energies

Article
Application of Predictive Feedforward Compensator
to Microalgae Production in a Raceway Reactor:
A Simulation Study
Andrzej Pawłowski 1, * ID , José Luis Guzmán 2 ID
, Manuel Berenguel 2 , Francisco G. Acíen 3 ID

and Sebastián Dormido 1 ID

1 Departamento de Informática y Automática, ETSII, UNED, 28040 Madrid, Spain; sdormido@dia.uned.es
2 Departamento de Informática, University of Almería, CIESOL-ceiA3, 04120 Almería, Spain;
joguzman@ual.es (J.L.G.); beren@ual.es (M.B.)
3 Departamento de Ingeniería, University of Almería, CIESOL-ceiA3, 04120 Almería, Spain; facien@ual.es
* Correspondence: a.pawlowski@dia.uned.es; Tel.:+34-9-1398-7147

Received: 24 November 2017; Accepted: 2 January 2018; Published: 4 January 2018

Abstract: In this work, the evaluation of a predictive feedforward compensator is provided in order
to highlight its most important advantages and drawbacks. The analyzed technique has been applied
to microalgae production process in a raceway photobioreactor. The evaluation of the analyzed
disturbance rejection schemes were performed through simulation, considering a nonlinear process
model, whereas all controllers were designed using linear model approximations resulting in a
realistic evaluation scenario. The predictive feedforward disturbance compensator was coupled with
two feedback control techniques, PID (Proportional-Integral-Derivative) and MPC (Model Predictive
Control) that are widely used in industrial practice. Moreover, the classical feedforward approach
has been used for the purpose of comparison. The performance of the tested technique is evaluated
with different indexes that include control performance measurements as well as biomass production
performance. The application of the analyzed compensator to microalgae production process allows
us to improve the average photosynthesis rate about 6% simultaneously reducing the energy usage
about 4%.

Keywords: disturbance compensation; feedforward control; process control; PID; MPC

1. Introduction
Disturbance compensation is a very important aspect that needs to be considered in control system
design for a particular process [1]. In such a case, the control technique should be able to maintain the
controlled variable close to the reference despite the external disturbances that influence the controlled
process. For this reason, most industrial processes need a personalized control scheme to achieve
the performance requirements [2]. From a control system point of view, the process disturbances can
be grouped as unmeasurable and measurable quantities depending on their origin. The measurable
disturbances can be handled by the feedforward structure that can compensate the disturbance before
its effect appears on the controlled variable. However, in many industrial applications, the control
system consists of a feedback controller, since it is able to provide a set point tracking, reduces the
influence of plant-model mismatch as well as compensates for process disturbances [1,3]. Due to this
simple structure, the control system focuses only on one of these issues and provides weak performance
for the other problems [4]. In the classical approach, the feedforward compensator is obtained as
the quotient between the disturbance and the process dynamics. Nevertheless, this structure is
rarely realizable, e.g., due to dead time inversion issues. This problem has been analyzed by several
researchers during the past decades and drives to design more advanced tuning rules for feedforward

