the transformed input and corre- linearization literature because they assume state feedback, that
sponding transformed system become is, they assume C is a P-controller.
vA = −A−1 f (u, y, d) + y (24a)
dy 4.3. Transformed input in terms of measured state variables w
= A(y − vA ) (24b)
dt
The expressions for the ideal transformed inputs (v0 and vA )
4.2.2. Choice of tuning parameter A
are the same also when we include additional measured depen-
The choice of the parameter A is important as it determines the
dent variables w (states) in the model equations (in f0 and f ). This
dynamics of the transformed system. However, the importance
allows us to use simpler models, because a measured w replaces
should not be overemphasized, since we can change the closed-
a model equation for w . For example, a measured flow w = F
loop dynamics by design of the outer controller C . Note that we
must choose A < 0 for the transformed system to be stable. We may replace a valve Eq. (6).
discuss below three choices for the tuning parameter A. Following the same derivations as above, we find that the
1. Keep the original dynamics, A = Ã. In most cases we expressions v = g(u, w, y, d) for the ideal transformed inputs
propose selecting remain the same, except that we add w in the argument list.
Then, assuming that we measure d and w perfectly, have a perfect
∂f
( )
A = Ã ≡ (25) model, are able to generate the exact inverse g −1 and that the
∂y ∗ resulting transformed system is internally stable, the resulting
where the derivative is evaluated at the nominal point ∗ of oper- transformed system becomes as before linear and independent
ation. This makes the dynamics of the transformed system equal of disturbances.
to the linearized dynamics of the original system. This choice also Ideal transformed input with w -variables for static case.
minimizes the effect of the measurements y on the transformed Specifically, for the static model
variables vA (see Appendix). This seems reasonable because the
y = f0,w (u, w, d) (26)
outer controller C in any cases makes use of the measurements
y. the ideal transformed input becomes
In the multivariable case, to obtain decoupling, we may choose
A = diag(Ã). v0 = B− 1
f0,w (u, w, d) (27)
0
2. Make the transformed system faster: |A| > |Ã|. To speed
 
g(u,w,d)
up the response from v to y, one may use larger magnitudes
for the elements in A than that resulting from (25). However, resulting in the transformed system
note that the presence of a time delay in the measurement of
y = B0 v0 (28)
y (or other dynamics that result in an effective delay) may give
instability if we choose the elements in A too large in magnitude. Ideal transformed input with w -variables for dynamic case.
Alternatively, note that it is always possible to select A = Ã Similarly, for the dynamic model
as in (25) and instead ‘‘speed up’’ the response with the outer
controller C , which can be designed based on the experimental dy
= fw (u, w, y, d) (29)
response from vA to y and for which established robust design dt
methods are available, for example, the SIMC PID-rules [12]. the ideal transformed input becomes
3. Make the system integrating: A = 0. The choice A = 0
is recommended in the standard feedback linearization litera- vA = B−1 (fw (u, w, y, d) − Ay) (30)
dy
ture [5]. This results in an integrating transformed system, dt =
  
g(u,w,y,d)
BvA , where usually one selects B = I, corresponding to selecting
the transformed input as the right-hand side of the differential resulting in the transformed system
equation model, vA = f (u, y, d). However, although being very dy
= Ay + BvA (31)
dt
2 Interestingly, since y = I v at steady state, where I is the identity matrix,
A
In most cases, we choose B0 = I and B = −A which give
the choice B = −A gives decoupling at steady state even if A (and thus B) is
not diagonal. However, to also get dynamic decoupling, we must choose A to y = v0 and y = v at steady state. Choosing A diagonal also gives
be diagonal. decoupling for the dynamic case.
121
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

4.3.1. Dynamics of transformed system with measured state (w ) dynamically, if we initially are at steady state. This surprising fact,
variables which is observed in several of the examples and proved in the
If the w -variables depend dynamically on the inputs u, then discussion in Section 6.3, holds because of the particular simple
the implicit feedback through the measurements w will affect dynamics assumed in (29). The assumption about being initially
the dynamics of the transformed system. In some cases, this may at steady state is not limiting, because it is desirable that the
result in internal instability as discussed below. system stays close its steady state setpoint.
For the static case, the dynamics of the transformed system Also note that when v0 is independent of w , i.e., we have
(from v0 to y) will no longer be the same as that of the original v0 = f0 (u, d), then there is no feedback from the outputs or states,
system (from u to y). and the transformed dynamic system (from v0 to y) retains the
The reason is that the feedback from w through v0 may either dynamics of the original system (from u to y) without needing
speed up or slow down the response. Note that for the static any tuning parameter. These two facts makes it very tempting
case, we have no tuning parameter to change the dynamics of for an engineer to apply v0 also to dynamic systems.
the transformed system. However, there are some advantages of instead applying the
dynamic transformed input vA (30) to the dynamic system dt =
dy
4.3.2. Possible unstable zero dynamics and internal instability f (u, w, y, d). The main advantage is that the transformed system
Note that we are essentially treating the variable w as a from vA to y is linear dynamically, whereas the transformed
measured disturbance when deriving the transformed variables
system from v0 to y is linear only at steady state. This is seen in
v0 and vA in (27) and (30). Is there any problem in doing this? several of the examples. Second, for the case when the ideal static
Yes, there will be a problem with the exact inverse, if the w -
transformed input v0 depends on measured states w , the system
variables result in unstable zero dynamics3 in the dynamic map
dynamics change because of the resulting feedback from w to u,
from the physical input u to the transformed input v (which
but we have no parameter in v0 to affect it. On the other hand,
may be v0 or vA or any other transformed input), The reason
when we use the ideal dynamic transformed variable vA , we can
is that unstable zero dynamics will result in internal instability
choose the dynamics of the linear transformed system through
for the transformed system if we implement the exact inverse
the parameter A.
u = g −1 (v, w, y, d). This follows because the unstable zeros of
the original map become unstable poles of the inverse map. A
simple example is given in the discussion section. This means 5. Examples
that the implementations with the exact inverse (Alternative A) in
Fig. 8(a) may yield internal instability in some cases. Fortunately, 5.1. Example 3. Ideal transformed inputs for mixing process (Moti-
as argued in the discussion section, it is not very likely to happen vating Example 1, continued)
in practice, because unstable zero dynamics require that the
indirect dependency of u on v through w is strong. Consider the mixing process in Fig. 2 with the following inputs,
The internal instability can in any case be avoided if we use the outputs and disturbances
alternative implementation B in Fig. 8(b) with a slave v -controller [ ] [ ] [ ]
Cv , but Cv then needs to be tuned sufficiently slow so that the F F T1
u= 1 ; y= ; d= (32)
unfavorable zero dynamics do not cause closed-loop instability. F2 T T2
Thus, linearization, decoupling and disturbance rejection will not This is the same process as in Example 1, where to obtain
be perfect dynamically in this case. decoupling we used engineering insight to propose a sum and a
Note that these problems with unstable zero dynamics are ratio of the flows as transformed inputs, see (4). In this section,
fundamental and do not arise because we use transformed inputs. we will derive the ideal transformed inputs using systematic
In particular, for the ideal transformed inputs (v0 and vA ), the zero
methods. Note that we assume that the flows F1 and F2 are the
dynamics from u to v are the same as the zero dynamics from u to
physical inputs u, that is, we are implicitly assuming that we have
y. This follows because of the direct relationship between v and y
two flow controllers.
in this case, for example, y = v0 for the static case. Our problem
For simplicity, we assume that the mass m [kg] of the system is
with internal stability with the exact inverse, is thus a special case
constant, which is a good assumption in many cases. The dynamic
of the well-known fact that perfect control of the output y cannot
mass balance dm = F1 + F2 − F then gives by setting dm = 0, the
be achieved for a system with unstable zero dynamics (RHP-zeros dt dt
following static mass balance
in the linear case) from u to y, no matter how good the model is
or what we measure (e.g., [13]). F = F1 + F2 (33)

4.4. Ideal static transformed variable v0 applied to a dynamic system Assuming perfect mixing, the dynamic energy balance becomes
dT
If we apply v0 in (27) (derived from a static model) as a m = F1 T1 + F2 T2 − −FT (34)
dt
transformed input to a dynamic system (but with the same static
model), then with the choice B0 = I, the transformed system We have here assumed constant and equal heat capacities so that
becomes y = v0 and we clearly get linearization, decoupling and cP drops out of the energy balance. Substituting the mass balance
disturbance rejection at steady state. (33) into the energy balance (34) and using the more general
But what happens dynamically? We cannot say anything in notation in (32), then gives the following model equations for the
general, but fortunately, if we apply v0 to the particular dynamic mixing process
system in (29), then we get perfect disturbance rejection also
y1 = u1 + u2 (35a)
  
3 Unstable zero dynamics go by many names. They are the same as RHP-zeros f0,1 (u,d)

for linear systems, and linear systems with RHP-zeros and/or time delay are also dy2 1
called non-minimum phase systems. In the linear scalar case, RHP-zeros always = (u1 (d1 − y2 ) + u2 (d2 − y2 )) (35b)
dt m
give inverse response in the time domain. More generally, for nonlinear systems  
the unstable zero dynamics from u to v correspond to the unstable dynamics f2 (u,y,d)
of the inverse map from v to u [5].

