Automatica 45 (2009) 2214–2222

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Linear offset-free Model Predictive ControlI

Urban Maeder a,∗ , Francesco Borrelli b , Manfred Morari a
a
Automatic Control Lab, ETH Zurich, CH-8092 Zurich, Switzerland
b
Department of Mechanical Engineering, University of California, Berkeley, 94720-1740, USA

article info abstract

Article history: This work addresses the problem of offset-free Model Predictive Control (MPC) when tracking an
Received 19 March 2008 asymptotically constant reference. In the first part, compact and intuitive conditions for offset-free MPC
Received in revised form control are introduced by using the arguments of the internal model principle. In the second part, we
16 March 2009
study the case where the number of measured variables is larger than the number of tracked variables. The
Accepted 8 June 2009
Available online 13 August 2009
plant model is augmented only by as many states as there are tracked variables, and an algorithm which
guarantees offset-free tracking is presented. In the last part, offset-free tracking properties for special
Keywords:
implementations of MPC schemes are briefly discussed.
Model predictive control © 2009 Elsevier Ltd. All rights reserved.
Reference tracking
No offset
Integral control

1. Introduction type of disturbance model used and the assumptions which guar-
antee offset-free control.
The main concept of Model Predictive Control (MPC) is to This work addresses the problem of offset-free MPC when
use a model of the plant to predict the future evolution of tracking a constant reference and using linear system models. It
the system (Borrelli, Bemporad, Fodor, & Hrovat, 2006; Garcia, represent a thorough study of the main conditions and design
Prett, & Morari, 1989; Hrovat, 1996; Mayne, Rawlings, Rao, & algorithms which guarantee offset-free MPC for a wide range of
Scokaert, 2000; Morari & Lee, 1999). At each time step t a certain practically relevant cases.
performance index is optimized over a sequence of future input We distinguish between the number p of measured outputs, the
moves subject to operating constraints. The first of such optimal number r of outputs which one desires to track (called ‘‘tracked
moves is the control action applied to the plant at time t. At time outputs’’), and the number nd of disturbances. We divide the paper
t +1, a new optimization is solved over a shifted prediction horizon. into three parts. The first part builds on the work in Pannocchia
In order to obtain offset-free control with MPC, the system (2003); Pannocchia and Rawlings (2003) and summarizes, in a
model is augmented with a disturbance model which is used to compact and intuitive manner, the conditions that need to be
estimate and predict the mismatch between measured and pre- satisfied to obtain offset-free MPC by using the arguments of the
dicted outputs. The state and disturbance estimates are used to ini- internal model principle. A simple proof of zero steady-state offset
tialize the MPC problem. The MPC algorithms presented in Badg- is provided when nd = r = p.
well and Muske (2002), Liu and Brosilow (1987), Maciejowski The second part considers the case where the number of mea-
(1998), Magni, De Nicolao, and Scattolin (2001), Meadowcroft, sured variables p is greater than the number of tracked variables r.
Stephanopoulos, and Brosilow (1992), Pannocchia (2003), Pannoc-
In this case, the approaches in Pannocchia and Rawlings (2003) and
chia and Bemporad (2007), Pannocchia and Rawlings (2003) and
Muske and Badgwell (2002) allow one to freely choose the distur-
Zafiriou and Morari (1988) guarantee offset-free control when no
bance models, but they require the number of added disturbance
constraints are active. They differ in the MPC problem setup, the
states nd to be at least equal to the number of measured variables
p, rather than tracked variables r. Our contribution is twofold. First,
a simple algorithm for computing the space spanned by the offset
I This paper was not presented at any IFAC meeting. This paper was is provided when nd < p. Second, we show how to construct a
recommended for publication in revised form by Associate Editor Bart De Schutter controller/observer combination such that zero offset is achieved
under the direction of Editor Ian R. Petersen.
∗ Corresponding address: ETH Zurich, Physikstrasse 3, 8092, Zurich, Switzerland. when the plant model is augmented only by as many additional
Tel.: +41 44 632 55 37; fax: +41 446321211. states as there are tracked variables (nd = r < p), yielding an MPC
E-mail addresses: [email protected] (U. Maeder), with minimal complexity. This effect is particularly useful in the
[email protected] (F. Borrelli), [email protected] (M. Morari). area of explicit Model Predictive Control (Bemporad, Morari, Dua,
0005-1098/$ – see front matter © 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.automatica.2009.06.005
 U. Maeder et al. / Automatica 45 (2009) 2214–2222 2215

& Pistikopoulos, 2000), where the number of parameters affects the Proposition 1 (Badgwell & Muske, 2002; Morari & Stephanopoulos,
complexity of the problem. 1980,a; Pannocchia & Rawlings, 2003). The augmented system (4) is
In the last part, we provide insights on zero steady-state offset observable if and only if (C , A) is observable and
when one/infinity norm objective functions, δ u formulation and  
explicit MPC are implemented. We conclude with two illustrative A−I Bd
has full column rank. (5)
examples. C Cd

2. Preliminaries Proof. From the Hautus observability condition system (4) is ob-
servable iff
In this section, we present the standard MPC design flow: the A T − λI CT
 
0
model choice, the observer design and the controller design. Con- has full row rank ∀λ. (6)
BTd I − λI CdT
sider the discrete-time time-invariant system
Again from the Hautus condition, the first set of rows is linearly
xm (t + 1) = f (xm (t ), u(t )) independent iff (C , A) is observable. The second set of rows is lin-
(
ym (t ) = g (xm (t )) (1) early independent from the first n rows except possibly for λ = 1.
z (t ) = Hym (t ) Thus, for the augmented system the Hautus condition needs to be
with the constraints checked for λ = 1 only, where it becomes (5). 