Energies 2018, 11, 123; doi:10.3390/en11010123 www.mdpi.com/journal/energies

Energies 2018, 11, 123 2 of 17

schema [5]. Other approach to measurable disturbance compensation consists in the application of
the MPC (Model Predictive Control) technique. Due to simultaneous solution for feedback as well as
feedforward control, this control technique is frequently used in industrial processes [6,7]. Regarding
the feedforward action, the most relevant advantage of the MPC algorithm is related to the prediction
mechanism that can handle future disturbance information [8,9]. Nevertheless, considering the internal
relationship, the MPC feedforward and feedback co-design needs a compromise, where the design is
oriented considering the most relevant aspects [10,11].
Taking into account all these properties, a new predictive feedforward compensator has been
recently proposed in [12], merging advantages of classical MPC-based solutions. In this case,
the compensator can be easily coupled with any feedback controller and its design is independent of
other elements in the control loop. Moreover, it is also able to consider future disturbance estimations
to improve a disturbance attenuation.
The disturbance compensation aspects are also very important in bio-processes, where the control
technique has to reduce the influence of external disturbances [13,14], and this is the case of microalgae
production in raceway reactors. In such a system, the main task of the controller is to compensate for
the influence of external disturbances that affect the photosynthesis rate. The microalgae culture is
influenced by many factors, where the most important variables are related to: solar light viability,
temperature of the medium, dissolved oxygen and medium pH [15]. In this context, a proper
disturbance attenuation could improve the control system accuracy and, as a consequence, the biomass
production performance. The remaining variables of the microalgae cultivation process such as
dissolved oxygen and pH need to be regulated using a proper control system [16]. Both variables,
dissolved oxygen and pH, are characterized by highly changing dynamics (influenced by solar
radiation and photosynthesis rate) and their values should be maintained within optimal values
for each strain. The photosynthesis rate is influenced mostly by solar irradiance, temperature and
pH among others. It is well known that injection of CO2 has a strong influence on the pH level
since it affects the growth medium. Moreover, the usage of CO2 represents important operational
expense for microalgae culture and its supply excess should be avoided [17]. Taking advantage of this
relationship, the control approach uses the pH value to define the amount of CO2 and time instance
for its injection [18,19]. Considering all previously mentioned aspects, the disturbance compensation
scheme can play a decisive role in boosting control performance and improving the rentability in
biomass production process.
In this study, we provide a practical evaluation of a predictive disturbance compensation technique
introduced in [12] in order to highlight the most important advantages and drawbacks of such a
scheme. As was shown in previous works [12,20], the main advantage of such a technique is the
ability to handle the complex dynamics required for perfect disturbance compensation. In the analysis
performed previously, a complete knowledge of the process model was assumed and the future
disturbance signal was known a priori [20]. From the theoretical point of view, these assumptions
allow us demonstrate that predictive feedforward compensator is able to compensate completely the
disturbance before its effect appears on the controlled variable. Nevertheless, this assumption can be
rarely achieved in an industrial control system, where the usage of simplified models is a common
practice. Moreover, the future values of disturbance signals need to be estimated, which also introduces
additional uncertainty resulting in performance degradation. Additionally, in this work, the original
approach based on a GPC (Generalized Predictive Control) algorithm is extended, with a constraints’
handling mechanism in the feedforward controller being a new future and very important aspect
from practical implementation point of view. Moreover, state of the future disturbances required
for predictive mechanism is estimated using a Double Exponential Smoothing (DES) technique that
was proposed in [8] for MPC control techniques where future information on disturbance signal
is required. In the performed study, we use PID and MPC techniques as a feedback controllers
that are coupled with the predictive feedforward compensator to show its influence on the control
performance in two cases. The evaluation is performed on the biomass production process using a
Energies 2018, 11, 123 3 of 17

raceway photobioreactor model developed in [21]. The test bed is built using a first principle raceway
reactor model as a process plant and all controllers are developed using linear model approximations
obtained close to its operating point. This approach considers a plant-model mismatch allowing us
create a realistic control scenario. Taking into account all these features, it is possible to obtain results
similar to those obtained in the real plant. However, the implementation in simulation environments
has the advantage of the repeatability of the external parameters that influence the control system.
In such a way, the changes in production performance depend only on the applied control techniques
and giving measurable changes in biomass production. All the analyzed control systems are evaluated
using control performance as well as biomass production indexes. Additionally, the test bed system is
evaluated in terms of energy usage efficacy.

2. Feedforward Compensators
This section is devoted to introducing the disturbance compensation approaches used in the
performed analysis. First, the classical approach is introduced and combined with a PID controller.
Next, the GPC-based implicit feedback/feedforward structure is described. Afterwards, predictive
feedforward compensator originally introduced in [12] is summarized. Moreover, the previous
approach is extended with constraints handling mechanism and the future disturbance estimation
technique is also provided.
It needs to be highlighted that, in this paper, we analyze the disturbance rejection problem in a
process that is affected by dead time. This issue makes the classical feedforward compensator less
effective and a more complex compensator is required. This is due to dead time inversion issues
(see [20] for more details). Moreover, for simplicity, the process, Pu , and the disturbance dynamics, Pv ,
are represented by a first order system with dead time using the following representation:

Ku Kv
Pu (s) = e−sdu , Pv (s) = e−sdv .
1 + sTu 1 + sTv

2.1. Classical Feedforward Structure for Disturbance Compensation

The block diagram of a classical feedforward disturbance compensation problem is shown in
Figure 1 and consists of a feedback controller C, process Pu , a feedforward controller C f f and process
Pv , which relates a measurable load disturbance v with the process output y. The remaining signals
are: set point w, control signal u (composed of u FF and u FB provided by feedforward and feedback
controllers). In this scheme, the load disturbance is fed through a feedforward compensator whose
output is added to the feedback control action. From the process control point of view, the goal of this
structure consists in a feedforward compensator design in a way that the influence of the measurable
disturbance on the controlled process is minimized.

v (t )
- Cff u (t)
Pv
FF

u(t) y (t )
+
C u (t)
+ Pu +
w(t ) FB

-1

Figure 1. Feedback plus feedforward control scheme.

Energies 2018, 11, 123 4 of 17

Using block algebra, it can be seen that the analytical expression for disturbance compensation is
given by:
y Pv − C f f Pu
= , (1)
v 1 + CPu
and, from this expression, it can be observed that the disturbance can be completely attenuated when
C f f = Pv /Pu . Nevertheless, the resulting compensator is rarely realizable, since it could contain
unstable or non-causal dynamics, and thus other developments should be investigated.
In a first approach, low-order feedforward and feedback controllers are considered. In this case,
we consider a PI or PID controller as the feedback controller, such that:
 
1
C (s) = K 1 + + sTd , (2)
sTi

where K refers to proportional gain, Ti is the integral time and Td is the derivative time (Td = is set
to 0 if a PI controller is used). Note that, in a real implementation, features like filters, anti-windup
and limitations must be added [1]. In addition, we also assume a low-order feedforward compensator
(a lead-lag filter with a dead time), such that:

1 + sB f f −sL
C f f (s) = K f f e ff, (3)
1 + sT f f

and K f f , −1/B f f , −1/T f f and L f f refers to the static gain the zero, the pole and the dead time,
respectively. Note that more complex structures are seldom used [22] in practice.
Taking into account the aforementioned process transfer functions, it is possible to achieve perfect
disturbance compensation by making K f f = Kv /Ku , B f f = Tu , T f f = Tv , L f f = −ρ and ρ = du − dv ,
where du is process time delay between input and output, and dv is time delay disturbance and output.

2.2. Generalized Predictive Controller

This section briefly describes the GPC algorithm with intrinsic feedforward capabilities, that is
when the GPC takes explicitly into account the dynamics of the measurable disturbances [23]. As is
well known, GPC consists of applying a control sequence that minimizes a multistage cost function of
the form:
N Nu
J= ∑ δ( j)[ŷ(t + j|t) − w(t + j)]2 + ∑ λ( j)[∆u(t + j − 1)]2 , (4)
j= N1 j =1

where ŷ(k + j|t) is an optimal system output prediction sequence performed with data known up
to discrete time t, ∆u(t + j − 1) is a future control increment sequence obtained from cost function
minimization with ∆ = (1 − z−1 ), N1 and N are the minimum and maximum prediction horizons,
respectively, Nu is the control horizon and δ( j) and λ( j) are weighting sequences that penalize the
future tracking errors and control efforts, respectively, along the horizons. The horizons and weighting
sequences are design parameters used as tuning knobs. The reference trajectory w(k + j) can be the set
point or a smooth approximation from the current value of the system output y(t) towards the known
reference by means of a determined filter [23]. In Equation (4), the j-step ahead prediction of system
output with data up to time t, ŷ(k + j|t), is computed using the following CARIMA model [23]:

e(t)
A ( z −1 ) y ( t ) = z − d u B ( z −1 ) u ( t − 1 ) + z − d v D ( z −1 ) v ( t ) + , (5)
∆

where v(t) is the measured disturbance at time t, e(t) is a zero mean white noise, A, B and D are
adequate polynomials in the backward shift operator z−1 and du and dv are the dead time in samples
of the process and disturbance dynamics, respectively. Notice that the effect of the measurable
disturbances is represented in Equation (5) by the term z−dv D (z−1 )v(t). Consider a set of N-ahead
 Energies 2018, 11, 123 5 of 17