122
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

5.1.1. Ideal transformed input v0 from static model in (39) is not truly decoupled, because we see from (39b) that v0,1
We first derive the ideal transformed inputs for a purely static also affects output y2 . However, for practical purposes, we have
dy
model. By setting m dt2 = 0 we derive from (35b) the following decoupling, because if we start from a steady-state operating
static equation for the temperature y2 = T point, where we have y2 = v0,2 , then (39b) tells that a change
u1 d 1 + u2 d 2 in v0,1 will not affect y2 .
y2 = (36)
u1 + u2
5.1.4. Ideal transformed input vA from dynamic model
  
f0,2 (u,d)
Actually, since the static transformed input v0 gives both dis-
With the standard choice B0 = I, the ideal static transformed turbance rejection and decoupling also for the dynamic process,
inputs v0 are simply the right-hand side f0 of the static model there is little reason to consider the dynamic transformed input
equations. Thus, the ideal static transformed inputs for the mixing
vA , but we will do it for completeness and comparison.
tank are
The first model equation is static, so the first transformed
v0,1 = u1 + u2 (37a) input is unchanged
  
g1 (u)=f0,1 (u,d)
v0,1 = u1 + u2 (40)
u1 d1 + u2 d2
v0,2 = (37b) To derive the second ideal transformed input, we use the right-
u1 + u2
   hand side f2 (u, y, d) of the dynamic energy balance in (35b), and
g2 (u,d)=f0,2 (u,d) from (24a) we derive

The static model for the transformed system becomes y = v0 , vA,2 = y2 − A−1 f2 (u, y, d) (41a)
or equivalently y1 = F = v0,1 and y2 = T = v0,2 . As ex- 1
pected, the transformed static system is decoupled, independent = y2 − A−1 (u1 (d1 − y2 ) + u2 (d2 − −y2 )) (41b)
m
of disturbances and linear (with gain equal to the identity matrix,   
I). g2 (u,y,d)
For implementation using the exact inverse, we need to invert Note that we have chosen B = −A which gives y2 = vA,2 at steady
the expressions (37) for v0 = g(u, d) to find the physical inputs
state. From (40) and (41b) the inverse transformation becomes
(flows) u. We get u = g −1 (v0 , d) where
v0,1 (v0,2 − d2 ) v0,1 (y2 − d2 ) − Am(vA,2 − y2 )
u1 = (38a) u1 = (42a)
d1 − d2 d1 − d2
v0,1 (d1 − v0,2 ) v0,1 (d1 − y2 ) + Am(vA,2 − y2 )
u2 = (38b) u2 = (42b)
d1 − d2 d1 − d2

There is a singularity in the inverse transformation when the two The transformed system from the ideal transformed inputs v =
inlet flows have the same temperature, d1 = d2 . This is not a [v0,1 vA,2 ] to the outputs y = [y1 y2 ] then becomes
limitation of the proposed method, because it is then physically
y1 = v0,1 (43a)
impossible to freely set the temperature y1 = T of the mixed
flow. dy2
= A(y2 − vA,2 ) (43b)
dt
5.1.2. Comparison with engineering-based variables from Motivat-
which is decoupled, independent of disturbances and linear since
ing Example 1
A is a constant.
Comparing the ideal static transformed inputs in (37) with the
The constant A is a tuning parameter. To eliminate the feed-
engineering-based variables in (4), we see that v0,1 is the sum
u1 + u2 as before. The second variable v0,2 is very similar to the back from the output y2 = T to the transformed variable v2
ratio v2 = u +1u in (4b), except that v0,2 includes feedforward
u in (41b) at the nominal operating point, we choose A such that
1 2
action from disturbances d1 = T1 and d2 = T2 . we keep the nominal linearized dynamics of the original system,
which from (25) gives
5.1.3. Applying the ideal static transformed input v0 to the dynamic (
∂ f2
)
F∗ v0∗,1
system A= =− =− (44)
What happens if we apply the static transformed input v0,2 to ∂ y2 ∗ m m
the dynamic system in (36)? Substituting u1 and u2 from (38a) where F ∗ = u∗1 + u∗2 = v0∗,1 is the nominal total flowrate.
and (38b) into (35b) gives after a little algebra the following Note that the expression for the transformed system in (39b) in
transformed dynamic system terms of v0,2 (static model) is very similar to (43b) in terms of
y1 = v0,1 (39a) vA,2 (dynamic model) if we choose A as given in (44). The main
dy2 v0,1 difference is that the transformed dynamic system (43b) for vA,2
= (v0,2 − y2 ) (39b) is linear, whereas the transformed dynamic system (39b) for v0,2
dt m
is nonlinear because of the multiplication with the term v0,1 .
We see that both disturbances d1 and d2 drop out, so the In summary, the only advantage of using the more complex
transformed system in (39) based on v0,2 is independent of distur- variables vA rather than v0 derived from a static model, is that
bances, also dynamically.4 We note that the transformed system the transformed system is linear dynamically. This benefit is small
compared to the added complexity, for example, of having to
4 Generally, when we apply static transformed inputs v to a dynamic system select the tuning parameter A. Thus, most likely an engineer will
prefer to use the static variables v0 . If the disturbances d1 = T1
0

of the form dt = f (u, y, d), we need to make the assumption that the system is
dy
and d2 = T2 are not measured or do not change frequently, then
initially at steady state to get perfect dynamic disturbances rejection. However,
this assumption is not necessary for this particular case since the disturbances this is equivalent to the simple sum and ratio proposed in (4) in
drop out completely in the transformed system. the Motivating example.
123
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

Fig. 11. Flowsheet of the nonlinear level process.