Exm (t ) + Lu(t ) ≤ M . (2) Remark 1. Note that for condition (5) to be satisfied the number
of disturbances in d needs to be smaller or equal to the number of
In (1)–(2), xm (t ) ∈ R , u(t ) ∈ R and ym (t ) ∈ R are the state,
n m p
available measurements in y, nd 6 p. Condition (5) can be nicely
input, measured output vector, respectively. The controlled vari- interpreted. It requires that the model of the disturbance effect on
ables z (t ) ∈ Rr are a linear combination of the measured variables. the output d → y must not have a zero at (1, 0). Alternatively we
Without any loss of generality we assume H to have full row rank. can look at the steady state of system (4)
The matrices E, L and M define state and input constraints.     
The objective is to design an MPC (Garcia et al., 1989; Mayne A−I Bd x∞ 0
et al., 2000) based on a linear system model of (1) in order to
= (7)
C Cd d∞ y∞
have z (t ) track r (t ), where r (t ) ∈ Rp is the reference signal,
which we assume to converge to a constant, i.e. r (t ) → r∞ as where we have denoted the steady state values with a subscript
t → ∞. Moreover, we require zero steady-state tracking error, ∞ and have omitted the forcing term u for simplicity. We note
i.e., (z (t ) − r (t )) → 0 for t → ∞. that from the observability condition (5) for system (4) Eq. (7) is
required to have a unique solution, which means, that we must
be able to deduce a unique value for the disturbance d∞ from a
2.1. The plant model measurement of y∞ in steady state.

The MPC scheme will make use of the following linear time- The state and disturbance estimator is designed based on the
invariant system model of (1): augmented model as follows:
x̂(t + 1) x̂(t )
      
A Bd B
x(t + 1) = Ax(t ) + Bu(t ) u(t )

= +
(3) d̂(t + 1) 0 I d̂(t ) 0
y(t ) = Cx(t ),
 
Lx
where x(t ) ∈ Rn , u(t ) ∈ Rm and y(t ) ∈ Rp are the state, input + (−ym (t ) + C x̂(t ) + Cd d̂(t )) (8)
Ld
and output vector, respectively. We assume that the pair (A, B) is
controllable, and the pair (C , A) is observable. Furthermore, C is where Lx and Ld are chosen so that the estimator is stable. We
assumed to have full row rank. remark that the results of this paper are independent on the choice
of the method for computing Lx and Ld . We then have the following
2.2. The observer design property.

The plant model (3) is augmented with a disturbance model in Proposition 2. Suppose the observer (8) is stable. Then, rank(Ld ) =
order to capture the mismatch between (1) and (3) in steady nd .
state. Several disturbance models have been presented in the Proof. From (8) it follows
literature (Badgwell & Muske, 2002; Liu & Brosilow, 1987;
x̂(t + 1) x̂(t )
    
Meadowcroft et al., 1992; Pannocchia, 2003; Pannocchia & A + Lx C B d + Lx C d
=
Rawlings, 2003; Zafiriou & Morari, 1988). In this note we d̂(t + 1) Ld C I + Ld Cd d̂(t )
follow (Pannocchia & Rawlings, 2003) and use the form:    
B L
+ u(t ) − x ym (t ). (9)
x(t + 1) = Ax(t ) + Bu(t ) + Bd d(t ) 0 Ld
(
d(t + 1) = d(t ) (4)
y(t ) = Cx(t ) + Cd d(t ) By stability, the observer has no poles at (1, 0) and therefore
 
with d(t ) ∈ Rnd . With abuse of notation we have used the same det
A − I + Lx C Bd + Lx Cd
6= 0. (10)
symbols for state and outputs of system (3) and system (4). Later Ld C Ld Cd
we will focus on specific versions of the model (4).
For Eq. (10) to hold, the last nd rows of the matrix have to be of full
The observer estimates both states and disturbances based on
row rank. A necessary condition is that Ld has full row rank. 
this augmented model. Conditions for the observability of (4) are
given in the following proposition. In the rest of this section, we will focus on the case nd = p.
2216 U. Maeder et al. / Automatica 45 (2009) 2214–2222

Proposition 3. Suppose the observer (8) is stable. Choose nd = p. Note that we distinguish between the current input u(t ) to
The steady state of the observer (8) satisfies: system (3) at time t, and the optimization variables uk in the
     optimization problem (15). Analogously, x(t ) denotes the system
A−I B x̂∞ −Bd d̂∞ state at time t, while the variable xk denotes the predicted state
= (11) at time t + k obtained by starting from the state x0 = x(t ) and
C 0 u∞ ym,∞ − Cd d̂∞
applying to system (3) the input sequence u0 , . . . , uk−1 .
where ym,∞ and u∞ are the steady state measured output and input Let U ∗ (t ) = {u∗0 , . . . , u∗N −1 } be the optimal solution of (15)–(16)
of the system (1), x̂∞ and d̂∞ are state and disturbance estimates from at time t. Then, the first sample of U ∗ (t ) is applied to system (1)
the observer (8) at steady state, respectively.
u(t ) = u∗0 . (18)
Proof. From (8) we note that the disturbance estimate d̂ converges
Denote by c0 (x̂(t ), d̂(t ), yref ) = u∗0 (x̂(t ), d̂(t ), r (t )) the control
only if Ld (−ym,∞ + C x̂∞ + Cd d̂∞ ) = 0. As Ld is square by assumption
law when the estimated state and disturbance are x̂(t ) and
and nonsingular by Proposition 2 this implies that at steady state,
the observer estimates (8) satisfy d̂(t ), respectively. Then the closed loop system obtained by
controlling (1) with the MPC (15)–(16)–(18) and the observer (8)
− ym,∞ + C x̂∞ + Cd d̂∞ = 0. (12) is:

Eq. (11) follows directly from (12) and (8).  x(t + 1) = f (x(t ), c0 (x̂(t ), d̂(t ), r (t )))

