Journal of Process Control 91 (2020) 1–11

Contents lists available at ScienceDirect

Journal of Process Control

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

Multi-input single-output control for extending the operating range:

Generalized split range control using the baton strategy
∗
Adriana Reyes-Lúa, Sigurd Skogestad
Department of Chemical Engineering, Norwegian University of Science and Technology (NTNU), Sem Sælands vei 4, 7491, Trondheim, Norway

article info a b s t r a c t

Article history: Split range control is used to extend the steady-state operating range for a single output (controlled
Received 3 November 2019 variable) by using multiple inputs (manipulated variables). The standard implementation of split range
Received in revised form 1 May 2020 control uses a single controller with a split range block, but this approach has limitations when it comes
Accepted 1 May 2020
to tuning. In this paper, we introduce a generalized split range control structure that overcomes these
Available online xxxx
limitations by using multiple independent controllers with the same setpoint. Undesired switching
Keywords: between the controllers is avoided by using a baton strategy where only one controller is active at a
Split range control time. As an alternative solution we consider model predictive control (MPC), but it requires a detailed
Control structure dynamic model and does not allow for using only one input at a time.
PID © 2020 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license
Tuning
(http://creativecommons.org/licenses/by/4.0/).
Anti-windup
MISO
MPC

1. Introduction valve sequencing [7]. Split range control has been extensively
applied in industry [7,8], but except for basic descriptions and
Classical advanced control extends the single-loop PID- examples of applications (see [3,9–16]), we have not found a
controller to cover more difficult control tasks and includes, systematic design procedure, and there are almost no academic
for example, cascade control, feedforward control, decoupling, studies. Therefore, in a previous paper [17], we proposed a sys-
selectors, split range control, parallel control, and valve posi- tematic procedure to design the standard (classical) split range
tion control (also called input resetting or mid-ranging control) controller in Fig. 1. However, as we explain in Section 2, standard
(e.g. [1–3]). When we need more than one input (ui ) to cover the split range control has limitations in terms of tuning. For example,
steady-state operating range for a single output (y), we can use we must use the same integral time for all inputs, which is
three alternative classical control structures: generally not desirable for dynamic performance.
To allow for independent controller tunings, one alternative is
1. Standard split range control (Fig. 1),
to use multiple controllers with different setpoints (Fig. 2). For
2. One controller for each input, each with a different setpoint
example, when controlling the temperature in a room (y = T ),
for the output (Fig. 2),
one may use ysp,1 = 23 ◦ C as the setpoint for cooling (u1 )
3. Input (valve) position control (Fig. 3).
and ysp,2 = 21 ◦ C as the setpoint for heating (u2 ) [18]. Then,
The strategies in Figs. 1–3 can be used to extend the steady- on hot days, we use cooling (u1 ) and keep the temperature at
state range when the primary input u1 reaches its limit (ulim 1 ). ysp,1 = 23 ◦ C. If we have a disturbance in the outdoor temperature
For example, we may have two sources of heating and we use so that it decreases, say below 20 ◦ C, the controller will reduce
the second most expensive source only when the first one has the cooling until it reaches its lower limit, umin1 = 0, and we
reached its maximum. In other cases, the available inputs have temporarily lose control of the output (y = T ). Eventually, the
opposite effects on the controlled variable; for example, a process room temperature will decrease to ysp,2 = 21 ◦ C and the second
that requires both heating and cooling. In this case, switching controller will start using the heating (u2 ). The use of different
occurs when heating or cooling reach their lower limit of zero. setpoints is to avoid undesired switching between the controllers
Split range control (Fig. 1) has been in use for more than and possible non-uniqueness when using two controllers with
75 years [4,5]. Some other names that have been used for split integral action to control the same output [14].
range control are dual control agent [4], range extending [6] and The third classical control structure for extending the steady-
state range is shown in Fig. 3. Here, the output (y) is always
∗ Corresponding author. controlled with the same input (u1 ), but if u1 approaches its limit
E-mail address: [email protected] (S. Skogestad). (ulim
1 ), then input u2 is activated and keeps u1 away from its limit.

https://doi.org/10.1016/j.jprocont.2020.05.001
0959-1524/© 2020 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
2 A. Reyes-Lúa and S. Skogestad / Journal of Process Control 91 (2020) 1–11

operating range, including model predictive control (MPC). We

conclude the paper in Section 6.

2. Standard split range controller

As shown in Fig. 1, in standard split range control there is one

common controller (C ) which computes the internal signal (v ) to
the split range block (SR-block), which assigns the value (e.g., the
valve opening) for each input (ui ). Importantly, at any particular
Fig. 1. Classical structure 1: Standard implementation of split range control with time, only one input (ui ) is being used to control the output (y),
two inputs (ui ) and one output (y). A typical SR-block is shown in Fig. 4. Note
whereas the remaining inputs are fixed at the values given by ulim ,
that v is a non-physical internal signal, whereas ui is the physical input. ulim
contains information about the maximum and minimum input values, which the
typically at their maximum or minimum values.
SR-block uses to decide on the input switching.
2.1. The split range block

The split range block has also been called characterization

function [16], splitter block [19], and function generator [20]. Fig. 4
depicts a typical split range block for two inputs (u1 and u2 ) for
a case when u1 has a positive effect on the output (y) and u2 has
a negative effect. For example, normally we may want to control
room temperature (y) with heating (u1 ) and with the ventilation
rate (u2 ) set at its maximum. However, on a very cold day we
may reach maximum heating (umax 1 ), and to maintain temperature
control (y), we can reduce the ventilation rate (u2 ).
Fig. 2. Classical structure 2: Two controllers and two inputs for the same output
(y), each controller with a different setpoint (ysp,1 and ysp,2 ). 2.2. Slopes (αi ) in split range block