Fig. 10. Process flowsheet of tank heated with an electric coil. We note that there is a singularity at v0 = T0 . This may be
a problem, because it may happen that the outer controller C
makes a large decrease in v0 (possibly to speed up the response)
so that v0 drops below T0 . This will cause the input u to jump
5.2. Example 4. Heated tank
from a large positive value (in practice, with constraints, from
Consider the continuous process in Fig. 10 with an electric u = umax ) to a large negative value (in practice, to u = umin = 0).
heater. Assuming perfect mixing, constant heat capacity cP [kJ/°C] We simulated this (not included), and found that it made the
and constant mass holdup m [kg], the energy balance gives the system drift away from the desired steady state, and it did not
following dynamic model recover. Fortunately, there are ways of handling the singularity.
dT 1 One is to use logic and set u = umax when the computed value of
= f (u, y, d) = (FcP (T0 − T ) + Q ) (45) u is negative. Another way is to use the cascade implementation
dt mcP
in Fig. 8(b) with a slave v -controller.
The objective is to control the outlet temperature y = T using A similar singularity occurs with the ideal transformed input
the inlet flowrate u = F [kg/s] as the manipulated input (we vA . Inverting (47b) gives
assume that we have a fast slave flow controller so that we can
consider u = F to be the physical input). Q and T0 (heat input and Q + AmcP (vA − y)
u=F = (50)
inlet temperature) are measured disturbances. Setting dTdt
= 0, we cp (y − T0 )
derive the corresponding static model for the outlet temperature
The singularity at y = T0 may happen in situations with large
Q dynamic variations in d = T0 or y = T , but can be handled in a
T = f0 (u, d) = T0 + (46)
FcP similar way as mentioned for v0 .
From (27) and (30), the ideal transformed inputs v0 and vA , based
on a static and dynamic model, respectively, become 5.3. Example 5: Simple nonlinear level process
Q
v0 = f0 (u, d) = T0 + (47a)
FcP We consider the control of level (volume) in a tank with two
inflows and one outflow, as shown in Fig. 11. We assume that the
( )
F Q
vA = −A−1 f (u, y, d) + y = −A−1 (T0 − T ) + +T level measurement is delayed or infrequent, so we would like to
m mcP
make use of feedforward action from measured disturbances.
(47b)
Assuming constant liquid density, we derive from the mass
We have here chosen the parameters B0 = I and B = −A so that balance the following dynamic model
we have at steady state y = v0 and y = vA , respectively.
dV
If we apply these two transformed variables to the dynamic = F1 + F2 − F3 (51)
system in (45) then the transformed dynamic system becomes dt
for the two cases The variables are defined as:
dT F
= (v 0 − T ) (48a) y = V = volume of fluid (measured)
dt m u = F1 = inflow (manipulated)
dT d1 = F2 = inflow (measured disturbance)
= −A(vA − T ) (48b)
dt F3 = outflow (not measured)
For both cases, we find that the transformed system is indepen-
( ∂f ) d2 = p0 = pressure difference between before and after tank
dent of disturbances (in Q and T0 ). If we choose A = − ∂ T ∗ = (measured).
∗
− Fm as recommended in (25), then we see that transformed
We assume that the inflow u = F1 can be manipulated directly,
systems in terms of v0 and vA are identical close to the nominal
for example, because we have a fast slave flow controller. This
operating point (∗), but note that the transformed system ((47a))
in terms of v0 is nonlinear, whereas the transformed system is not an integrating process, because the outflow F3 depends
((47b)) in terms of vA is linear. However, the nonlinearity in (48a) on the level (V ) and provides self-regulation. The outflow is
is small unless F varies a lot, so most likely an engineer will prefer not measured, but from the Bernoulli equation we derive the
to use the simpler static variable v0 . following model
For implementation using the exact inverse in Fig. 8(a), we √ √
F 3 = Cv ρ gV /A + p0 = c1 V + c2 p0 (52)
need to invert the expression for the transformed input v to find
the physical input u = F . For v = v0 , based on a static model, where c1 = Cv ρ g /A and c2 = Cv are constants. The dynamic
2 2
(47a) gives the inverse transformation model then becomes
Q dy √
u=F = (49) = f (u, y, d) = u + d1 − c1 y + c2 d2 (53)
cp (v0 − T0 ) dt
124
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

Transformed variable v0 based on static model. Solving 0 = gain is inversely proportional to the value of the outflow F3 .
f (u, y, d) in (53) with respect to y gives the static model However, the transformed system with v0 is linear at steady state,
so if the outer loop is slow, the dynamic nonlinearity will not
(u + d1 )2 − c2 d2
y = f0 (u, d) = (54) matter. Thus, again an engineer probably will most likely prefer
c1 the static transformed input v0 , which also does not depend on
The ideal transformed variable (selecting B0 = 1) is the right- the delayed or infrequent measurement of the volume y.
hand side of the static model Transformed variable vA based on dynamic model for the case
with measurement of w = F2 . From the examples so far, it
(u + d1 )2 − c2 d2
v0 = f0 (u, d) = (55) seems that there is not much benefit of using the ideal dynamic
c1 transformed input vA instead of the simpler static transformed
For a given v0 , the corresponding input is given by the inverse input v0 . However, we here consider a case where this is not true,
transformation simply because the system is integrating and thus has no steady
√ state.
u = f0−1 (v0 , d) = c1 v0 + c2 d2 − d1 (56) Consider again control of level (volume V ) in Fig. 11, with the
same dynamic model (51), but assume here that the outflow F2
At steady state the transformed system is y = v0 , so it is linear
is measured, that is,
and independent of the disturbances d1 and d2 . What happens if
we apply the ideal static transformed input v0 to the dynamic sys- w = F2 .
tem? Substituting (56) into (53) gives the transformed dynamic
If this measurement is fast and accurate, then
√ it is better to use
system
it rather then the model equation F3 = k k1 V + p0 which is
dy √ √ inaccurate and also depends on measuring the disturbance d2 =
= c1 v0 + c2 d2 − c1 y + c2 d2 (57)
dt p0 . Introducing w = F2 , the dynamic model dVdt
= F1 + F2 − F3
Thus, the transformed dynamic system is independent of d1 but it becomes
depends on d2 . However, for practical purposes we have perfect dy
= fw (u, w, d1 ) = u + d1 − w (63)
disturbance rejection also for d2 . To see this, note that y = v0 at dt
steady state. It then follows that if we are initially at steady state Note that the right-hand side (fw ) is independent of y so this
and we keep v0 constant, then from (57) we have dt = 0 for
dy
model is integrating with no steady-state value for y. Therefore,
any disturbance d2 . Thus, y will remain at v0 and we have perfect we cannot derive a corresponding transformed input v0 . Instead,
disturbance rejection for d2 also dynamically. we derive the ideal dynamic transformed input vA . Selecting B =
To see this even more clearly, we linearize the transformed 1 and A = 0 in (30) gives the transformed input
system model (57) around the steady-state (where y∗ = v0∗ ) to
get vA = fw (u, w, d1 ) = u + d1 − w (64)

d ∆y c1 which is simply the net inflow to the process. The transformed

= √ (∆v0 − ∆y) (58) system becomes dt = vA , which is linear, integrating and in-
dy
dt ∗
2 c1 y∗ + c2 d2 dependent of disturbances. Given vA , we get from (64) that the
where the ∆-variables represent deviations from the√steady state. corresponding input u = F1 is
Note that d2 drops out in (58). Also note that F3∗ = c1 y∗ + c2 d∗2
u = vA − d1 + w (65)
so the initial dynamic response will vary as a function of F3∗ and
thus depend on the operating point. Of course, we need an outer controller C to adjust vA and stabilize
Transformed variable vA based on dynamic model. For complete- the integrating transformed system, but otherwise the implemen-
ness and comparison, we also here consider the ideal dynamic tation is very simple.
transformed input vA . From (19), we have vA = B−1 (f (u, y, d) − Let us finally remark that it is possible to select vA = fw (u, w,
Ay). We choose B = −A so that y = vA for the transformed system d1 ) − Ay = u + d1 − w − Ay where A < 0. In this case, the
transformed system dt = vA + Ay is stable, although the original
dy
at steady state. With this choice we get
√ system in (63) is integrating. However, selecting A ̸ = 0 introduces
vA = y − A−1 f (u, y, d) = y − A−1 (u + d1 − c1 y + c2 d2 ) (59) feedback from y in the input transformation, which will come in
addition to the feedback from y in the outer controller C , which
  
g(u,y,d)
anyway could be used to stabilize the system. Thus, in this case
and the resulting transformed dynamic system becomes as ex- there is little benefit of selecting a more complicated vA with
pected A ̸ = 0.
dy
= A(y − vA ) (60) 5.4. Example 6: Heat exchanger
dt
The exact inverse of (59) gives the corresponding input Temperature control using heat exchangers may benefit from
√ the use of input transformations, both to reduce nonlinearity
u=g −1
(vA , y, d) = A(y − vA ) − d1 + c1 y + c2 d2 (61)
and to introduce feedforward control. Consider the process in
To eliminate the feedback from the y to vA at the nominal point, Fig. 12 where the objective is to control the outlet temperature
we may from (25) choose of stream 1 (which may be the process side) by exchanging heat
with stream 2 (which may be the utility side). We assume that
∂f
( )
c1 the input (manipulated variable) is the utility flowrate, u = F2 . In
A= =− √ (62)
∂y ∗ 2 c1 y∗ + c2 d∗2 summary, we have for this example
As expected, the linearized transformed system (58) for v0 u = F 2 , y = T1
is identical to dt = A(y − vA ) in (60) if we select A as in
dy
Measured disturbances are the inlet temperatures and the
(62). However, there is an important difference. With vA , the
flowrate of stream 1,
transformed system is linear dynamically because A is a constant,
whereas with v0 the dynamic response is nonlinear as the initial d = [T10 T20 F1 ]⊺
125
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