predictions in Equation (5) where different horizons are used for output prediction and disturbance
estimation, N, and future control action, Nu , respectively [23]. The prediction equation can be
expressed as:
b = Gu + Hv + f;
y f = Gp ∆u(t − 1) + Hp ∆v(t) + Sy(t), (6)
Version December
Version December
22,
Version
2017December
22,
submitted
2017
Version
submitted
22,
toDecember
Energies
Version
2017 tosubmitted
December
Energies
22, 2017 to submitted
Energies
22, 2017 submitted
to Energies to Energies 6 of 20 6 of 20 6 of 20
where Gp , Hp and S are matrices of polynomials that are used to calculate the contribution to future
Version December 22, 2017 submitted to Energies 6 of 20
process outputs of the past and present values of the process output, disturbance and control signal
where upwhere up =
= Δu(t where
− Δu(t
1), vupp−= where
=1), vp =u
Δv(t)
Δu(t −andwhere
pΔv(t)
1), and
=ypvΔu(t
p== −
upy(t) for
Δv(t)
= =vand
yp1),
Δu(t simplicity.
y(t)
p =−yfor
p =simplicity.
1),
Δv(t)
vpy(t)Hence,
and for
yp simplicity.
= Δv(t) the
=Hence,
andsolution
y(t)ypfor
the
=Hence,
simplicity.
solution
of thefor
y(t) the
GPC
of
simplicity.
solution
Hence,
the GPC the
ofHence,
the
solution
GPC th
values (see [23] for more details). Then, Equation (6) becomes:
controllercontroller
when no when
controller
constrains
no constrains
when
arecontroller
considered
no constrains
are controller
when
considered
is obtained
no
areconstrains
when
considered
is obtained
minimizing
no constrains
areisminimizing
considered
obtained
J withare considered
minimizing
respect
is
J obtained u,isJminimizing
withtorespectwhich to u,
obtained
with respect
leads
which
minimizing
Jto:withu, respect
toleadswhich
to:
J with u, to:
leads
to respe
wh
where up = Δu(t − 1), vp = Δv(t) and yp = y(t) for simplicity. Hence, the solution of the GPC
Version December
Version December
22, Version
2017 submitted
22,
December
2017 Version
submitted
to 22,
Energies
December
2017Version
tosubmitted
Energies
22,
December
2017to Energies
submitted
22, 2017tosubmitted
Energies to Energies 6 of 20 6 of 20 6 of 20
controller when no constrains are considered is=obtained u,which
u = K−1u(G y
b= KGu
w −(G u =minimizing
−1 + Hv
G w+
Sy
K −
p−
G
−1
G(Gu
pG pw
Sy
u G+
=
pp
H
J− with
uppG
K respect
+=
v−p(G
−1
G
Sy
u ppu−
wp−
Hv
K ,to−
Sy p −1
G G G
(G 
Hv
Sy
pwuHp −
p
leads
−vpG
G) Sy
H
G
Hvto:
p u−
pv−pp)−
GGG 
GHpHv
pu
(7)
vpp(8)
)−
−G GHvH(8)
p v−
p )G H

(8)
pv

where
where uuppu
where K= ∆u
where
==Δu(t
= K
where
−1
Qλ(uKp−=
t(G
+

where
w)Q
11),
where
G −
,G.vppG
Δu(t
λ + K u
=
G

=
−p=Sy
=
Finally,∆v −
Δv(t)
1), vwhere
p Δu(t
where
Q
G. G=G
)and G
(Finally,
pt+ and
K−
λaccording
u=
Δv(t)
yG.
1),uyQ
where
p =p=where
v−
ppaccording
pto and
= G
Δu(t
y(t)
+=G
Finally,
λpthe
K