In Fig. 4(a), the split value is located at the mid-point (v ∗ =

50%) and the slopes have the same magnitude (|α1 | = |α2 |). This
choice is used in most examples in the literature (see [3,13,16,21–
28]). However, each input (ui ) has a different dynamic and static
effect on the output, and the split value v ∗ (or equivalently, the
slopes αi ) should generally be located at some other value to
compensate for this, as illustrated in Fig. 4(b) (see [7,11,17,19,29–
31]). For example, with a PI-controller (with parameters KC and
τI ), we can get the desired controller gain1 for each input (KC ,i ) by
selecting the slopes (αi ) such that αi KC = KC ,i . However, we need
to use a common integral time (τI ) which must be a compromise
Fig. 3. Classical structure 3: Input (valve) position control to control one output
among the desired τI ,i for every ui .
(y) with two inputs (u1 and u2 ).
3. Generalized split range control structure

Normally, input u2 is not used, that is, we have u2 = ulim2 , where

3.1. Proposed baton strategy
typically ulim
2 = 0%. The advantage with input (valve) position
control is that we always use the same input (u1 ) to control The dynamic behavior of each input is generally different and
y, but the disadvantage is that we cannot utilize the full range using a common controller (C (s)), as in standard split range con-
for u1 as we need a back-off (∆u1 ) from the limit. For example, trol (Fig. 1), represents a compromise. Fig. 5 depicts our proposed
we may select to always use cooling (u1 ) to control the room generalized control structure for split range control applications,
temperature (y = T ). On cold days, we use heating (u2 ) to avoid where each input has its own controller (Ci (s)). Here, Ci (s) can be
that the cooling reaches its lower limit (umin = 0). Thus, even any type of controller, but it is commonly a PID controller.2 Each
1
controller produces a suggested input u′i , and the baton strategy
on cold days we will use a little cooling (u1 ). Comment: the term
logic block in Fig. 5 selects and computes the actual inputs (ui ).
‘‘valve position control’’ is more commonly used for another case
In order to use multiple controllers for the same output, we
than in Fig. 3, namely when u1 is used to improve the dynamic
want to make sure that only one input (ui ) is actively controlling
performance of y. In this case u2 is the main manipulated variable
the output (y) at any given time. The other inputs are required to
for steady-state control and u1 is always controlled to its setpoint be at fixed values (umin or umax ), as given in ulim . We propose to
sp i i
(u1 ), which is typically a ‘‘midrange’’ value. do this by using a baton strategy logic, similar to what is used by
In the present paper, we propose a generalized split range runners in a relay (Fig. 6), where only the runner who holds the
control structure (Fig. 5), where the controller for each input can baton is active at any given time, and the active runner decides
be designed independently. To avoid the use of different setpoints when to pass the baton. This avoids the need for a centralized
(Fig. 2), we use a baton strategy, in which undesired switching is supervisor. In other words, we let the active input decide when to
avoided by allowing only one controller to be active at a time. switch to another input. The active input remains active as long as
This paper is organized as follows. In Section 2 we briefly its not saturated (umin
i < ui < umax
i ) and will only pass the baton
describe standard split range control and its limitations with to another input once it becomes saturated (reaches umin or umax ).
i i
respect to tuning. In Section 3 we present the new generalized
structure, which overcomes these limitations. In Section 4 we use 1 The desired tunings for these controllers can be found, for example, from
a case study to illustrate our proposed generalized structure and the SIMC PID tuning rules [32].
compare it to standard split range control. Then, in Section 5 we 2 Having independent controllers (C (s)) allows one to individually tune the
i
discuss possible alternative implementations and control strate- controller for each input i, without any compromise. To design Ci (s) we suggest
gies for multiple-input single-output control for extending the using a systematic tuning procedure, such as the SIMC PID tuning rules [32].
 A. Reyes-Lúa and S. Skogestad / Journal of Process Control 91 (2020) 1–11 3

Fig. 4. Typical split range (SR) block for Fig. 1.

Fig. 5. Proposed generalized split range control using the baton strategy. Each controller computes a suggested input u′i and the baton logic decides on the actual
input ui .

(a) Inputs for which the value of the output (y) increases
when we move ui away from its desired operating
value (fully opened or fully closed).
(b) Inputs for which the value of the output (y) decreases
when we move ui away from its desired operating
value (fully opened of fully closed).

Fig. 6. Baton strategy for relay. D2.3 Within each group, (a) and (b), order the inputs according
to which one should be used first (less expensive) to which
should be used last (more expensive).
3.2. Sequencing of inputs D2.4 In our experience, it is usually helpful to graphically sum-
marize the final sequence in a standard split range block, as
the one in Fig. 4 (and Fig. 8 in the case study), but note that
Before actually designing the baton strategy logic, we need the slopes and the split values have no significance when
to make some initial decisions. First, we need to define the we use the generalized split range control structure that
minimum and maximum values for every input (umin i , umax
i ). This we are proposing.
is decision D1. Then, we need to choose the sequence of use
of the inputs (decision D2). This should be defined considering 3.3. Baton strategy logic
their effect on the output (y) as well as economic aspects. In
some cases, operational aspects may be taken into account. The Once that the sequence of inputs is defined, we can formulate
following steps are used for decision D2: the logic for the baton strategy. Consider that input k is the active
input (has the baton). The proposed baton strategy is then:
D2.1 Define the desired or most economical operating value for
each input (e.g. fully closed or fully open valve). B.1 Controller Ck computes u′k , which is the suggested value for
D2.2 Consider the effect of every input (ui ) on the output (y). the input k.
Then group the inputs into: B.2 If umin
k < u′k < umax
k
4 A. Reyes-Lúa and S. Skogestad / Journal of Process Control 91 (2020) 1–11