5.4.1. Ideal transformed input v0 based on full static model

We assume that the fluids do not change phase and have
constant heat capacity (cp1 , cp2 ). Assuming ideal countercurrent
flow, the steady-state behavior is then given by the following
three equations for the heat transfer Q from stream 1 to stream
2:
Q = F1 cp1 (T10 − T1 ) (68a)
Q = F2 cp2 (T2 − T20 ) (68b)
( ) ( )
T10 − T2 − T1 − T20
Q = UA ( ) (68c)
Fig. 12. Heat exchanger (Example 6) where the objective is to control the outlet T10 −T2
ln
temperature T1 of stream 1 (process side) by exchanging heat with stream 2 T1 −T20
(utility side).
This gives 3 equations in 3 unknowns (Q , T1 , T2 ) which can be
solved analytically to find the following analytic expression for T1
as a function of the input and the disturbances (e.g., [15])
In the simulations, we will also consider an unmeasured dis-
turbance in the UA-value, for example, caused by flow changes, y = T1 = T10 + ϵ (T20 − T10 ) (69)
fouling or gas bubbles in the streams,   
f0 (u,d)
dunmeasured = UA where
A possible extra measurement (in addition to F2 ) which depends 1−E
on the input u is the utility outlet temperature ϵ=
C −E
w = T2 F1 cp1
C =
F2 cp2
The dynamic and steady-state behaviors of heat exchangers are ( ( ))
1 1
highly nonlinear. For example, for small values of u = F2 (relative E = exp UA −
dy F1 cp1 F2 cp2
to F1 ), the process gain k = du is large and relatively constant, but
for large values of u = F2 , the gain k approaches 0 and makes it From (69) the corresponding ideal static transformed input be-
difficult to control y = T1 . This is because we get a pinch for T1 comes
(constant value) with y = T1 approaching the inlet temperature
T20 . v0 = f0 (u, d) = T10 + ϵ (T20 − T10 ) (70)
An ideal countercurrent heat exchanger is modeled by partial where it should be noted that ϵ is a nonlinear function of F1 , F2
differential equations, but we use a cell model with n = 100 well- and UA.
mixed cells on each side; see [14] for model equations. In total,
this gives 200 differential equations to represent the temperature 5.4.2. Transformed input v0,w based on parts of static model and
dynamics, so this model clearly cannot be written in the form measured state w = T2
dy
dt
= f (u, y, d) in (18) which allows for only one differential The second transformed variable, v0,w , follows by using the
equation. This leads us to consider transformed inputs based on measured state w = T2 to replace the heat transfer Eq. (68c) for
a static model of the heat exchanger, which we will then apply Q . We use (68a) to find
to the dynamic heat exchanger cell model. Thus, in this example,
Q
we analyze by simulation the effect of model uncertainty. We will T1 = T10 +
consider two transformed inputs F1 cp1

v0 = f0 (u, d) (66) and then we substitute Q using (68b) to get

F2 cp2
v0,w = f0,w (u, w, d) (67) y = T1 = T10 + (T20 − T2 ) (71)
F1 cp1
The first is the ideal transformed input v0 that follows from the   
detailed static model y = f0 (u, d). Note that this model does not f0,w (u,w,d)

depend on the measured state variable w = T2 and use of the From (71) the corresponding ideal static transformed input be-
transformed variable v0 will therefore retain the dynamics of the comes
original system (the heat exchanger).
F2 cp2 0
The second transformed variable, v0,w , is inspired by an actual v0,w = f0,w (u, w, d) = T10 + (T − T2 ) (72)
industrial implementation, where we make use of the measured F1 cp1 2
variable w = T2 . This allows us to use a much simpler model, which depends on w = T2 but not on the UA-value.
without having to include a model for the heat transfer. For
example, whereas v0 depends on the UA-value, v0,w does not use 5.4.3. Implementation
this information. For both transformed inputs, we will use the pure feedback-
Both transformed inputs are based on steady-state expressions based implementation in Fig. 8(b) with a slave v -controller (Cv ).
for y = T1 and give y = v s at steady state. Thus, both transformed An alternative would be to use an exact model-based inverse
inputs will provide perfect disturbance rejection and linearity at (analytical or numerical) to compute u = F2 plus use a slave
steady state. However, this assumes that the model parameters flow controller to implement F2 (Fig. 8(c)). The use of the slave
do not change and we have that v0 gives an offset if we change v -controller avoids the flow controller, and has an additional ad-
the value of UA, whereas v0,w gives no offset because it uses the vantage in both cases. For v0 , we avoid implementing a numerical
measurement w = T2 instead of a model equation where UA is solution to generate u = f0−1 (v0 , d), and instead we generate the
included. inverse by the slave controller Cv . For v0,w there is no problem
126
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

in generating the exact inverse from (72) to compute u = F2 , Table 1

but with a v -controller we avoid worrying about the potential Nominal operating conditions for the heat exchanger (Example 6) from [16].
internal instability or excessive inputs of using the exact inverse Variable Value Unit
(although this is not a problem in this case). F1 3 kg/s
The transformed input v0,w depends on the measured variable F2 5 kg/s
w = T2 . This feedback will change the dynamics, such that the T1s 297.2 K
dynamics of the transformed system will be different to that of T20 293 K
the original system. On the other hand, the use of v0 has no T10 343 K
feedback from any output (y) or state (w ), and the dynamics will U 150 W/m2 /K
not change (except for the dynamics of the slave loop, which are A 90 m2
negligible in this case because the slave loop for v0 is fast). V1 = V2 0.45 m3
For tuning the slave v0 -controller, we note that the ‘‘process’’ cp1 1200 J/kg K
cp2 1500 J/kg K
(calculation block) from u = F2 to v0 in (70) is static, and a pure
ρ1 980 kg/m3
I-controller is recommended [12]. The tuning of the slave v0,w - ρ2 890 kg/m3
controller is more difficult, because w = T2 depends dynamically
on u = F2 which results in an overshoot in the dynamic response
from u = F2 to v0,w (corresponding to a stable (LHP) zero). For Table 2
Slave v -controller (Cv ) tunings for heat exchanger example.
simplicity, we use I-controllers for both v0 and v0,w , tuned based
on the initial gain, and with the same closed-loop time constant Transformed input KI τC [s]
(τC = 10 s); see Table 2. v0 −0.125 10
v0,w −0.01 10
5.4.4. Simulations
We consider a cooler with data as given in Table 1, and
represented by a dynamic process model with 200 states. The
simulations in Fig. 13 compare the two alternative transformed 6. Discussion
inputs (with the setpoint v s for v0 or v0,w set at the desired
value for y = T1 ). The results are also compared to the open- 6.1. Outer controller C for case with ideal transformed inputs
loop (OL) response with no input transformation (with the input
u = F2 kept constant). The setpoint of the transformed input v s is For the case with ideal transformed inputs (Fig. 9), the trans-
initially at 297 K and changes to 302 K at time t = 167 min. The formed system includes in theory perfect disturbance rejection
simulations show responses to step disturbances in F1 , T10 and T20 and perfect steady-state setpoint tracking. For this reason all
(all measured) and to a step change in the heat transfer parameter simulations in this paper are without the outer controller C , and
UA (unmeasured which introduces uncertainty). to handle setpoint changes in ys we made direct changes in the
From the response for the controlled variable (y = T1 ) in transformed input v . This was implemented simply by setting
Fig. 13(a), we see that there is a clear benefit of using transformed v = ys (or v s = ys for the feedback implementations), since we
inputs. From the theory, both transformed inputs give perfect in most of the examples choose the tuning parameters (B0 = I or
control (y = v ) at steady state for measured disturbances and this B = −A for v0 or vA , respectively) such that we have y = v at
is confirmed by the simulations. However, disturbance rejection steady state.
is not perfect dynamically because the process dynamics are In practice, an outer controller C that adjusts v is needed to
quite complex and not described by a first-order model. For the handle model uncertainty and unmodeled or unmeasured dis-
unmeasured disturbance in UA (towards the end of the simulation
turbances. In many cases, the outer controller is a rather slow
in Fig. 13(a)), we see, as expected, that we get a steady-state offset
PI-controller and the ability to make fast setpoint changes for y
for y = T1 when we use v0 (red curve) as the transformed input,
may be lost. Thus, to retain the direct (and fast) effect from the
but not when we use v0,w (blue curve) which makes use of the
setpoint ys to the transformed input v , we may add feedforward
state measurement w = T2 .
Dynamically, we find for this example that the responses are action from the setpoint ys . A simple solution is to use the
best (fastest) when we use v0 as the transformed input (red following two degrees-of-freedom PI-controller C :
curves). The dynamics with v0 are similar to the quite fast dy- t
∫
namics of the uncontrolled heat exchanger (green curves). This v (t) = v (t0 ) + Ks y (t) − Kc y(t) + KI
s
(ys (t) − y(t))dt (73)
is expected because there is no feedback to v0 which can change t0