Next we particularize the conditions in Proposition 1 to special x̂(t + 1) = (A + Lx C )x̂(t ) + (Bd + Lx Cd )d̂(t )
disturbance classes. The results will be useful later in this paper. (19)
+ Bc0 (x̂(t ), d̂(t ), r (t )) − Lx ym (t )
The following two corollaries follow directly from Proposition 1.
d̂(t + 1) = Ld C x̂(t ) + (I + Ld Cd )d̂(t ) − Ld ym (t ).
Corollary 1. The augmented system (4) with nd = p and Cd = I is
observable if and only if (C , A) is observable and
  3. Number of disturbance states nd equal to number of
A−I Bd measured outputs p
det = det(A − I − Bd C ) 6= 0. (13)
C I
Often in practice, one desires to track all measured outputs
with zero offset. Choosing nd = p = r is thus a natural choice.
Remark 2. We note here clearly how the observability require- Such disturbance models have already been shown to yield offset-
ment restricts the choice of the disturbance model. If the plant has free control (Badgwell & Muske, 2002; Muske & Badgwell, 2002;
no integrators, then det (A − I ) 6= 0 and we can choose Bd = 0. Pannocchia & Rawlings, 2003). Such zero-offset property continues
If the plant has integrators then Bd has to be chosen specifically to to hold if only a subset of the measured outputs are to be tracked,
make det (A − I − Bd C ) 6= 0. i.e., nd = p > r.
Next we provide a very simple proof for offset-free control
2.3. The MPC design when nd = p (Badgwell & Muske, 2002; Muske & Badgwell, 2002;
Pannocchia & Rawlings, 2003).
Denote by z∞ = Hym,∞ and r∞ the tracked measured outputs
and their references at steady state, respectively. For offset-free Theorem 1. Consider the case nd = p. Assume that for r (t ) → r∞
tracking at steady state we want z∞ = r∞ . The observer condition as t → ∞, the MPC problem (15)–(16) is feasible for all t ∈ N+ ,
(11) suggests that at steady state the MPC should satisfy unconstrained for t ≥ j with j ∈ N+ and the closed-loop system (19)
     converges to x̂∞ , d̂∞ , ym,∞ , i.e., x̂(t ) → x̂∞ , d̂(t ) → d̂∞ , ym (t ) →
A−I B x∞ −Bd d̂∞ ym,∞ as t → ∞. Then z (t ) = Hym (t ) → r∞ as t → ∞.
= (14)
HC 0 u∞ r∞ − HCd d̂∞
Proof. Consider the the MPC problem (15)–(16). At steady state
where x∞ is the controller state at steady
h state.
i For x∞ and u∞ to u(t ) → u∞ = c0 (x̂∞ , d̂∞ , r∞ ), x̄t → x̄∞ and ūt → ū∞ . Note that
A−I B
exist for any d̂∞ and r∞ the matrix HC 0
must be of full row the steady state controller input u∞ (computed and implemented)
rank which implies m ≥ r. might be different from the steady state target input ū∞ .
The MPC is designed as follows The asymptotic values x̂∞ , x̄∞ , u∞ and ū∞ satisfy the observer
conditions (11)
N −1
X
kxN − x̄t k2P + kxk − x̄t k2Q + kuk − ūt k2R
    
min A−I B x̂∞ −Bd d̂∞
u0 ,...,uN −1 = (20)
k=0 C 0 u∞ ym,∞ − Cd d̂∞
subj. to Exk + Luk ≤ M , k = 0, . . . , N
xk+1 = Axk + Buk + Bd dk , k = 0, . . . , N (15) and the controller requirement (16)
dk+1 = dk , k = 0, . . . , N     
x0 = x̂(t ) A−I B x̄∞ −Bd d̂∞
= . (21)
d0 = d̂(t ), HC 0 ū∞ r∞ − HCd d̂∞

with ūt and x̄t given by Define δ x = x̂∞ − x̄∞ , δ u = u∞ − ū∞ and the offset  = z∞ − r∞ .
Notice that the steady state target values x̄∞ and ū∞ are both
−Bd d̂(t )
    
A−I B x̄t functions of r∞ and d̂∞ as given by (21). Left multiplying the second
= (16)
HC 0 ūt r (t ) − HCd d̂(t ) row of (20) by H and subtracting (21) from the result, we obtain
and where kxk2M , xT Mx, Q  0, R  0, and P satisfies the Riccati (A − I )δ x + Bδ u = 0
(22)
equation HC δ x = .
P = AT PA − (AT PB)(BT PB + R)−1 (BT PA) + Q . (17) Next we prove that δ x = 0 and thus  = 0.
 U. Maeder et al. / Automatica 45 (2009) 2214–2222 2217

Consider the MPC problem (15)–(16) and the following change 4.1. Steady-state tracking error
of variables δ xk = xk − x̄t , δ uk = uk − ūt . Notice that Hyk − r (t ) =
HCxk + HCd dk − r (t ) = HC δ xk + HC x̄t + HCd dk − r (t ) = HC δ xk The controller (15)–(16) and observer (8) remain unchanged.
from condition (16) with d̂(t ) = dk . Similarly, one can show that However, since nd < p, Proposition 3 is replaced by the following
δ xk+1 = Aδ xk + Bδ uk . Then, the MPC problem (15) becomes: proposition.

N −1
X Proposition 4. The steady state of the observer (8) satisfies
min kδ xN k2P + kδ xk k2Q + kδ uk k2R
Lx ym,∞ − (Bd + Lx Cd )d̂∞
    
δ u0 ,...,δ uN −1 A − I + Lx C B x̂∞
k=0 = (25)
subj. to E δ xk + Lδ uk ≤ M2 , k = 0, . . . , N (23) Ld C 0 u∞ Ld ym,∞ − Ld Cd d̂∞
δ xk+1 = Aδ xk + Bδ uk , k = 0, . . . , N
where ym,∞ and u∞ are the steady state output and input of the sys-
δ x0 = δ x(t ),
δ x(t ) = x̂(t ) − x̄t . tem (1), respectively, x̂∞ and d̂∞ are state and disturbance estimates
from the observer (8) at steady state, respectively.
Denote by KMPC the unconstrained MPC controller (23), i.e., δ u∗0 =
Proof. From (8) at steady state we obtain
KMPC δ x(t ). At steady state δ u∗0 → u∞ − ū∞ = δ u and δ x(t ) →
x̂∞ − x̄∞ = δ x. Therefore, at steady state, δ u = KMPC δ x. From (22) x̂∞ = Ax̂∞ + Bu∞ + Lx (−ym,∞ + C x̂∞ + Bd d̂∞ )
(26)
(A − I + BKMPC )δ x = 0. (24) 0 = Ld (−ym,∞ + C x̂∞ + Cd d̂∞ ).