= ufor
y(Hv
yreceding
ptto
−
Δv(t) )Q=the
according
pG. =−
for
1), G
y(t)
+ Δu(t 
simplicity.
and for
H
vhorizon
λ Finally,
Gp =
yto
recedingG.p simplicity.
vΔv(t)
simplicity.
p− = Hence,
1), vpand
y(t)
according
the
pFinally,
horizon for
)strategy,
=
recedingthe
Hence, Hence,
simplicity.
yΔv(t)
according
to strategy,
only the =the
phorizon (8)
solution
and
the
y(t)
the
receding
only
first
to for
of
ysolution
solutionHence,
strategy,
the the
pvalue
the simplicity.
= GPC
ofthe
y(t)
horizon
receding
first
only
of of for
the
value
u the
solution
strategy,
is
the simplicity.
Hence,
horizonGPC
of
first isofthe
uonly
value the
strategy,
the Hence
ofsoluti
GPC
first
uonl
is
controller
GPC controller
when
controller
computed, no
when
computed, constrains
when
controller
Δu(t). Hence,no no constrains
are when
considered
constraints
computed,
Δu(t).ifHence,k Δu(t).controller
no
iscomputed,are
constrains
are
kHence,
theiffirst considered
is when
obtained
controller
considered
is row
the
computed, are
no
if kmatrix
first
Δu(t).of row is
isHence,
the constrains
considered
minimizing
ofK
Δu(t). obtained
iswhen
firstobtained
−1 no are
kHence,
if,row
matrix isK
Δu(t) minimizing
constrains
is
−1
theJ obtained
considered
with respect
minimizing
isif k is
,first
of matrix given K
row
Δu(t) are
minimizing
−1
J
by: with
isoffirst
the is
given
, matrix
Δu(t)u,
considered
to
obtained
J respect
which
with
K
by:is
row −1
J respect
ofgiven u,
minimizing
with
isto leads
obtained
, Δu(t)
matrix respect
which
by: to:
to
−1 u,
K is ,givenJ u,
minimizing
leads
to with
Δu(t)by:which
to:
is given by: re
respect
J leads
with
to u,
to
where K which G G.
= Qλ +leads to: Finally, according to the receding horizon strategy, only the first value of u is
computed, Δu(t). Hence, if k isΔu(t)
uthe
u = first
= K
K=−−11
row
(G
(uG=
Δu(t)
kG of

w
wK=matrix
−−1 
−kGG(Gu
Δu(t)
kG
G

Sy
wKw
Sy =−1
Sy −p−
p= −
K,kG
p− −
kG
−1
G
G 
Δu(t)
(G
Sy
Gppis
G
Δu(t)
GkG w

pp− w
Gu−given
uu p=−−
=G
−
pkG
pu
kG K
kGG
G−1
−G

pΔu(t)
G by:

Sy
SyG(G
Hv
kG
pw uu
ppp−
Hv =−
pu

−
−
p−
w
= −
kG
Hv
kG G
G
GK


−1 
−kG


−GHv
H
w
Sy
H
 
(G
kG
p−
pp
p vv−
uSy
pHv
)
)p−
pppkG
p w
G
H−G
,kG−

kG
pv
G
H

kG
Sy

GHv

pG
G
vp−
ppHv Sy
H
pu)−
p u
pp−
pkG
−
G
vp− −
 
G
pkG
G
kG
G

H(8)
G
uvHv
ppH (9)
(8)
ppp
)uvpp−
p−
−kG
kGGG 
(8)
 HHv

(9)
HvHpv −pv−
p
)pkG 
(8
G(9)