the dynamics. On the other hand, when we use v0,w (blue curves), With ideal transformed inputs, we have y = v at steady state
which contains an indirect feedback from w = T2 , the dynamics (with no model error). In such cases, we may choose the ‘‘feedfor-
for the return to the steady state are much slower. The effect ward gain’’ to be Ks = I to get the same initial setpoint response
of this indirect feedback is case dependent and in this case the as for the transformed system without the outer controller C . On
feedback from w slows down the response. It does not help to the other hand, Kc is usually smaller than I. For example, we may
change the tunings of the slave controller (Cv ), because even with choose Kc = 0 to get pure I-control in C . The integral gain KI in
a perfect inverse, the dynamics caused by the feedback from w = C will correct for model error and unmeasured disturbances on a
T2 will be present. It may be possible to use the outer controller
slower time scale. The gain matrices Kc and KI should be diagonal
C to speed up the response, but one must be careful because this
because the ideal transformed system from v to y is decoupled.
may cause instability if there is a measurement delay for y = T1 .
In summary, since the implementation of v0 based on the
full static model is complex and gives steady-state offset for 6.2. Time scale separation for feedback implementation (alternative
disturbances in the UA-value, it is likely that the simpler imple- B)
mentation using v0,w with a measured w = T2 is preferred in
practice. The feedback implementation in Fig. 8(b) (alternative B) gen-
Additional examples, including a steam network, CSTR reactor erates only an approximation of the exact inverse g −1 in Figs. 8(a)
and a pH control problem may be found in the PhD thesis of Zot- and 9, but the error can be neglected if the inner loop (with Cv )
ica [17] and in the Master theses of Bjorvand [18] and Kingstree is sufficiently fast. By ‘‘sufficiently fast’’ we mean that the time
[19]. scale separation τc /τc ,v is sufficiently large. Here, τc ,v denotes the
127
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

Fig. 13. Dynamic simulation of heat exchanger (Example 6) using the cascade implementation of the transformed input v with a controller Cv (Fig. 8(b)). Two choices
for the transformed input, v0 and v0,w (with setpoint v s set equal to the setpoint for y), are compared with the open-loop (OL) case with no input transformation
(constant u). The simulations are with no outer controller (C = 0) for the following step disturbances: F1 from 3 to 4 kg/s at t = 8 min, T20 from 293 to 288 K at
t = 80 min, T10 from 343 to 328 K at t = 117 min, setpoint change in v s = ys from 297 to 302 K at t = 167 min, and disturbance in U from 150 to 100 W/m2 /K at
t = 217 min.

closed-loop time constant of the inner loop involving the slave for the multivariable case, one may insert a filter for the signal
controller Cv , and τc denotes the closed-loop time constant for u that goes to the process (but not on the signal u that goes to
the outer loop involving the master controller C . Note that the the block that computes v in Fig. 8(b), because we do not want
slave controller Cv generates the inverse by iteration, so reaching to unnecessary delay this dependency). For example, a first-order
complete convergence (steady state) will take infinite time. As- filter may be used, F = τ s1+1 where τF is about 5 times larger than
F
suming a linear first-order response, the approach to convergence the closed-loop time constant τc ,v for the slave loop involving Cv .
(or steady state) within the desired overall response time τc is
(1 − eτc /τc ,v ). Thus, the approach to convergence increases from 6.3. Decoupling and disturbance rejection when ideal static trans-
63% to 95.0% to 99.3% as the time constant ratio τc /τc ,v increases formed input v0 is applied to the dynamic system in (18)
from 1 to 3 to 5. Since, convergence (or steady state) for practical
purposes is reached at 99.3%, this gives the rule of thumb of In many cases the process is dynamic, but nevertheless we
requiring a time scale separation between the control layers of may want to apply the ideal static transformed input v0 =
at least 5 [13]. f0 (u, w, d) in (13) or (27), where f0 is the right-hand side of the
If the time scale separation gets too small, typically 3 or less, static process model. Note here that we have here chosen B0 = I
the layers will start interacting and we may experience undesired so that we have y = v0 at steady state.
oscillatory behavior or even instability [20]. A larger value (larger When the transformed input v0 is applied to a dynamic pro-
than 5) allows for robustness to process gain variations which cess, we have that the transformed system is linear, decoupled
and independent of disturbances at steady state. However, dy-
may affect the closed-loop time constants of both the inner and
namically we generally do not know what happens; the response
master control loops. Therefore, a time scale separation of 10
may be nonlinear, coupled and dependent on disturbances. How-
or larger is usually recommended in most practical cases. The
ever, assume that we apply the static transformation v0 =
limiting case of infinite time scale separation corresponds to ϵ =
τc ,v f0 (u, w, ..d) to the particular dynamic system
t
→ 0, which is the singular perturbation condition in the
mathematical literature. dy
= f (u, w, y, d)
The main fundamental limitation in having a fast response in dt
the inner loop (and thus making τc ,v small) is a possible unstable in (18) then we get perfect disturbance rejection and in many cases
zero (or more generally an effective delay) in the dynamic map decoupling, also dynamically, if we make the reasonable assumption
gu from u to v for cases when v depends on the output (state) that we are initially at steady state.. This is an important result,
variables w or y. This is discussed in more detail below. although the class of dynamic systems dt = f (u, w, y, d) in (18) is
dy

However, note that unstable zero dynamics from u to v are not somewhat limited as it only includes low-order dynamic models
a particular problem for the feedback implementation in Fig. 8(b), with as many differential equations as inputs u and outputs y.
but rather it is a fundamental limitation. To avoid excessive To prove that we retain disturbance rejection for this class of
changes (spikes) in the value of u sent to the process, in particular dynamic systems, note that we at steady-state, where f (u, w, y, d)
128
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

Fig. 14. Chain of input transformations.

= 0, have the static relationship y = f0 (u, w, d). Assume now that 0 for the static model case (v0 ) and 1 for the dynamic model case
we apply the ideal static transformed input v0 = f0 (u, w, d) to (vA ). By introducing measured state variables w , we may extend
the dynamic system in (18). Furthermore, assume that we gen- the use of transformed inputs to systems of higher relative order,
erate the exact inverse u = f0−1 (v0 , w, d) and that the resulting but this will require the use of feedback (cascade) control for
transformed dynamic system becomes performing the inversion, which implies that the inversion is not
dy perfect as it depends on the tuning of the slave controller (Cv ).
= ft (v0 , w, y, d) (74) However, if we have a dynamic model for how w depends on
dt
u, it is possible in some cases to get exact inversion for systems
An example of such a transformed system is given by (57) for the of higher relative order by using a chain of input transformations,
nonlinear level process. Since we have y = v0 at steady state, or rather by a chain of inversions on the form (see Fig. 14)
dy
and since dt = 0 at steady state, we must have that ft = 0 when
y = v0 , independent of the value of d (and w ). v2 = g1−1 (v1 , y, d1 ), u = g2−1 (v2 , w, d2 )
Assume now that we are initially at steady-state, so that we
Here, v1 is the transformed input related to the control of y,
have y = v0 and thus ft = 0. Next, consider a disturbance d and
whereas v2 is the transformed input related to the control of w .
assume that we keep the transformed input v0 constant. Then,
since we start from y = v0 and we keep v0 constant, we have Note that g2 is based on a dynamic model for w , for example,
dw
dy
that ft remains at 0 and from (75) we have that dt = 0, and the dt
= f2 (u, w, d2 ) (process 2), whereas g1 may be based on either
a dynamic or static model for y, for example, dt = f1 (w, y, d1 )
dy
system will remain at steady state with y = v0 . In conclusion,
we have perfect dynamic disturbance rejection for the dynamic or y = f0 (w, d1 ) (process 1). With a chain of transformations we
model dt = f (u, w, y, d) in (18) with the use of ideal transformed
dy can get perfect disturbance rejection for all disturbances at steady
variables v0 based on a static model. state. We can also get perfect dynamic disturbance rejection for
However, if we are not initially at steady state then we will not disturbances that enter the model in the same way as the input
have perfect dynamic disturbance rejection with v0 . For example, u, that is, for disturbances d2 with the same relative order as the
if we are in a transition between steady states, due to a change input, but not for disturbances d1 that have a lower relative order
in the transformed input v0 (made by the outer controller), then than u.
an output yi which is not at steady-state, will not be dynamically For example, assume that we have a continuous process with
independent of disturbances. However, if the changes for v0 are two mixing tanks in series and we want to control the outlet
infrequent or on a slow time scale, then for practical purposes we temperature in the second tank (y = T2 ) using the heat input to
will have perfect disturbance rejection when applying v0 to the the first tank (u = Q1 ). Also assume that we have an extra state
dynamic system (18). measurement of the temperature in the first tank (w = T1 ). In this
To prove that we in many cases have decoupling, note the case the relative order from u to y is 2 and perfect disturbance
Eqs. (74) is a set of equations of the form rejection is not possible for disturbances d1 directly affecting y
dyi (e.g., from another inlet stream to tank 2). However, since the
= ft ,i (v0 , w, y, d) (75) relative order from u to w is 1, it is possible, by using a chain of
dt
transformations, to get perfect control for disturbances d2 directly
We make the additional assumption that for each equation, ft ,i = 0 affecting w (e.g., the inlet feed temperature to tank 1). A more
when yi = v0,i . This may not always hold, but it is satisfied detailed treatment of a similar example is given in the Master
for the models studied in this paper. Next, consider a change Thesis of Kingstree [19].
in a single transformed input v0,i with all the other transformed
inputs v0,j (j ̸ = i) constant. Since we start from a steady state with
6.5. Comparison with feedback linearization
yj = v0,j and we keep v0,j constant, we then have that ft ,j = 0 and
dyj
from (75) we have that dt = 0(j ̸ = i). This means that we have The use of ideal transformed inputs vA based on a dynamic
a decoupled response where only yi changes in response to the model is a special case of feedback linearization to systems of
change in v0,i . relative order 1 from u to y. For the scalar case (with one input
In summary, if we initially are at steady state, then v0 achieves u and one output v ) it only allows for one nonlinear differential
disturbance rejection and in many cases decoupling, also dy- equation, dt = f (u, y, d) and it transforms it into a first-order
dy
namically. Thus, the main advantage of using vA (based on the
linear system, dt = Ay + BvA . Nevertheless, we have shown in
dy
dynamic model (18); similar to feedback linearization) rather
than v0 (based on a static model) when applied to the dynamic this paper that this may be very useful for practical applications
system (75) is that vA linearizes the transformed system, also in process control.
dynamically. Since vA depends on y, this gives justification for Compared to the traditional feedback linearization literature,
referring to this approach as ‘‘feedback linearization’’. we have put the main emphasis on the nice properties of vA
related to feedforward control and decoupling. In the feedback
6.4. Chain of input transformations linearization literature, the main emphasis is usually on the lin-
earization effect. In any case, an important advantage of the feed-
The use of the exact inverse u = g −1 (v, w, y, d) is limited to back linearization literature is that it provides a rich theoretical
cases where the relative order from the input u to the output y is basis for introducing the transformed variables vA .
129
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