Since P satisfies the Riccati equation (17), KMPC is a stabilizing Eq. (25) follows directly from (26). 
control law, which implies (A − I + BKMPC ) is nonsingular and hence In the proof of Theorem 1 we have shown that at steady state,
δ x = 0.  the MPC controller (15)–(16) satisfies

Remark 3. Theorem 1 was proven in Pannocchia and Rawlings u∞ − ū∞ = KMPC (x̂∞ − x̄∞ ), (27)
(2003) by using a different approach. where KMPC is the unconstrained MPC (15)–(16). By combining the
Eqs. (16), (25), (27) and using the offset equation  = Hym,∞ − r∞
Remark 4. Theorem 1 can be extended to prove local Lyapunov we find that the observer and controller steady state values satisfy
stability of the closed-loop system (19) under standard regularity the following equation:
assumptions on the state update function f in (19) (Mayne et al., M1 v + M2  = 0 (28)
2000).
with v = T T T T
[x̂∞ u∞ x̄∞ ū∞ d̂∞ yTm,∞ T T
] and
Remark 5. The proof of Theorem 1 assumes only that the models
 
A + Lx C − I B 0 0 Bd + Lx Cd −L x
used for the control design (3) and the observer design (4) are  Ld C 0 0 0 Ld C d −L d 
identical in steady state in the sense that they give rise to the same M1 =  0 0 A−I B Bd 0 
 
relation z = z (u, d, r ). It does not make any assumptions about the  0 0 HC 0 HCd −H 
behavior of the real plant (1), i.e. the model-plant mismatch, with KMPC −I −KMPC I 0 0
the exception that the closed-loop system (19) must converge to
M2 = [0 0 0 I 0]T .
a fixed point. The models used in the controller and the observer
could even be different as long as they satisfy the same steady state The set of all possible offsets  belongs to the projection of the
relation. subspace K , {(v, )|M1 v + M2  = 0} on the offset space, this is
denoted by Π (K ).
Remark 6. If condition (16) does not specify x̄t and ūt uniquely, The following Algorithm 4.1 describes a standard procedure to
it is customary to determine x̄t and ūt through an optimization determine the smallest subspace H = Π (K ) ⊆ Rr spanned by
problem, for example, minimizing the magnitude of ūt subject to all possible offsets  .
the constraint (16) (Pannocchia & Rawlings, 2003). Algorithm 4.1. Step 1. if rank(M1 ) = rank([M1 M2 ]) then
Π (K ) is full dimensional, i.e., H ← Rnd ; end
4. Number of disturbance states nd different from number of Step 2. if rank([M1 M2 (:, 1 : j − 1) M2 (:, j + 1 : nd )]) 6=
measured outputs p rank([M1 M2 ]) for all j = 1, . . . , nd then we have zero
steady state offset and H ← 0.
Step 3. else let U = [u1 , . . . , uk ] span the kernel of M1T . Then
This section deals with the case when the number of measured
H ← {|Z  = 0}, where Z = U T M2 ; end.
variables p is greater than the chosen number of disturbance
states nd . If only a few measured variables are to be tracked
We emphasize that the synthesis of MPC controllers with zero
without offset, choosing nd = p as in the previous section, might
steady state offset is not an easy problem. In fact, Z (and thus H ) is
introduce a possibly large number of disturbance states which in a function of Lx , Ld , H and KMPC . In the next section we provide an
turn increases the complexity of the MPC problem. algorithm for computing Lx , Ld such that zero steady state offset is
From the internal model principle, it is clear that we have to obtained. First we make two important remarks.
add at least one disturbance state for every output to be tracked
without offset. Hence, the number of additional disturbances nd is Remark 7. The case when the p − r measured variables which are
chosen to be equal to the number of tracked outputs, i.e., nd = r < not tracked are discarded from the observer design:
p. This choice, in general, does not guarantee offset-free tracking
x̂(t + 1) x̂(t )
      
A Bd B
when the controller is designed as in Section 3, since Proposition 3 = + u(t )
does not hold. In the following, we first derive a characterization d̂(t + 1) 0 I d̂(t ) 0
of the tracking error at steady state. Then, we present a method for
 
Lx
constructing the observer such that offset-free tracking is obtained. + H (−y(t ) + C x̂(t ) + Cd d̂(t )) (29)
Ld
2218 U. Maeder et al. / Automatica 45 (2009) 2214–2222

falls in the class of systems studied in Section 2.3. Therefore, the Remark 9. A simple choice in (36) is to set Lx = 0 and thus
conditions presented in this section are relevant only if all the H̄ = H. This is clearly equivalent to the case discussed in Remark 7,
measurements y are used for the observer design and r < p. since the number of measurements used by the observer y0m (k) =
Hym (k) = z (k) is equal to the number of disturbances. Next we are
Remark 8. By defining e∞ = ym,∞ − C x̂∞ − Cd d̂∞ , (28) can be interested in the case when Lx 6= 0 since neglecting measurements
rewritten as follows might negatively affect the observer performance.

M̃1 ṽ + M2  = 0 (30) Theorem 2 suggest a direct construction method for the

estimator.
with ṽ = [x̂T∞ uT∞ x̄T ūT d̂T∞ eT∞ ]T and
Algorithm 4.2. Consider the linear system (4) with the definitions

A−I B 0 0 Bd −L x
 (35). Assume (Cm , Am ) detectable. Suppose KMPC is given.
 0 0 0 0 0 −L d  Step 1. Choose Lx such that A + Lx C is stable and (H̄Cm , Ā)
M̃1 =  0 0 A−I B Bd 0 . (31) detectable, where H̄ = H (I − C Φ −1 Lx ) and Ā = Am −
 
 −HC 0 HC 0 0 −H  [LTx 0]T Cm .
KMPC −I −KMPC I 0 0 Step 2. Choose L̄ such that Ā − L̄H̄Cm is stable with L̄T = [L̄Tx L̄Td ]T .
Step 3. Choose the final estimator gain
By using direct substitution, the following equation can be derived    
from Eq. (30) Lx L̄
L= + x H̄ .
    0 L̄d
Ld 0
e = . (32)
H (I − C (I − A − BKMPC )−1 Lx ) ∞ I