(a) keep uk active, with uk ← u′k Table 1

Parameters for Gp,i (s) from ui to y = T and Gd (s) from d = T amb to y = T .
(b) keep the remaining inactive inputs i at the relevant
limit value (umin or umax ). Gi K p ,i τi (min) θi (min)
i i
GAC −5 8 2
B.3 If u′k ≤ umin
k or u′k ≥ umax
k
GCW −10 15 3
GHW 12 10 3
(a) Set uk = umink or uk = umax
k , depending on which
GEH 8 5 1
limit is used, and pass the baton to the new active Gd 1 15 6

input j. The new active input is selected according to

the predefined sequence, depending on which limit Table 2
is met (j = k + 1 or j = k − 1). Tuning parameters in room temperature control.
(b) Set k = j and go to step B.1. ui τc ,i (min) K C ,i τI ,i (min)
uAC θAC −0.4000 8
3.4. Anti-windup strategy uCW θCW −0.2500 15
uHW θHW 0.1389 10
uEH θEH 0.3125 5
One needs to avoid windup for the inputs which are not active.
Thus, when switching, one needs to decide on how to initialize
the new active controllers. There are several alternatives. Since
we only want one controller to be active at a time, the simplest
and most obvious strategy is to set all the states of the non-active
controllers to zero. For a PI controller (Eq. (1)), this means that
the integral action starts at the time of the switching, tb , when
the baton is passed.
t
( ∫ )
1
′
uk (t) = u0k + KC ,k e(t) + e(t) (1)
τI ,k tb

The value of the bias u0k is equal to umin

k or umax
k , depending on from
which side the baton was received. Note that the integration in
Eq. (1) starts from tb and not from 0. This simple strategy is used
in the case study. Alternative anti-windup implementations are Fig. 7. Block diagram for standard split range control for room temperature
described in the discussion (Section 5.4). control. The SR block is shown in Fig. 8.

4. Case study: Control of room temperature

In this section, we demonstrate the implementation and per- Table 1 shows the gains (Kp,i ), time constants (τi ) and de-
formance of our proposed generalized split range control struc- lays (θi ) for Gp,i (s) and Gd (s), modeled as first-order transfer
ture with a temperature control case study. We compare our functions.
proposed generalized control structure with the standard split Note that since the gain for the disturbance in ambient tem-
range control described in Section 2. perature (d = T amb ) is 1 and the inputs (ui ) are scaled in the
We want to control room temperature (y = T ) using four range 0 to 1, the gains Kp,i tell us the disturbance range that
inputs (ui ), two sources of cooling and two sources of heating: each input can handle. For example, since Kp,HW = 12 we can
handle ambient temperatures down to T amb = T0amb − Kp,HW =
• uAC : air conditioning
18 ◦ C-12 ◦ C = 6 ◦ C before we must switch from hot water (HW)
• uCW : cooling water
to electric heating (EH). Furthermore, since Kp,EH = 8 we can
• uHW : hot water (district heating) handle ambient temperatures down to 6 ◦ C-8 ◦ C = −2 ◦ C before
• uEH : electric heating. we lose control of room temperature (y = T ) because both
The setpoint for the room temperature is T sp = 18 ◦ C. The heating sources (HW and EH) are at their maximum. In the other
main disturbance is ambient temperature (d = T amb ), which is direction, we can handle ambient temperatures up to T0amb −
not measured and is nominally the same as the setpoint; thus, Kp,CW − Kp,AC = 18 ◦ C + 5 ◦ C + 10 ◦ C = 33 ◦ C before we lose
T0amb = 18 ◦ C. This means that no heating or cooling is required control of y = T because both cooling sources (AC and CW) are
at the nominal operating point (ui = 0 ∀i), which is desired for at their maximum.
economic reasons. In this example, all four inputs (ui ) are scaled
from 0 to 1. 4.2. Standard implementation of split range control

4.1. Model Fig. 7 shows the block diagram for the standard implementa-
tion of split range control for this process, using one common PI
For simplicity, we model the room as a linear system: controller (C ) and the split range block in Fig. 8. For the common
PI controller we choose KC = 0.0592 and τI = 15 min. Table A.1
y(s) = Gp (s) u(s) + Gd (s) d(s) (2)
in Appendix A.1 summarizes the parameters for the standard split
where: range block in Fig. 8. The details about the design and tuning of
this control structure can be found in [17].
y=T (3a)
u = [uAC uCW uHW uEH ] ⊺
(3b) 4.3. Generalized implementation of split range control
amb
d=T (3c)
Fig. 9 shows the block diagram for the new proposed gen-
Gp (s) = [GAC (s) GCW (s) GHW (s) GEH (s)] (3d) eralized split range control structure. We use PI controllers for
 A. Reyes-Lúa and S. Skogestad / Journal of Process Control 91 (2020) 1–11 5

reach the limit of the range within which we can control y = T . As

there is no other input to pass the baton, they remain the ‘‘active’’
input. In those cases, we lose control of T because all inputs are
constrained.