Feedback linearization allows for considering higher-order Generally, with MPC, the measurement w of a state x can only
system with m nonlinear differential equations (m state variables be exploited if we have a model equation that relates the state
x), and it transforms it into a m’th order chain of m first-order x to the input u. The measurement w is then used to correct the
systems from v to y. Actually, in the feedback linearization lit- prediction by the model, for example, using a Kalman filter. Thus,
erature, it is often assumed to be a chain of linear integrators, we cannot easily substitute model equations by measured states
but for process control applications it is better to generalize and w with MPC.
use a chain of linear first-order systems. However, even with this For the heat exchanger case study, it is not clear how we with
generalization, there are few, if any, reports of using such chains MPC can use the measurement w = T2 to eliminate the heat
in practice. There are some reasons for this. First, the feedback transfer Eq. (68c) (involving the UA-value). At least it cannot be
linearization theory assumes that all the states x are measured, done if MPC is based on an experimental model (which is used
which is often not satisfied in process control applications. Sec- for most MPC applications) because we cannot here treat T2 as a
ond, even if we can measure all the states, the resulting derivation measured disturbance. With transformed inputs, we could do this
and implementation tend to become complicated, and may not be by introducing the transformed input v0w in (72) which depends
worth the effort. on w = T2 .
An alternative approach, which is simpler but somewhat less
general than feedback linearization, is to use a chain of input 6.7. Transformed outputs
transformations, as mentioned above, based on using a model
for the measured state variables w . However, also this imple- It is also possible to define transformed outputs
mentation gets rather complex, and other model-based control
methods, like nonlinear model predictive control (MPC), may be z = h(y, w, d) (76)
better for more complicated dynamic models. where y are the outputs that we want to control at a given
setpoint yS and h is a static function of our choice. When we
6.6. Comparison with nonlinear MPC introduce transformed outputs z, the transformed input vz will
be in terms of z rather than y, that is, vz = gz (u, w, z , d). Two
MPC provides a generic model-based framework for handling
alternative implementations are shown in Fig. 15.
complex control problems that require linearization, decoupling
However, we have already shown that we, by use of trans-
and feedforward control. There are some similarities between
formed inputs v alone, can make the transformed system from v
nonlinear MPC and the use of ideal transformed inputs. Both
to y linear, decoupled and independent of disturbances. There is
methods are based on ‘‘open-loop’’ (feedforward) thinking and
therefore no fundamental advantage of introducing transformed
feedback from the measurements y is only introduced as a cor-
outputs. However, in some cases the use of transformed outputs z
rection. An advantage with MPC is that it includes weights that
may simplify the implementation of the transformed inputs u and
makes it possible to deviate from the exact inverse to address
this may be an important advantage. One example is to introduce
more directly the overall control problem to get a trade-off be-
the enthalpy z = H(T , . . . ) as a transformed output when the
tween output performance, input usage and robustness. Thus,
objective is to control temperature y = T [21]. The reason is
with MPC, we are not forced to use exact model inversion, which
that the model may be formulated much simpler in terms of
is the basis of the input transformations and also of feedback
z = H rather than in terms of y = T , leading to a much simpler
linearization. MPC also allows for a much larger class of dynamic
expression for the ideal transformed input vz = gz (u, w, z , d) for
models. Constraints on inputs and outputs may also be included
z than for the transformed input v = g(u, w, y, d) for y.
more directly with MPC.
On the other hand, the ideal input transformations used in this
6.8. Internal instability with exact inversion
paper have the advantage that they are more explicit and po-
tentially simpler to understand and implement, especially when
For the exact implementation of the transformed input v , we
based on physical static models. They can easily be simplified to
include only decoupling or only parts of the disturbance rejection. must for a given value of v, y, w and d, invert the static map
They can also easily be combined with conventional feedback v = g(u, w, y, d) to generate (analytically or numerically) the
controllers C (e.g., single-loop PID controllers) which may used corresponding value of u
to correct for model error and unmeasured disturbances. Unlike u = g −1 (v, w, y, d) (77)
MPC, input transformations can be easily implemented within
the standard DCS or PLC control platforms used at chemical Note that (77) is a purely static expression and therefore by itself
processing facilities. does not contain any instabilities. However, when generating the
A second more fundamental advantage of input transforma- inverse in (77) we are in effect treating w and y as measured
tions is that we may easily use extra measured variables (states) disturbances, whereas they in reality depend on u. This depen-
w to replace model equations. This is generally more difficult dency may generate internal instability because of the feedback
with MPC, except for cases where w is treated as a measured generated when the static inverse u = g −1 (v, w, y, d) in (77) is
disturbance, which means that we neglect the dependency of w applied to the real dynamic system. We use the term internal
on u. With transformed inputs, we have the possibility of gen- instability because the map from the transformed input v to the
erating the input u by feedback using a cascade implementation output y may appear to be stable, but this may not be true if
(Fig. 8(b)). We then do not need to explicitly know how the we consider the input u, and this ‘‘hidden’’ internal instability
measured state w depends on the input u as this goes into the will eventually appear also in y, either because of model error
response from u to v , which we may obtain experimentally and or because infinite inputs u are not physically realizable.
use to tune the slave feedback controller Cv . Note that we only For the analysis that follows, there is no difference between
need a simple dynamic model (e.g., first-order plus delay) to tune the variables w and y, so to simplify we will for now assume that
Cv and it does not to be exact as the inversion (from v to u) is the transformed input v is independent of y.
handled by feedback. We have the following general result: Let v = gu (u, d) repre-
For example, this makes it possible with transformed inputs sent the dynamic map from u to v when the internal variables w
to replace a valve equation by a flow measurement. This is not in the static map v = g(u, w, d) have been eliminated by using
possible with MPC. a dynamic model w = g w (u, d) for how w depends on u. That is,
130
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

Fig. 15. System with both transformed input v and transformed output z = h(y, w, d).