Clearly Algorithm 4.1 can be applied to Eq. (32) as well. Remark 10. The construction of the estimator gain L in (36) can be
From Eq. (32) we conclude that zero steady state offset is nicely interpreted when L̄x is equal to the identity matrix. In this
obtained if  = 0 for all e∞ solving (32), i.e., if H (I − C (I − A − case, during the transient Lx is used to generate a state estimation
BKMPC )−1 Lx )∞ = 0 for all e∞ satisfying Ld e∞ = 0. This can be which is based on all measurements. At steady state, the corrective
rewritten in the following null space condition term H̄ cancels the effect of Lx and the steady state estimation relies
only on the tracked measurement (which guarantees zero offset,
N (Ld ) ⊆ N (H (I − C (I − A − BKMPC )−1 Lx )). (33) see Remark 9). In order to demonstrate the last point, consider Eq.
(26). Then, ẑ∞ = HC x̂∞ = HC Φ −1 L̃x ∞ since u∞ = KMPC x̂∞ .
Eq. (33) is the main equation used in Pannocchia and Rawlings Choosing L̃x = Lx + H̄ and H̄ = H (I − C Φ −1 Lx ) we obtain
(2003). ẑ∞ = HC Φ −1 Lx ∞ − HC Φ −1 Lx ∞ + HC Φ −1 H ∞ . Therefore at
steady state ẑ∞ = HC Φ −1 H ∞ .
4.2. Algorithm for offset-free tracking when nd = r < p An alternative way to construct the estimator is described
in the following modified algorithm, which allows one to move
In the following, we propose a method for constructing Lx and Ld the closed-loop poles associated with the disturbance estimates
when nd = r < p such that condition (33) holds. We assume that independently from the observed states modes.
the MPC is defined as in Eqs. (15) and (16), and the unconstrained
Algorithm 4.3. Step 1. Choose Lx such that A − Lx C is stable and
MPC controller gain KMPC is given. We introduce the following
notation for brevity
(H̄Cm , Ā) detectable, where H̄ = H (I + C Φ −1 Lx ) and
Ā = Am − [LTx 0]T Cm .
Φ = I − A − BKMPC , (34) Step 2. Apply the linear transform

−(I − A − Lx C )−1 (Bd + Lx Cd )

 
and I
T =
  0 I
A Bd
Am = , Cm = [C Cd ] . (35) which brings the system to block-diagonal form
0 I
 
A + Lx C 0
Āt = T ĀT −1
= ,
Theorem 2. Consider the augmented system model (4) with nd = r 0 I
and the estimator with a gain of the form C̄t = H̄Cm T −1
    = H̄ [C C (I − A − Lx C )−1 (Bd + Lx Cd ) + Cd ].
L L̄
L = x + x H̄ (36)
0 L̄d Step 3. Choose L̄d such that

where H̄ = H (I − C Φ Lx ). Assume the closed-loop observer I − L̄d H̄ C (I − A − Lx C )−1 (Bd + Lx Cd ) + Cd

−1
 
dynamics Am + LCm is stable. Then, controller (15)–(18) yields offset-
is stable.
free tracking.
Step 4. Compute the estimator gain for the original system
Proof. Substituting (36) into (33) yields    
Lx 0
L= + T −1 .
N L̄d H̄ ⊆ N H (I − C Φ −1 (Lx + L̄x H̄ )) ,



0 L̄d H̄

N L̄d H̄ ⊆ N H̄ − HC Φ −1 L̄x H̄ ,

N L̄d H̄ ⊆ N (I − HC Φ −1 L̄x )H̄ .

Remark 11. If L is computed as in Algorithm 4.3, then the closed-
 are the eigenvalues of A − Lx C and
loop poles of the estimator  the
The last inclusion holds true since L̄d is of full row rank as eigenvalues of I − L̄d H̄ C (I − A − Lx C )−1 (Bd + Lx Cd ) + Cd . They
established in Proposition 2.  can be assigned independently.
 U. Maeder et al. / Automatica 45 (2009) 2214–2222 2219

Remark 12. Algorithms 4.2 and 4.3 require that Lx can be chosen 5.2. Explicit controller
such that (H̄Cm , Ā) is detectable. From Algorithm 4.3, it is clear that
detectability holds if and only if In the last few years there has been growing interest to
apply MPC to systems where the computational resources are
H̄ C (I − A − Lx C )−1 (Bd + Lx Cd ) + Cd
 
(37) insufficient to solve the optimization problem (15) and (16) on-line
is full row rank. We note that this is not a restrictive assumption. in real time. Methods have been developed (Bemporad, Borrelli,
By noting that & Morari, 2002; Borrelli, 2003; Borrelli, Baotic, Bemporad, &
Morari, 2005; Kvasnica, Grieder, & Baotić, 2004) to solve (15)
H̄C = HC Φ −1 (I − A − BKMPC − Lx C ) (38) and (16) explicitly to obtain a state feedback control law u(t ) =
c0 (x̂(t ), d̂(t ), r (t )) in the form of a look-up table. The applicability
we observe two cases where (37) might lose rank:
of these methods is limited by the complexity of the control law
(1) If (I − A − BKMPC − Lx C ) is not full rank and span(HC Φ −1 ) ∩ c0 (·) which is greatly affected by the number of parameters, e.g. the
N ((I − A − BKMPC − Lx C )T ) 6= {0}, then (37) looses rank. Because number of elements in the vectors x̂(t ), d̂(t ) and r (t ).
of the degree of freedom we have in choosing KMPC and Lx , we
Thus it is of interest to examine the proposed control
may safely assume that this rank deficiency
can be avoided. formulations and disturbance models from this perspective.
(2) If span (I − A − Lx C )−1 (Bd + Lx Cd ) + Cd ∩ N (H̄C ) 6= {0},
then (37) looses rank. In this case, either the disturbance model Examining (15) and (16) we note that the control law depends
Bd , Cd or the estimator gain Lx can be modified in order for (37) on x̂(t ), d̂(t ) and r (t ). For instance, nd = p is a popular choice
to have full rank. of disturbance model dimension, since it yields offset-free control
by default, as was seen in Section 3. However, this leads to p + r
Note that Algorithms 4.2 and 4.3 also require that Lx is chosen such
additional parameters. In Section 4 a method was proposed to
that A − Lx C is stable. Conditions on A, C , Cm , H̄, Ā guaranteeing both
obtain offset-free control for models with nd < p and in particular,
stability and a detectability properties are subject of current study.
with minimum order disturbance models nd = r. The total size
Remark 13. In Algorithm 4.3, the matrix H̄ depends on Φ and thus of the parameter vector can thus be reduced to n + 2r. This is
on the controller gain KMPC . Assume that Algorithm 4.3 has been significant only if a small subset of the plant outputs are to be
executed for a given MPC tuning yielding L̄d . If the same MPC is controlled.
redesigned with a different tuning with corresponding KMPC ,1 and A greater reduction of parameters can be achieved by the
H̄1 , then in Step 3 of Algorithm 4.3 one can choose following method. By Corollary 1, we are allowed to choose Bd = 0
in the disturbance model if the plant has no integrators. Recall the

−1
L̄d,1 = (I − Λ) H̄1 C (I − A − Lx C )−1 · (Bd + Lx Cd ) + Cd target conditions (16) with Bd = 0

(39)
    
where A−I B x̄t 0
= . (43)
HC 0 ūt r (t ) − HCd d̂(t )
Λ = I − L̄d H̄ [C (I − A − Lx C )−1 (Bd + Lx Cd ) + Cd ]. (40)
This will ensure identical observer performance regardless of Clearly, any solution to (43) can be parameterized by r (t )−HCd d̂(t ).
the controller used. Note that for L̄d,1 to exist, the detectability The explicit control law is written as u(t ) = c0 (x̂(t ), r (t ) − HCd
condition as discussed in Remark 12 needs to hold. d̂(t )), with n + r parameters. Since the observer is unconstrained,
complexity is much less of an issue. Hence, a full disturbance model
5. Special MPC classes with nd = p can be chosen, yielding offset-free control by default.