4.3.3. Simulations
The standard and the generalized split range control schemes
are tested for rejection of disturbances in T amb , which is nominally
18 ◦ C. T sp is kept constant at 18 ◦ C. At t = 10 min, T amb increases
to 20 ◦ C and at t = 80 min to 29 ◦ C. Then, at t = 140 min,
T amb decreases to 24 ◦ C and at t = 180 min to −1 ◦ C. T amb
then increases to 17 ◦ C at t = 280 min, and finally to 22 ◦ C at
t = 350 min.
Fig. 8. Standard split range block for room temperature control with air From Fig. 10, we observe that both the standard and the
conditioning (AC), cooling water (CW), hot water (HW), and electric heating generalized implementation maintain T = T sp at steady-state, but
(EH); SR block in Fig. 7. the generalized structure is better as it reaches steady-state much
faster, except for the disturbances at t = 10 min and t = 140 min
when CW (cooling water) is the active input. This is expected
each input and tune each loop ‘‘tightly’’, according to the SIMC because the integral time for the common controller for standard
tuning rules [32]. This is achieved by selecting the closed-loop split range control is τI = 15 min, which is the same as for
time constant for each input equal to the time delay (τc ,i = θi ). CW with generalized SRC (see Table 2). For the other inputs, the
Table 2 gives the PI tuning parameters for each Ci (s). integral time for generalized SRC is smaller (8, 10, and 5 min),
We next design the generalized split range control structure resulting in a faster return to the setpoint.
according to the procedure in Section 3.
5. Discussion
4.3.1. Sequencing of outputs
5.1. Alternative implementations of generalized split range control
D1 The inputs are normalized, and the operating range for
every input is ui = [0, 1]. In standard split range control, we can use the slopes in the
amb sp split range block to adjust the controller gain for each input, but
D2.1 The most economical operating point is when T =T ,
we have to use the same value for the other controller settings,
and we can have all inputs fully closed (ui = 0).
like the integral or derivative times. By ‘‘generalized split range
D2.2 To maintain T = T sp , we need to cool the room if T amb > control’’ we mean an implementation where the controllers for
T sp , and to heat the room if T amb < T sp . We can group the each input can be tuned independently. Various statements on
inputs according to their effect on the room temperature using independent controllers have appeared in the literature
into: [11,12,33] but we did not find any details on how it should be
(a) Inputs for which y = T increases when we open them implemented or whether it had been used in practice.
(move away from the desired operating condition, During the work with this project, we tried several alternative
fully closed). These are the two heating sources: HW implementations. Our first attempt was to use a common inte-
and EH. grator and put the dynamics after the split range block in Fig. 1.
For example, to change the PI-tunings from the original set 1 (in
(b) Inputs for which y = T decreases when we open
C ) to the set 2, we may add a block KC ,2 /KC ,1 (1 + 1/τ2 s)/(1 +
them (move away from the desired operating condi-
1/τ1 s) on the signal u2 exiting the split range block. However, the
tion, fully closed). These are the two cooling sources:
signal u2 is a physical signal, which already includes its maximum
CW and AC.
or minimum value, and adding dynamics to the signal creates
D2.3 As CW is less expensive than AC, we prioritize the use of non-uniqueness in the switching.
CW over AC for decreasing room temperature. Likewise, we Our next attempt was to have one controller C (s), as in Fig. 1,
prioritize the use of HW over EH. and use different sets of parameters in C (s) based on the output
D2.4 The final sequence can be summarized in the split range from the split range block, which tells which input is active.
block in Fig. 8. However, note that when using the gener- Åström and Wittenmark [33] and Hägglund [11] refer to this
alized control structure the values of the slopes (αi ) have no idea as a special type of gain scheduling. However, the term
significance except for the sign, which determines whether gain scheduling is generally used for the case where the inputs
we start from umin or umax . and outputs are fixed and we change the controller parameters
i i
depending on the operating parameters, for example, the setpoint
(ysp ) or the disturbance (d). On the other hand, split range control
4.3.2. Design of the baton strategy
is used to extend the steady-state range of y by using a sequence
We consider the block diagram in Fig. 9 and use Fig. 8 to define
of different inputs. In any case, we encountered problems with
the sequence and the choice of bias. The proposed baton strategy
implementing this approach. This is because when we change the
logic in steps B.1 to B.3 is written out in detail in Table 3.
controller parameters for C (s), the signal v from C (s) changes,
When an input receives the baton, the integrator of its cor-
which may cause the selector block to change the active input,
responding PI controller (Ck (s)) is reset to zero, according to
resulting in cycling and non-uniqueness in the switching.
Eq. (1). Thus, the initial value for uk (at time t = tb ) will be the We therefore decided to use independent controllers. How-
proportional term plus u0k ever, only one controller should be active at the time, and to
uk (tb ) = u0k + KC ,k e(tb ). select which one, we introduced the baton strategy. The baton
strategy has the advantage that the selection of the active input
Note here that the bias, u0k , is equal to umax
k or umin
k , depending is not centralized. Each active controller only needs to ‘‘know’’ to
on from which side the baton is coming (see Table 3). Note than which two controllers it should give the baton if it reaches its
when u1 = uAC or u4 = uEH reach their corresponding umax i , we maximum or minimum value, respectively.
6 A. Reyes-Lúa and S. Skogestad / Journal of Process Control 91 (2020) 1–11

Fig. 9. Generalized split range control solution for room temperature control.