gu (u, d) = g(u, g w (u, d), d). We then get internal instability when (83) with respect to u to get
we apply the exact inverse u = g −1 (v, w, d) if the map gu (u, d)
4s + 1
contains unstable zero dynamics (RHP-zeros in the linear case). This u= (v − d) (84)
result follows trivially because the inverse will have unstable 4s − 1
poles at the unstable zeros of gu . which as expected is unstable due to the unstable (RHP) pole at
p = 1/4. The response from v to w is also unstable
6.8.1. Example: Simple linear system with unstable zeros −2
As a simple example, consider the system (seemingly static) w= (v − d) (85)
4s − 1
y=u+w+d (78) The two instabilities in Eqs. (84) and (85) cancel each other when
for which we from propose from (27) (with B0 = I) to use as the put into (78) to give y = v . The system from v to y therefore
ideal transformed input the right-hand-side of (78) appears to be stable, but this is not true if we consider the input
u.
v = g(u, w, d) = u + w + d (79)

For implementation, (79) may be solved with respect to u to get 6.8.2. Linear stability analysis of transformed system
the ‘‘inverse input transformation’’ We now want to discuss more generally how we may get in-
stability when applying the exact inverse transformation
u = g −1 (v, w, d) = v − w − d (80)
(Fig. 8(a)). We allow for any transformed input v = g(u, w, y, d),
This may be implemented as in Fig. 8(a) and it gives the (ideal) not only the ones derived systematically from a model as shown
transformed system in Section 4. For simplicity we consider linear systems. Nonlinear
systems may have additional stability problems, but the local
y=v (81)
stability can be assessed from a linear analysis.
which has no dynamics and therefore appears to be stable. So When is the transformed system stable? We know from the
far we have not said anything about how w depends on u. In examples above that it is not sufficient to study the stability from
effect, we have treated w as a measured disturbance and we have v to y because of the possibility for internal instability when the
counteracted the effect of w on v by use of the ‘‘feedforward con- transformed input depends on the dependent (state) variables
troller’’ (inverse) in (80). However, there is a potential ‘‘hidden’’ w. Therefore, to eliminate the issue of internal stability, we will
instability because of the dynamic response from u to w . As an consider the stability of the ‘‘closed-loop’’ map from v to u, which
example, assume that it is first-order with a steady state gain of will be unstable if we have internal instability, e.g., see (84).
−2 By ‘‘closed-loop’’ we mean that we have included the effect of
−2u dw feedback through the dependent variables w and y.
w = g w (s) = or = −0.25(2u + w) (82) For the analysis that follows, there is no difference between
4s + 1 dt
the variables w and y, so to simplify we will for now assume that
Note from (79) that the direct static effect of u on v has a gain of 1,
the transformed input is independent of y
whereas from (82) the indirect dynamic effect of u on v (through
w) has a steady-state gain of −2, that is, it is larger and in the v = g(u, w, d) (86)
opposite direction. The combined effect causes an unstable (RHP)
zero from u to v . To see this, eliminate w from (79) using (82) to Here all the variables may be vectors. The transformation in (86)
get is static so it can be linearized to give
4s − 1 dv = Ku du + Kw dw + Kd dd (87)
v = gu (u, d) = u+d (83)
4s + 1
The linearized process dynamics from u to w can be written
which has an unstable (RHP) zero at z = 1/4. This gives internal
instability if we use the exact inverse in (79). To see this, solve dw = Gw (s)du (88)
131
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

We here set dd = 0 since d does not matter for local (lin- The starting point is often a nonlinear static model, y =
ear) stability considerations. Substituting (88) into (87) gives the f0 (u, d). From (13) with B0 = I, the ideal static transformed
response from u to v input is simply the right-hand side of the model, v0 = f0 , that
is, we have g = f0 where we note that g in the static case
dv = (Ku + Kw Gw (s))du = (I + Lw (s))Ku du (89)
does not depend on the outputs y. For the ideal case, where all
where disturbances d that enter the model are measured and there is no
model error, this gives at steady state the transformed system y =
Lw = Kw Gw (s)Ku−1 (90)
v0 , which is linear, decoupled and independent of disturbances.
Physically, Lw (s) represents the ratio of the indirect effect Kw Gw In practical cases, the ideal transformed inputs v0 may be
(through w ) and the direct effect Ku of u on the transformed input used as a starting point to suggest simpler transformed inputs
v . Solving (89) respect to the input u gives the ‘‘closed-loop’’ input v . In such cases, some of the ideal properties are lost, but the
that achieves the desired perfect control of v transformed input v may still be very useful and greatly sim-
plify the design of the outer controller C , which in any case is
du = Ku−1 (I + Lw (s))−1 dv (91)
needed to handle model uncertainty and unknown or uncertain
Mathematically, we note that the linearized transformed system disturbances.
in (91) corresponds to the closed-loop response from dv to du of a For implementation, we need to invert the transformation g
system with negative feedback and a loop transfer function Lw (s). to generate the physical input u = g −1 (v0 , . . . ). In some cases,
This is very useful because of the existence of stability conditions we may use the exact model-based inverse in Fig. 8(a), but if the
for such systems (e.g. [13]). These stability conditions include the equations are complex, we may use feedback control as a ‘‘trick’’
small gain theorem as well as the more general Bode stability to solve the equations by using the cascade implementation in
condition. Fig. 8(b) with a slave v -controller.
The small gain theorem says that we have stability if the norm The model may often be written in a simpler form, y =
of Lw is less than 1 (e.g., |Lw (jω)| is less than 1 at all frequencies if f0w (u, w, d), by replacing parts of the model by a measured de-
we use the H-infinity norm). For a scalar system, this says that if pendent variable (state) w . The most common is to use a mea-
the indirect effect Kw Gw of u on v (via w ) is less in magnitude than sured flow to replace a valve-like equation, like (6) or (52), or
the direct effect Ku , then the transformed system is internally stable, to use a measured temperature to replace the energy balance or
that is, there are no unstable zeros in the map from u to v . This is parts of it. In such cases, the cascade implementation in Fig. 8(b)
a useful condition which yields significant insight. However, the is often preferred to invert the transformation v0 = f0w (u, w, d).
small gain theorem gives only a sufficient condition for stability, First, it may happen that the function f0w does not depend explic-
but not necessary. The Bode stability condition is tighter because itly on the input u and then it is not possible to use a model-based
it provides a necessary and sufficient conditions for stability. inverse. Second, there is a potential problem with internal insta-
Consider again the example in (79) and (82). From (79) we
bility if we use the model-based inverse, but instability will not
have Ku = 1, Kw = 1, and thus Lw (s) = Gw (s) = − 4s2+1 . The be occur with the cascade implementation because the tuning of
steady-state gain |Lw (0)| = 2 is larger than 1, but from the small the inner controller Cv is based on the experimental response
gain theorem we cannot conclude that the system is unstable. from u to v0 . Internal stability with the exact inverse occurs
However, the Bode condition says that we need to look at the gain if the indirect (dynamic) effect of u on v0 through w is large
at the frequency where the phase shift is −180◦ . Since the steady compared to the direct (static) effect of u on v0 , such that the
state gain of Lw (s) is negative we have that phase shift at steady overall response from u to v0 has unstable zero dynamics (inverse
state is −180◦ . Thus |Lw (jw180 )| = |Lw (0)| = 2 and we conclude response in the linear scalar case). This is not very common, and
from the Bode stability condition that the transformed system is with an ideal transformed input v0 it only happens when the
(internally) unstable, as expected. original system response from the input u to the output y has
unstable zero dynamics (inverse response). In such cases, control
6.8.3. Relationship between internal stability and stable zero dy- of y is fundamentally difficult with any control approach.
namics for output y (for systematic transformed inputs v0 and It is also possible to derive ideal transformed inputs, vA , based
vA ) on a dynamic model, dt = f (u, y, d); see (30). This approach is
dy
We consider here the special case when the ideal transformed closely related to the theory of feedback linearization. At first
input has been derived from a model using (27) or (30). We sight, this seems to be a much more powerful approach than with
can then instead look at the zero dynamics for the output y the static variables v0 , as it gives a transformed system dt = Ay +
dy
(rather than for v ). This follows because y and v are directly
BvA which is linear, decoupled and independent of disturbances,
related in this case. In the static case, we have from (15) that the
also dynamically. However, the benefit is usually small. First, the
transformed system is y = B0 v0 (even dynamically for the case
class of dynamic systems described by dt = f (u, y, d) is limited.
dy
when w depends dynamically on u). Similarly, for the dynamic
For example, for a single-input single-output processes, it allows
case, we see from (21) that vA and y are directly related. In
particular, the zero dynamics will be the same. It then follows, for only one differential equation with no direct effect from the
that to have stability of the transformed system when v is derived input u to the output y (that is, no zeros are allowed). Second, we
using Eqs. (13) or (19) we must require that the map from the have found that the ideal static transformed input v0 performs
physical input u to the output y has stable zero dynamics, that is, almost as well as vA for this class of systems. In particular, it
the process has stable zero dynamics [5]. maintains perfect dynamic disturbance rejection if the system
initially is at steady state. Third, a disadvantage with vA is that
7. Summary and conclusion it is more complex and requires choosing a reasonable value for
the tuning parameter A. Simply setting A = 0, as is normally
In this paper we use the concept of transformed inputs v = recommended in feedback linearization, is normally not a good
g(u, w, y, d) to provide a systematic approach to analyze, imple- choice as the resulting transformed system will be drifting for
ment and derive model-based nonlinear calculation blocks and unknown disturbances. The main advantage with vA compared
cascade control schemes which are frequently used in practice to v0 is that it linearizes the system dynamically, which may
for industrial processes. simplify the design of the outer controller C in some cases. In
132
S. Skogestad, C. Zotică and N. Alsop Journal of Process Control 122 (2023) 113–133