5.1. Different norm in the objective function Remark 14. The choice of Bd = 0 might be limiting in practice.
In Vibhor and Borrelli (2008), the authors have shown that for a
If the 2-norm in the objective function of (15) is replaced with a wide range of systems, if Bd = 0 and the observer is designed
1 or ∞ norm (kP (xN − x̄t )kp + k=0 kQ (xk − x̄t )kp + kR(uk −
PN −1
through a Kalman filter, then the closed loop system might suffer
ūt )kp , where p = 1 or p = ∞), then the results of Section 3 a dramatic performance limitation.
continue to hold. In particular, Theorem 1 continues to hold. In fact,
the unconstrained MPC controlled KMPC in (23) is piecewise linear
around the origin (Bemporad, Morari, Dua, & Pistikopoulos, 2002). 5.3. Delta input (δ u) formulation
In particular, around the origin, δ u∗ (t ) , δ u∗0 = KMPC (δ x(t )) is a
continuous piecewise linear function of the state variation δ x: In δ u formulation, the MPC scheme uses of the following linear
time-invariant system model of (1):
KMPC (δ x) = F i δ x if H i δ x ≤ K i , i = 1, . . . , N r (41)
x(t + 1) = Ax(t ) + Bu(t )
(
i i
where H and K in Eq. (41) are the matrices describing the ith
u(t ) = u(t − 1) + δ u(t ) (44)
polyhedron CRi = {δ x ∈ Rn |H i δ x ≤ K i } inside which the
y(t ) = Cx(t ).
feedback optimal control law δ u∗ (t ) has the linear form F i δ x(k).
The polyhedra CRi , i = 1, . . . , N r are a partition of the set of System (44) is controllable if (A, B) is controllable. The δ u
feasible states of problem (15) and they all contain the origin. For formulation often arises naturally in practice when the actuator is
Theorem 1 to hold, it is sufficient to require that all the linear subject to uncertainty, e.g. the exact gain is unknown or is subject
feedback laws F i δ x(k) for i = 1, . . . , N r are stabilizing. For the case to drift. In these cases, it can be advantageous to consider changes
nd = r < p, condition (33) extends to in the control value as input to the plant. The absolute control value
Nr
\ is estimated by the observer, which is expressed as follows
N (Ld ) ⊆ N H (I − C (I − A − BF i )−1 Lx ) .


(42)
x̂(t + 1) x̂(t )
      
A B B
i=1 = + δ u( t )
û(t + 1) 0 I û(t ) I
As a consequence, Algorithms 4.2 and 4.3 cannot be applied  
directly. Therefore, a full disturbance model with nd = p has to Lx
be chosen for 1 and ∞ norms. + (−ym (t ) + C x̂(t )). (45)
Lu
2220 U. Maeder et al. / Automatica 45 (2009) 2214–2222

The MPC problem is readily modified min (u0 − ūt )2 + (x1 − x̄t )2
u0

min kyk − rk k2Q + kδ uk k2R subj. to −1 ≤ u0,t ≤ 1

δ u0 ,...,δ uN −1
x1 = ax̂(t ) +u0,t + d̂(t ) (50)
subj. to Exk + Luk ≤ M , k = 0, . . . , N − 1  
= −d̂(t ) .
a − 1 1 x̄t
xk+1 = Axk + Buk , k ≥ 0
(46) 1 0 ūt 0
yk = Cxk k ≥ 0
uk = uk−1 + δ uk , k ≥ 0 The closed form solution to (50) can be easily computed:
u−1 = û(t )
1

x0 = x̂(t ). 
1, −d̂(t ) − ax̂(t ) > 1,
2


The control input applied to the system is

u? (t ) = −1, 1
−d̂(t ) − ax̂(t ) < 1, (51)
u(t ) = δ u∗0 + u(t − 1). (47) 2

a


−d̂(t ) − x̂(t ) otherwise.


The input estimate û(t ) is not necessarily equal to the actual 2
input u(t ). This scheme inherently achieves offset-free control, Note that KMPC = − 2a .
there is no need to add a disturbance model. To see this, we first Since the number of measured variables equals the number
note that δ u∗0 = 0 in steady-state. Hence, the analysis presented of disturbances used in the model, any stabilizing observer will
in Section 3 applies as the δ u formulation is equivalent to a achieve offset-free control. We choose
disturbance model in steady-state. This is due to the fact that any  
plant/model mismatch is lumped into û(t ). Indeed this approach L=−
a
. (52)
is equivalent to an input disturbance model (Bd = B, Cd = 0). If 1/4
in (46) the measured u(t ) is substituted to its estimate, i.e. u−1 = The dynamics of the controller in the unconstrained case is thus
u(t − 1), then the algorithm would show offset. given by the piecewise affine system
In this formulation the computation of a target input ūt and
Ãc x̃(t ) − Ly(t ) + f , hT x̃(t ) > 1,

state x̄t is not required. A disadvantage of the formulation is that it
is not applicable when there is an excess of manipulated variables x̃(t + 1) = Ãc x̃(t ) − Ly(t ) − f , hT x̃(t ) < 1, (53)
in u compared to measured variables y, since detectability of the
Ãu x̃(t ) − Ly(t ),

otherwise
augmented system (44) is lost.
with
5.4. Minimum-time controller  a  " #
0 0 1
Ãu =  21 , Ãc = 1 , (54)
In minimum-time control, the cost function minimizes the − 1
− 1 4
predicted number of steps necessary to reach a target region, 4
usually the invariant set associated to the unconstrained LQR  
controller (Keerthi & Gilbert, 1987). This scheme can reduce
 