Table 3
Baton strategy logic for case study.
Value of u′k Active input (input with baton, uk )
u1 = uAC u2 = uCW u3 = uHW u4 = uEH
umin
k < u′k < umax
k Keep u1 active Keep u2 active Keep u3 active Keep u4 active
u1 ← u′1 u1 ← umin
1 u1 ← umin
1 u1 ← umin
2
u2 ← umax
2 u2 ← u′2 u2 ← umin
2 u2 ← umin
1
u3 ← umin
3 u3 ← umin
3 u3 ← u′3 u3 ← umax
3
u4 ← umin
4 u4 ← umin
4 u4 ← umin
4 u4 ← u′4
u′k ≥ umax
k Keep u1 active Baton to u1 Baton to u4 Keep u4 active
(max. cooling) u01 = umin
1 u04 = umin
4 (max. heating)
u′k ≤ umin
k Baton to u2 Baton to u3 Baton to u2 Baton to u3
u02 = umax
2 u03 = umin
3 u02 = umin
2 u03 = umax
3

5.2. Comparison with multiple controllers with different setpoints However, depending on the selected weights, this scheme may
not always bring the output (y) to zero offset and dynamically
As mentioned in the introduction, an alternative to split range it may use more than one input simultaneously, which is not
control is to use multiple controllers with different setpoints. In necessarily the desired strategy.
this case, all controllers are active at any given time (although The simulation in Fig. 11 compares the proposed generalized
some inputs may be saturated), so to avoid undesired switching SRC with MPC. The generalized SRC is the same as the one
and fighting, one has to separate the setpoints. studied earlier (Figs. 9 and 10) and the details of MPC are given
Our new generalized split range controller may be viewed as in Appendix A.2. The MPC weight for setpoint deviation is ten
an extension of this, which avoids the use of different setpoints times higher than the weight for the use of the expensive inputs
(Fig. 2). The use of different setpoints has the advantage of avoid- (u1 = uAC and u4 = uEH ) and the weight for the expensive
ing the logic block in Fig. 5, as the sequence of the inputs is inputs is five times higher than the weight for the less expensive
indirectly given by the value of the setpoints. For example, for inputs (u2 = uCW and u3 = uHW ). In general, we see that MPC
our room temperature case study, we could have used four con- has better initial response, because it uses several inputs at the
trollers with setpoints 20 ◦ C for AC, 19 ◦ C for CW, 18 ◦ C for HW same time, but the settling towards the steady-state is slower
and 17 ◦ C for EH, assuming that we can have tight temperature than with generalized SRC. This is also seen from the values of
control so that a setpoint difference of 1 ◦ C is enough to avoid the integrated absolute error in Table 4. For example, consider
undesired switching. the response at t = 280 min, when d = T amb increases from
−1 ◦ C to 17 ◦ C and the room requires much less heating than
5.3. Comparison of split range control with model predictive control before but still no cooling. Indeed, split range control handles
this disturbance by only limiting the heating. It first turns off
One obvious design approach to handle MISO systems with the electrical heating (uEH ) and then controls the temperature by
input constraints is model predictive control (MPC) [34]. The reducing the hot water (uHW ). MPC also turns off uEH initially, but
standard approach in MPC is to use the weights in the objec- then it starts using the cooling water (uCW ) while at the same time
tive function to assign the priorities for the control objectives. reducing the hot water (uHW ). MPC uses cooling to speed up the
To assure that the controller uses the right input, we need to initial response, but this is not beneficial on a longer time scale
introduce penalties on deviations in the inputs (ui ) from the as seen from the simulations. Moreover, the input usage is also
desired value and the values of the weights should be higher for higher. It is not only dynamically that MPC may use more than
more costly inputs. As there is no systematic way of choosing one input; it also happens at steady state, at least with quadratic
the weights or tuning rules for MPC, we used trial and error. input weights, as in our case study.
 A. Reyes-Lúa and S. Skogestad / Journal of Process Control 91 (2020) 1–11 7

Fig. 10. Comparison of standard and generalized split range controller (SRC) for the case study. The structure for standard SRC (dashed line) is shown in Fig. 7 and
the structure for the generalized SRC (solid line) is shown in Fig. 9.