summary, we recommend that the engineer starts with static Thus, with the choice for A in (25), there is no feedback from
models when deriving transformed inputs. y on the transformed input vA at the nominal point.
The use of ideal transformed inputs v may in theory produce For the multivariable case, to get a decoupled response, we
no offset at steady state (y = v ), but this is based on feedforward may choose A equal to the diagonal elements of the A-matrix of
control and assumes an exact model and perfect measurements of the original system
the disturbances. This is not realistic, so we generally need to add
∂f
( )
an outer controller C which manipulates v to control the output A = diag(Ã) = diag (A.3)
y. Single-loop PID-controllers are usually sufficient because the ∂y ∗
response from v to y is linear and decoupled, at least at steady For the multivariable case, this will not exactly keep the original
state. The objective of the outer controller is to correct for er- dynamics and there will be some feedback from y to v at the
rors in the model and measurements and to reject unmeasured nominal point. However, it provides a good comprise between
or unmodeled disturbances. The outer controller should include decoupling and minimizing the feedback from y. In any case, the
integral action to get offset-free control at steady state. exact value for A should not be overemphasized, since we can
We have stressed the need to keep things simple. This is usu- change the closed-loop dynamics by design of the outer controller
ally not an objective in academic papers, but in practice simplicity C.
is important for many reasons. First, it makes it possible to build
a control system of smaller parts (blocks) which may designed References
and tuned independently. Second, it is easier to understand and
modify by engineers and operators, and it reduces errors in the [1] F.G. Shinskey, Controlling Multivariable Processes, Instrument Society of
implementation. America, 1981.
[2] E. Welfonder, Dynamic interactions between power plants and power
systems, Control Eng. Pract. 7 (1999) 27–40, http://dx.doi.org/10.1016/
Declaration of competing interest S0967-0661(98)00146-4.
[3] F.G. Shinskey, Process Control Systems, McGraw-Hill, 1967.
The authors declare that they have no known competing finan- [4] J. Lee, R. Mukherjee, H.K. Khalil, Output feedback performance recovery in
cial interests or personal relationships that could have appeared the presence of uncertainties, Systems Control Lett. 90 (2016) 31–37.
[5] A. Isidori, Nonlinear Control Systems, third ed., in: Communications and
to influence the work reported in this paper. Control Engineering, Springer London, London, 1995.
[6] C. Kravaris, C.-b. Chung, Nonlinear state feedback synthesis by global input
Data availability / output linearization, AIChE J. 33 (1987).
[7] G. Bastin, D. Dochain, On-Line Estimation and Adaptive Control of
Bioreactors, Elsevier Science Publishers B.V., 1990.
No data was used for the research described in the article.
[8] H.K. Khalil, Nonlinear Control, second ed., Pearson, Boston, 2015.
[9] M.A. Henson, D.E. Seborg (Eds.), Nonlinear Process Control, Prentice-Hall,
Acknowledgments 1997.
[10] A. Isidori, personal communication, 2020.
The authors gratefully acknowledge fruitful discussions with [11] C. Zotică, N. Alsop, S. Skogestad, Transformed manipulated variables
for linearization, decoupling and perfect disturbance rejection, IFAC-
Krister Forsman and Christian Holden. This work was partly PapersOnLine 53 (2) (2020) 4052–4057, http://dx.doi.org/10.1016/j.ifacol.
funded by HighEFF (Centre for an Energy Efficient and Competi- 2020.12.2331, 21st IFAC World Congress.
tive Industry for the Future). The authors gratefully acknowledge [12] S. Skogestad, Simple analytic rules for model reduction and PID controller
the financial support from the Research Council of Norway and tuning, J. Process Control 13 (4) (2003) 291–309, http://dx.doi.org/10.1016/
S0959-1524(02)00062-8.
user partners of HighEFF, an 8-year Research Centre under the
[13] S. Skogestad, I. Postlethwaite, Multivariable Feedback Control: Analysis and
FME-scheme (Centre for Environment-friendly Energy Research, Design, second ed., Wiley, 2005.
257632/20). [14] A. Reyes-Lúa, C. Zotică, T. Das, D. Krishnamoorthy, S. Skogestad, Changing
between active constraint regions for optimal operation: classical advanced
Appendix. Tuning parameter A for ideal dynamic transformed control versus Model Predictive Control, Comput.-Aided Chem. Eng. 43
(2018) 1015–1020.
inputs vA [15] N. Soave, M. Barolo, On the effectiveness of heat-exchanger bypass control,
2021, http://dx.doi.org/10.3390/pr9020244, Processes, Vol. 9 URL: https:
Consider the nonlinear system model
dy
dt
= f (u, y, d), which //www.mdpi.com/2227-9717/9/2/244.
can be linearized to get [16] S. Skogestad, Chemical and Energy Process Engineering, CRC Press, 2008.
[17] C. Zotica, Optimal Operation and Control of Thermal Energy Systems
dy (Ph.D. thesis), Norwegian University of Science and Technology, 2023, URL:
= df = Ãdy + B̃du + B̃d dd (A.1) https://folk.ntnu.no/skoge/publications/thesis/2023_zotica/.
dt
[18] S. Bjorvand, Control Structures with Embedded Process Knowledge (Master
where the ∼ variables correspond to the linearized dynamics of thesis), Norwegian University of Science and Technology, 2020, URL:
the original system. We have that à = (∂ f /∂ y)∗ , B̃ = (∂ f /∂ u)∗ https://folk.ntnu.no/skoge/diplom/diplom20/simen-bjorvand/.
and B̃d = (∂ f /∂ d)∗ . where the evaluation of the derivatives is [19] C.R. Kingstree, Application of Transformed Manipulated Variables to Cas-
cade Control Systems for Linearization and Disturbance Rejection (Master
performed at the nominal point of operation, denoted by ∗. Recall thesis), University of Edingborough and Norwegian University of Science
from (24b) that the dynamics of the transformed system are given and Technology, 2021, URL: https://folk.ntnu.no/skoge/diplom/diplom21/
by dt = Ay + BvA . Thus, if we choose A = Ã then the transformed
dy
kingstree/.
system will locally (close to the nominal operating point ∗) have [20] M. Baldea, P. Daoutidis, Control of integrated process networks—A multi-
time scale perspective, Comput. Chem. Eng. 31 (5) (2007) 426–444,
the same dynamics as the original system in (A.1).
http://dx.doi.org/10.1016/j.compchemeng.2006.05.017, URL: https://www.
Furthermore, from (19) the linearized transformed input be- sciencedirect.com/science/article/pii/S009813540600130X, ESCAPE-15.
comes [21] C. Zotica, R. Mocholí Montañés, A. Reyes Lua, S. Skogestad, Control of
steam bottoming cycles using nonlinear input and output transformations
dvA = B−1 (df − Ady) = B−1 (B̃du + B̃d dd) (A.2) for feedforward disturbance rejection, in: The 13th IFAC Symposium on
Dynamics and Control of Process Systems, Including Biosystems (DYCOPS),
and we find that dvA is independent of dy. Busan, Republic of Korea, 2022.

133