1 d̂(t )
the on-line computation time significantly, especially for explicit f = , h = − 1  (55)
0 ax̂(t )
controllers (Grieder & Morari, 2003). While minimum-time MPC 2
is computed and implemented differently from standard MPC
and x̃(t ) = [x̂(t )T d̂(t )T ]T .
controllers, there is no difference between the two control schemes
One can notice that the unconstrained dynamics Ãu contains
at steady-state. In particular, one can choose the target region to be
the unconstrained region of (15). When the state and disturbance an integrator, while the constrained dynamics Ãc is asymptotically
estimates and reference are within this region, the control law is stable (two poles in 0.5). Hence, when the system saturates, we
switched to (15). The analysis and methods presented in this paper obtain an anti-windup effect.
therefore apply directly.
6.2. Multivariable system
6. Examples
Consider the linearized airplane model discussed in Ma-
In this section, two examples are discussed. The purpose of the ciejowski (1998)
first one is to illustrate the anti-windup effect of the proposed 0

0 1.132 0 −1.0

controller, while the second example shows the application of 0 −0.054 −0.171 0 0.071 
Algorithm 4.3. ẋ(t ) = 0

0 0 1 0  x(t )

0 0.049 0 −.856 −1.01
6.1. Integral action and anti-windup 0 −.291 0 1.053 −.686
 
0 0 0 (56)
Consider the system
 −36 20 0 
x(t + 1) = ax(t ) + u(t ) + d(t ), + 0 0 0  u(t ),
 
(48)
y(t ) = x(t ) −57
1325 0
472 0 −3
with input constraints:
ym (t ) = [I3×3 03×2 ]x(t ).
|u(t )| ≤ 1. (49)
The state vector x = [x1 , . . . , x5 ] comprises altitude, horizontal
The reference value is assumed to be constant and equal to 0. The speed, pitch angle, pitch rate and vertical speed, respectively. The
goal is to design a controller which achieves zero offset, i.e. y(t ) → input variables u1 , u2 and u3 are spoiler deflection, engine thrust
0 as t → ∞. and elevator angle, respectively. The inputs are constrained as
The MPC is formulated as follows follows: −1 ≤ ui (t ) ≤ 1.
 U. Maeder et al. / Automatica 45 (2009) 2214–2222 2221

7. Conclusion

We discussed the problem of offset-free Model Predictive

Control when tracking an asymptotically constant reference. The
system was augmented by additional disturbance states, and a
linear disturbance observer was employed to obtain disturbance
estimates. Simple conditions for zero offset have been derived from
a steady-state analysis of both the estimator and controller. We
first treated the case when the number of disturbances is equal
to the number of measured variables (nd = p), which yields
zero-offset in a straightforward way. This approach may, however,
introduce more disturbance states than there are controlled
variables, and thus lead to more complex MPC problems than
necessary.
Then, the case when nd = r < p was discussed, which does in
general not yield zero offset (Muske & Badgwell, 2002; Pannocchia,
2003; Pannocchia & Rawlings, 2003). We have proposed an
algorithm for computing the observer in such a way that the offset
is removed in selected variables. Thus, the resulting controller has
fewer parameters and is less complex than with previous methods.
Insights were given into the important cases when the perfor-
mance objective is a 1 or ∞ norm, and when the MPC is computed
explicitly to reduce on-line computation time.
We remark that Reference Governor (RG) algorithms (Bem-
porad, 1998; Bemporad, Casavola, & Mosca, 1997; Bemporad &
Mosca, 1994a,b, 1995) provide an alternative, attractive way for
designing tracking controllers for constrained systems. Since RG
make use of a prediction model of the closed loop (plant + con-
troller), the results presented in this work can be easily applied to
RG design.

Acknowledgments

The authors would like to thank the anonymous reviewers

whose careful scrutiny and many useful suggestions considerably
improved the quality of this manuscript.

References

Fig. 1. Reference and disturbance step responses. Badgwell, T. A., & Muske, K. R. (2002). Disturbance model design for linear model
predictive control. In Proceedings of the American control conference, Vol. 2 (pp.
1621–1626).
Bemporad, A. (1998). Reference governor for constrained nonlinear systems.
The continuous-time model (56) is discretized with a sampling
Automatic Control, IEEE Transactions on, 43(3), 415–419.
period of Ts = 0.1. We design an MPC controller in order to Bemporad, A., Borrelli, F., & Morari, M. (2002). Model predictive control based on
track altitude and horizontal speed, hence H = [I2×2 02×3 ], but linear programming — the explicit solution. Automatic Control, IEEE Transactions
on, 47(12), 1974–1985.
zero offset is required on the speed only. According to the results Bemporad, A., Casavola, A., & Mosca, E. (1997). Nonlinear control of constrained
presented in Section 4.2 a disturbance model with nd = 1 is linear systems via predictive reference management. Automatic Control, IEEE
sufficient for obtaining zero steady-state offset. Transactions on, 42(3), 340–349.
Bemporad, A., Morari, M., Dua, V., & Pistikopoulos, E.N. (2000). The explicit solution
The MPC is posed as in (15) and (16) with Q = diag([100 of model predictive control via multiparametric quadratic programming. In
100 100 1 1]), R = 104 · diag([1 1 10]) and a prediction horizon Proceedings of the American control conference.
Bemporad, A., Morari, M., Dua, V., & Pistikopoulos, E. (2002). The explicit linear
of N = 3. The resulting feedback gain depends on n + r + nd = 8 quadratic regulator for constrained systems. Automatica, 38(1), 3–20.
parameters. The estimator Algorithm 4.2 is employed, where the Bemporad, A., & Mosca, E. (1994a). Constraint fulfilment in feedback control via
gains Lx and L̄ are the solutions of the steady-state Kalman filter predictive reference management. In Control Applications, 1994. Proceedings of
the third IEEE conference on, Vol. 3 (pp. 1909–1914).
with unitary weights. Since only one variable is to be controlled Bemporad, A., & Mosca, E. (1994b). Constraint fulfilment in feedback control via
without offset, in Algorithm 4.2 we set H̄ = [0 1 0 0 0](I − predictive reference management. In Control applications, 1994. Proceedings of
C Φ −1 Lx ). the third IEEE conference on, Vol. 3 (pp. 1909–1914).
Bemporad, A., & Mosca, E. (1995) Nonlinear predictive reference governor for
Fig. 1 depicts two simple tests which show offset free control on constrained control systems. In Decision and control, 1995, Proceedings of the
the horizontal speed. In both tests we simulate a model mismatch 34th IEEE conference on, Vol. 2 (pp. 1205–1210).
Borrelli, F. (2003). Constrained optimal control of linear & hybrid systems: Vol. 290.
by doubling the drag coefficient (i.e., multiply by two the element Springer Verlag.
(2, 2) in the matrix A of model (56)) and reducing all the actuator Borrelli, F., Baotic, M., Bemporad, A., & Morari, M. (2005). Dynamic programming
gains by 5%. In the first test (left section of Fig. 1) we simulate a for constrained optimal control of discrete-time hybrid systems. Automatica, 41,
1709–1721.
step change in the reference at time t = 5 s. In the second test Borrelli, F., Bemporad, A., Fodor, M., & Hrovat, D. (2006). An MPC/hybrid system
(right section of Fig. 1) we simulate an additive disturbance at time approach to traction control. IEEE Transactions on Control Systems Technology,
t = 5 s. The disturbance represents a wind gust with horizontal 14(3), 541–552.
Garcia, C., Prett, D., & Morari, M. (1989). Model predictive control: Theory and
and vertical components (headwind and downdraft). practice-a survey. Automatica, 25, 335–348.
2222 U. Maeder et al. / Automatica 45 (2009) 2214–2222