Note that the sampling time for the MPC is ∆t = 1 min, shows how input tracking with back-calculation can be imple-
whereas split range control is continuous. This partly explains mented for each input (ui ) with the generalized split range con-
why SRC is faster than MPC for the disturbance at t = 280 min. trol structure. In the block diagram in Fig. 12, the tracking con-
The actual performance of MPC will depend on the tuning. Nev- stant, KT , [35] is used to reset the integrator dynamically [14].
ertheless, the main disadvantages with MPC compared to SRC are With KT = 0, tracking is turned off and with a large value for
that it requires a detailed dynamic model and that it will increase KT , tracking is fast. If we implement this anti-windup scheme in
the input cost because it uses several inputs at the same time. For combination with the generalized split range controller proposed
example, as we observed, it may use cooling to avoid a sudden in this paper, all desired inputs (u′i ) are calculated at any time, this
is, we do not reset the integrator of the input that becomes active
temperature increase, although the disturbance could be handled
(receives the baton). Otherwise, the switching logic to transfer the
without cooling.
baton remains the same.
We implemented the back-calculation tracking scheme in
5.4. Anti-windup for generalized split range control Fig. 12, with KT = 1 for all inputs and with the same PI-tunings
(Table 2) and switching logic (Table 3) as before. Fig. 13 com-
In the proposed generalized structure for split range control pares back-calculation (dashed lines) with the strategy of integral
there are multiple controllers for the same output. In the case resetting in Eq. (1) (solid lines). The differences are quite small,
but as expected, we observe a somewhat less aggressive initial
study, windup is overcome by having only one controller active
response to the disturbances when we use back-calculation. For
at any time and resetting the integrator term to zero when a
example, at t = 100 min, uAC does not jump from 0 to 1 as it does
controller becomes active (see Eq. (1)). However, the proportional
with integral resetting. On the other hand, the integrated absolute
and derivative terms of the controller may potentially cause large
error (IAE) with back-calculation is somewhat higher than with
output changes when the switch occurs. This may be partly seen integral resetting, although it is still significantly lower than with
by the value for uAC in Fig. 10 at t = 100 min, which jumps from standard split range control (see Table 4).
0 to 1 for a short time, before settling at about 0.2. Thus, we do
not have bumpless transfer, which actually may be an advantage 5.5. Stability for controllers extending the operating range
because it may give a faster response.
Windup can be avoided by implementing other anti-windup All the structures considered in this paper (Figs. 1–3 and 5)
schemes, such as input tracking with back-calculation [14]. Fig. 12 involve switching between different active controllers. During
8 A. Reyes-Lúa and S. Skogestad / Journal of Process Control 91 (2020) 1–11

Fig. 11. Comparison of standard MPC with generalized split range control for room temperature. The dashed lines correspond to MPC and the solid lines correspond
to the strategy proposed in this paper, which is also depicted in Fig. 10.

Fig. 12. Antiwindup with bumpless transfer: input tracking with back-calculation for input ui used in Fig. 13.

Table 4
Integral absolute error (IAE) for the case study with alternative controllers.
Controller IAE
Standard split range control (Fig. 7): Fig. 10 448.6
Generalized split range control (Fig. 9), with integrator resetting (Eq. (1)): Figures 10, 11, and 13 202.4
Generalized split range control (Fig. 9), with back-calculation tracking (Fig. 12): Fig. 12 235.7
MPC (Eq. (6)): Fig. 11 327.9
 A. Reyes-Lúa and S. Skogestad / Journal of Process Control 91 (2020) 1–11 9

Fig. 13. Comparison of two anti-windup strategies for the generalized split range control structure in Fig. 9. The dashed lines correspond to back-calculation (Fig. 12)
and the solid lines correspond to the strategy of resetting the integrator (Eq. (1)), which is also depicted in Fig. 10.

normal operation, when there is no switching, we achieve robust- or can be perfectly estimated, which is not realistic in most pro-
ness by using the SIMC PI tuning with τc = θ (Table 2), which cess control problems and does not apply to systems with time
guarantees a gain margin of about 3 and a delay margin of about delay as in the case study in this paper. In addition, traditional
2.5 to 3 [32,36]. MPC does not allow for logic variables and therefore does not
In general, switching may result in oscillations, and indeed, we allow for switching such that only one input is used at the time.
encountered such problems with some of the other structures we However, Bemporad and Morari [38] developed an MPC strategy
tried, but not with our proposed structure (see Section 5.1). In which allows for logic variables with closed-loop stability guar-
practice, undesired oscillations may be overcome by introducing antees (again under the assumption that all states are measured),
something on top of the switching, like adding a delay [12]. An- but this assumes the control system has to be designed using
other option is to use a two-step approach with a compensation the approach proposed in the paper, which involves solving a
(auxiliary) loop to avoid undesired switching. For example, Garelli mixed-integer quadratic program (MIQP), for which there is no
et al. [37] propose sliding mode reference conditioning (SMRC), guarantee of convergence to a unique solution.
based on variable structure analysis (VSS) theory and sliding
mode (SM) related concepts, to improve robustness by shaping 6. Conclusions
the reference signal. In this method, the reference signal is shaped
by including a switching block and a first-order low-pass filter Split range control is widely used in industry, but it has not
in the auxiliary loop. This strategy can be implemented to avoid been studied much in academia. In this work, we introduce a
bumpy transfers when switching between different controllers. new generalized control structure using a baton strategy that
There exist no general analysis results for switched systems, for allows for using individual controllers for each available input
say analyzing whether undesired switching will occur and future without a centralized supervisor. The proposed baton strategy
theoretical work in this area will indeed be valuable for the is illustrated in Fig. 5 and Table 3. We demonstrated the feasi-
implementation of switching strategies using advanced PID-based bility of implementing this structure in a case study with four
control structures. available inputs and one controlled variable. This new generalized
Regarding MPC, there exist a number of stability results, al- structure has better dynamic performance than the standard split
though an important assumption is that all states are measured range controller, and also outperforms MPC in our case study.
10 A. Reyes-Lúa and S. Skogestad / Journal of Process Control 91 (2020) 1–11

Table A.1 Here, n = 5, m = 5; the states (x) do not have a physical

Values for αi , ∆vi and ui,0 . meaning and the input uk in Eq. (4) contains the manipulated
AC CW HW EH variables (Eq. (3b)) and the disturbance. In practice, only the room
αi −6.7600 −4.2250 2.3472 5.2813 temperature is measured, but for simplicity we assume that we
∆vi 0.1479 0.2367 0.4260 0.1893 have a perfect estimator so that we can have full state feedback.
u0i 1.0000 1.6250 −0.9028 −4.2813
Once that the system is discretized, the MPC problem can be
formulated as:
N N
∑ )2 ∑
CRediT authorship contribution statement w Tk − Tksp + Qu2
(
min (6a)
k=1 k=1
Adriana Reyes-Lúa: Conceptualization, Formal analysis, Inves-
s.t. discretized model, Eq. (4) and (5) (6b)
tigation, Methodology, Software, Validation, Visualization, Writ-
ing - original draft, Writing - review & editing. Sigurd Skogestad: u min
= 0 ≤ uk ≤ u max
=1 ∀k ∈ {1, . . . , N } (6c)
Conceptualization, Validation, Resources, Data curation, Writing -
review & editing, Supervision, Funding acquisition.
The objective function includes a term for temperature (y = T )
setpoint tracking as well as a term penalizing the use of the
Declaration of competing interest manipulated variables (u). The following weights were selected:
w = 50, Q = diag(5, 1, 1, 5).
The authors declare that they have no known competing finan- The prediction horizon is set to 100 min, and the control
cial interests or personal relationships that could have appeared horizon to 5 min. The problem is solved using the KWIK algo-
to influence the work reported in this paper. rithm [40].

Acknowledgment References

The authors acknowledge Cristina Zotic for her insights for this [1] P.S. Buckley, Techniques of Process Control, first ed., John Wiley & Sons,
paper. Inc., Delaware, 1964.
[2] F. Shinskey, Process Control Systems, third ed., McGraw-Hill, 1988.
Appendix [3] D.E. Seborg, T.F. Edgar, D.A. Mellichamp, Process Dynamics and Control,
second ed., John Wiley & Sons, Inc., 2003.
[4] D. Eckman, Principles of Industrial Control, John Wiley & Sons, New York,
1945, pp. 204–207.
A.1. Parameters for standard split range controller for case study [5] E.D. Fink, Monogram of Instruments and Process Control, Delmar
Publishers, Inc., Albany, NY, 1945, p. 120,149.
Table A.1 summarizes the information that describes the stan- [6] E. Bristol, After DDC idiomatic control, Chem. Eng. Progress 76 (11) (1980)
dard split range block in Fig. 8, where u0i corresponds to the bias, 84–89.
the slopes are αi and ∆vi is the range of the internal variable for [7] B.G. Lipták, Control and on-off valves, in: Instrument Engineers’ Handbook,
second ed., Chilton Book Company, 1985, pp. 410–412.
each input.
[8] B.X. Sun, A. Shah, J. Amalraj, A dual split-range control strategy for pressure
and flow processes, Control Eng. (2015) www.controleng.com/articles/a-
A.2. MPC implementation dual-split-range-control-strategy-for-pressure-and-flow-processes/.
[9] G. Stephanopoulos, Synthesis of control systems for chemical plants A
To implement MPC for the system described in the case study, challenge for creativity, Comput. Chem. Eng. 7 (4) (1983) 331–365, http:
the dynamic optimization problem is set up using Matlab [39]. //dx.doi.org/10.1016/0098-1354(83)80018-0.
The transfer function model, Eq. (2), relating the inputs with the [10] G. Stephanopoulos, Chemical Process Control: An Introduction to Theory
output, is converted to the discrete-time linear time-invariant and Practice, Prentice-Hall, 1984.
[11] T. Hägglund, Praktisk Processreglering, second ed., Studentlitteratur, Lund,
(LTI) system described by:
Sweden, 1997, p. 135.
xk+1 = Axk + Buk (4a) [12] T.E. Marlin, Process Control. Designing Processes and Control Systems for
Dynamic Performance, McGraw Hill, 2000.
yk = Cxk + Duk (4b) [13] B.W. Bequette, Process Control: Modeling, Design, and Simulation,
Prentice-Hall, 2002, p. 800.
where x ∈ R is the state vector, u ∈ R the input vector, y ∈ R1
n m
[14] K.J. Åström, T. Hägglund, Advanced PID Control, ISA, 2006.
the output vector, and A ∈ Rn×n , B ∈ Rn×m , C ∈ R1×n , D ∈ R1×m [15] N.P. Lieberman, Troubleshooting Process Plant Control, Wiley, 2008.
are constant matrices. With a sampling time of ∆t = 1 min: [16] C.L. Smith, Advanced Process Control. Beyond Single Loop Control, John
0.8825 Wiley & Sons, New Jersey, 2010, p. 450.
⎡ ⎤
0 0 0 0
[17] A. Reyes-Lúa, C. Zoticǎ, K. Forsman, S. Skogestad, Systematic design of
⎢ 0 0.9355 0 0 0 ⎥
split range controllers, in: 12th IFAC Symposium on Dynamics and Control
A=⎢ 0 0 0.9048 0 0 (5a)
⎢ ⎥
of Process Systems, Including Biosystems (DYCOPS), IFAC-Papers OnLine,
⎥
⎣ 0 0 0 0.8187 ⎦
Florianópolis, Brazil, 2019.
0 0 0 0 0.9355 [18] A. Reyes-Lúa, S. Skogestad, Multiple-input single-output control for
0.94
⎡ ⎤
0 0 0 0 extending the steady-state operating range—Use of controllers with dif-
⎢ 0 0.9674 0 0 0 ⎥ ferent setpoints, Processes 7 (12) (2019) 941, http://dx.doi.org/10.3390/
B=⎢ 0 0 0.9516 0 0 ⎥ (5b) pr7120941, https://www.mdpi.com/2227-9717/7/12/941.
⎢ ⎥
[19] T.L. Blevins, Control Loop Foundation: Batch and Continuous Processes,
⎣ 0 0 0 0.9063 0 ⎦
International Society of Automation (ISA), 2011.
0 0 0 0 0.2418
[20] C.L. Smith, Control of Batch Processes, John Wiley & Sons, Incorporated,
C = [−0.625 − 0.67 1.20 1.60 0.27] (5c) 2014.
[21] S. Mercer, Cryogenics: A technological tool for space scientists, Cryogenics
D = [0 0 0 0 0] (5d) 8 (2) (1968) 68–78, http://dx.doi.org/10.1016/0011-2275(68)90044-1.
 A. Reyes-Lúa and S. Skogestad / Journal of Process Control 91 (2020) 1–11 11