Grieder, P., & Morari, M. (2003). Complexity reduction of receding horizon control. Urban Maeder received his Master degree in electrical
In IEEE conference on decision and control (pp. 3179–3184). engineering from ETH Zurich in 2003. He is currently
Hrovat, D. (1996). MPC-based idle speed control for IC engine. In Proceedings FISITA working towards his Ph.D. degree at the Automatic Control
1996. Laboratory at ETH Zurich. His research interests include
Keerthi, S., & Gilbert, E. (1987). Computation of minimum-time feedback control reference tracking for state-feedback controllers, model
laws for discrete-time systems with state-control constraints. Automatic predictive control, sensor filtering in automotive and
Control, IEEE Transactions on, 32(5), 432–435. aircraft, and collision avoidance systems and algorithms.
Kvasnica, M., Grieder, P., & Baotić, M. (2004). Multi-Parametric Toolbox
(MPT) http://control.ee.ethz.ch/mpt/.
Liu, Y.-C., & Brosilow, C. B. (1987). Simulation of large scale dynamic systems—
I. Modular integration methods. Computers & Chemical Engineering, 11(3),
241–253.
Maciejowski, J. M. (1998). The implicit daisy-chaining property of constrained Francesco Borrelli received the ‘Laurea’ degree in com-
predictive control. Applied Mathematics and Computer Science, 8(4), 695–711. puter science engineering in 1998 from the University of
Magni, L., Nicolao, G. D., & Scattolini, R. (2001). Output feedback and tracking Naples ‘Federico II’, Italy. In 2002 he received the Ph.D.
of nonlinear systems with model predictive control. Automatica, 37(10), from the Automatic Control Laboratory at ETH-Zurich,
1601–1607. Switzerland. He is currently an Assistant Professor at the
Mayne, D. Q., Rawlings, J. B., Rao, C. V., & Scokaert, P. O. M. (2000). Constrained model Department of Mechanical Engineering of the University
predictive control: Stability and optimality. Automatica, 36, 789–814. of California at Berkeley, USA. He is author of the book
Meadowcroft, T. A., Stephanopoulos, G., & Brosilow, C. (1992). The Modular Constrained Optimal Control of Linear and Hybrid Systems
Multivariable Controller: 1: Steady-state properties. AIChE Journal, 38(8), published by Springer Verlag. His research interests in-
1254–1278. clude constrained optimal control, model predictive con-
Morari, M., & Lee, J. H. (1999). Model predictive control: Past, present and future. trol, robust control, parametric programming and automo-
Computers & Chemical Engineering, 23(4–5), 667–682. tive applications of automatic control.
Morari, M., & Stephanopoulos, G. (1980). Minimizing unobservability in inferential
control schemes. International Journal of Control, 31, 367–377.
Morari, M., & Stephanopoulos, G. (1980a). Studies in the synthesis of control Manfred Morari was appointed head of the Department
structures for chemical processes; Part III: Optimal selection of secondary of Information Technology and Electrical Engineering at
measurements within the framework of state estimation in the presence of ETH Zurich in 2009. He was head of the Automatic Control
persistent unknown disturbances. AIChE Journal, 26, 247–260. Laboratory from 1994 to 2008. Before that he was the
Muske, K. R., & Badgwell, T. A. (2002). Disturbance models for offset-free linear McCollum-Corcoran Professor of Chemical Engineering
model predictive control. Journal of Process Control, 12, 617–632. and Executive Officer for Control and Dynamical Systems
Pannocchia, G. (2003). Robust disturbance modeling for model predictive control at the California Institute of Technology. He obtained the
with application to multivariable ill-conditioned processes. Journal of Process diploma from ETH Zurich and the Ph.D. from the University
Control, 13(8), 693–701. of Minnesota, both in chemical engineering. His interests
Pannocchia, G., & Bemporad, A. (2007). Combined design of disturbance model are in hybrid systems and the control of biomedical
and observer for offset-free model predictive control. Automatic Control, IEEE systems. In recognition of his research contributions, he
Transactions on, 52, 1048–1053. received numerous awards, among them the Donald P. Eckman Award and the
Pannocchia, G., & Rawlings, J. B. (2003). Disturbance models for offset-free model John R. Ragazzini Award of the Automatic Control Council, the Allan P. Colburn
predictive control. AIChE Journal, 49(2), 426–437. Award and the Professional Progress Award of the AIChE, the Curtis W. McGraw
Vibhor, B., & Borrelli, F. (2008). On a property of a class of offset-free model Research Award of the ASEE, Doctor Honoris Causa from Babes-Bolyai University,
predictive controllers. In Proceedings of the American control conference. Fellow of IEEE, the IEEE Control Systems Field Award, and was elected to the
Zafiriou, E., & Morari, M. (1988). A general controller synthesis methodology National Academy of Engineering (US). Manfred Morari has held appointments with
based on the IMC structure and the H2 -, H∞ -and µ-optimal control theories. Exxon and ICI plc and serves on the technical advisory boards of several major
Computers & Chemical Engineering, 12(7), 757–765. corporations.