[22] R.M. Price, P.R. Lyman, C. Georgakis, Throughput manipulation in plantwide [31] P. Gupta, K. Rana, V. Kumar, P. Mishra, Split-range control of a Jacketed
control structures, Ind. Eng. Chem. Res. 33 (5) (1994) 1197–1207, http: CSTR using self-tuning fuzzy PI controller, in: 2015 International Confer-
//dx.doi.org/10.1021/ie00029a016. ence on Advances in Computer Engineering and Applications, IEEE, 2015,
[23] M. Piovoso, K. Dahl, K. Kosanovich, Control of a batch reactor using a pp. 527–533, http://dx.doi.org/10.1109/ICACEA.2015.7164750.
multivariate statistical controller design, IFAC Proc. Vol. 28 (9) (1995) [32] S. Skogestad, Simple analytic rules for model reduction and PID controller
357–361, http://dx.doi.org/10.1016/S1474-6670(17)47063-7. tuning, J. Process Control 13 (4) (2003) 291–309, http://dx.doi.org/10.1016/
[24] S. Wang, Intelligent Buildings and Building Automation, first ed., Spon S0959-1524(02)00062-8.
Press, 2010. [33] K.J. Åström, B. Wittenmark, Adaptive Control, second ed., Addison-Wesley
[25] R.R. Fonseca, J.E. Schmitz, A.M.F. Fileti, F.V. da Silva, A fuzzy–split range Series in Electrical Engineering: Control Engineering, 1995.
[34] S. Qin, T.A. Badgwell, A survey of industrial model predictive control
control system applied to a fermentation process, Bioresour. Technol. 142
technology, Control Eng. Pract. 11 (7) (2003) 733–764, http://dx.doi.org/
(2013) 475–482, http://dx.doi.org/10.1016/j.biortech.2013.05.083.
10.1016/S0967-0661(02)00186-7.
[26] D. Shen-Huii, Z. Gang, H. Mei-Rong, Research on regulator signal segment
[35] MathWorks, Anti-windup control using a PID controller, in: Matlab
match to control valve in split range control system, in: 2011 International
Documentation Center, 2019, https://se.mathworks.com/help/simulink/
Conference on Consumer Electronics, Communications and Networks (CEC-
slref/anti-windup-control-using-a-pid-controller.html;jsessionid=
Net), IEEE, 2011, pp. 4350–4353, http://dx.doi.org/10.1109/CECNET.2011.
0a5ce78fe614b664e892077af362.
5768801. [36] C. Grimholt, S. Skogestad, Optimal PI and PID control of first-order plus
[27] S.T. Fatani, R.S. Patwardhan, M.A.L. Andreu, A novel single-input two- delay processes and evaluation of the original and improved SIMC rules, J.
output (SITO) strategy for split range control, in: 2017 6th International Process Control 70 (2018) 36–46, http://dx.doi.org/10.1016/j.jprocont.2018.
Symposium on Advanced Control of Industrial Processes (AdCONIP), IEEE, 06.011.
2017, pp. 348–353, http://dx.doi.org/10.1109/ADCONIP.2017.7983805. [37] F. Garelli, J. Mantz, Ricardo, H. De Battista, Advanced Control for Con-
[28] A. Bahadori, Control valves, in: Oil and Gas Pipelines and Piping Systems, strained Processes and Systems, The Institution of Engineering and
Elsevier, 2017, pp. 483–571, http://dx.doi.org/10.1016/B978-0-12-803777- Technology. IET Control Engineering Series 75, 2011, p. 224.
5.00016-2. [38] A. Bemporad, M. Morari, Control of systems integrating logic, dynamics,
[29] B. Glemmestad, Optimal Operation of Integrated Processes - Studies on and constraints, Automatica 35 (3) (1999) 407–427, http://dx.doi.org/10.
Heat Recovery Systems, Norwegian University of Science and Technology 1016/S0005-1098(98)00178-2.
(NTNU), 1997. [39] MathWorks, Model predictive control toolbox, in: Design and Simulate
[30] F.B. Bastani, I.-R. Chen, The role of artificial intelligence in fault-tolerant Model Predictive Controllers, 2019, https://se.mathworks.com/help/mpc/
process-control systems, in: Proceedings of the First International Confer- gs/control-of-a-multi-input-single-output-plant.html.
ence on Industrial and Engineering Applications of Artificial Intelligence [40] C. Schmid, L. Biegler, Quadratic programming methods for reduced hessian
and Expert Systems - IEA/AIE ’88, vol. 2, ACM Press, New York, New York, SQP, Comput. Chem. Eng. 18 (9) (1994) 817–832, http://dx.doi.org/10.1016/
USA, 1988, pp. 1049–1058, http://dx.doi.org/10.1145/55674.55724. 0098-1354(94)E0001-4.