H
    
where
where This KΔu(t)== where
Kunconstrained
Q Q
Thisλ=λ+kG + KGG =w
unconstrained G. where
Q
G.
Thiscontrol Finally,
−Finally,
λ kG+ K
G Sy
unconstrained
law=according
G.
control Q
pThis−
of where
according
Finally,
λkG +
thelaw G G
K
unconstrainedto
control
GPC according
ofto
G.
Thisp= uthe
the pwhere
Finally,
Q− receding
considers kG
receding
unconstrained
law
λ GPC to
K
+ control
G
of according
Hvthe
= G.
considers
the horizon
Q receding
−
horizon
knowledge Finally,
GPC
λlaw kG
+ G
control to
of strategy,
H the
knowledge
considers horizon
according
G.p vlaw
strategy,
the
of receding
Finally,
pfuture
GPC only
of strategy,
only
knowledge
of toaccording
horizon
the
considers
the
measurablethe
the
futureGPC first (9)
receding
only
first strategy,
value
measurable
of to
considersthe
value
knowledge
future the
disturbances first
horizon
of receding
of
u only
value
is
uofisthe
measurable
knowledge strategy,
disturbances ofhorizon
future first
u isdisturbances
value
of only strategy,
measurable
future the
of ufir mi
∆u − 1 ∆u
computed,
computed,and references. computed,
Δu(t).( t ) .
and references. Hence,
Hence,
Notice computed,
Δu(t). if
if
and references.thatkkHence,
is
is
Notice the
the
an andΔu(t). first
optimization
thatNoticek
computed,
first
if is
row
Hence,
row
references. the
an optimizationand ofoffirst
that matrix
Δu(t).
problem k
computed,
matrix
if row
references.
an
Notice is K
the
Hence,
ofK
optimization
−1
matrix
problem
must first ,
Δu(t).
thatNotice,
Δu(t) ifrow
be k
an mustK
( t )−1
Hence,
is
solvedis
of is
optimization
problem ,
thegiven
matrix
thatbean given
Δu(t) first if
using solved k
by: Kis
rowis
by:
optimization
must
−1
given
the,
of
aproblem Δu(t)
quadratic
be
using first
matrix by:
solved row
amustis K
problem given
quadratic
cost
−1
of
using ,
matrix
befunction, by:
Δu(t)
solved
amust cost K is
quadratic
−1
be given
using ,
function, Δu(t)
solved by:
cost
a quadraticis
using given
function, a qcb
This unconstrained control law of the GPC considers knowledge of future measurable disturbances
when constraints when constraints are when taken constraints
are into
 taken when
account into
areconstraints
 takensince
account
whenthere intoconstraints
since
are account
istaken nothere explicitinto
are
sinceis
 taken noaccountthere
solution
explicit into issince
no account
solution
[24]. explicit
there For since[24].
isfuture
solution
nothere For
explicit
analysis future
[24].
is nosolution explicit
For
analysis
the
 future [24].solution
theanalysis
For [24]. futurethe
and references. Notice that Δu(t) an
∆u(optimizationt) = kG Δu(t) wproblem
=− kGΔu(t) wSy must −
p= −kG be
kG 
solved
Sy wp p−uppkG
Δu(t)
G using
= kG
− G
Sy appquadratic
Δu(t) uHvw −
p− kG=−kG kG G 
cost
HvSy
pu w
HHppp− function,
− v pkG kG . Sy HGpppvu−
Hv −
pp − kG
kG kG 
(9) GH(9) Hv
ppuvp − −(9) kGHv
kG Hp v− p (9 k
Version December unconstrained
22, unconstrained
Version
2017 control
submitted
December unconstrained
lawcontrol
to (equation
22,
Energies law
2017 unconstrained
control
(equation
9) is used.
submitted law
unconstrained
9)(equation
tocontrol
InisEnergies
consequence,
used.law 9)
In control (equation
is
consequence,
Version used. the
law Indifferent
December9)
(equation
consequence,
isthe used. different
terms
22, 9)In62017
isconsequence,
in
of the
used. (9)
terms
20 different
submittedare
In in consequence,
represented
(9)theterms
toaredifferentrepresented
in6 (9)
Energies of the are
20 terms
different
represented
in (9) termare
when constraints are taken into account since there is no explicit solution [24]. For future analysis the
on December 22, 2017 This as unconstrained
polynomials
submitted
This This
unconstrained unconstrained
astopolynomials
Energiesin control
theas This inunconstrained
polynomials
operator law
the control
of
operator
zasthat the
This
inlaw GPC
polynomials unconstrained
zcontrol
multiply
the of
operator
that
as considers
the This inGPC
polynomialslaw
multiply
considered zunconstrained
the of
that considers
knowledge
control
operator the
considered
multiply
in
signals GPC
the law knowledge
zoperator
that control
considers
of
(w(t), of
considered
signals
6in future
the
multiply
of law
z (w(t),
y(t),
20 GPC
thatmeasurable
of
knowledge
of
signalsfuture
considers
considered
Δu multiply the
and
y(t), (w(t),GPC
measurable
Δv(t)).
Δu of
disturbances
knowledge
considered
signals
and future
considers
y(t), Thus,
Δv(t)). measurable
disturbances
(w(t),
Δu signals
(9)
and knowledge
of
Thus,
y(t), future
(w(t),
Δv(t)). (9)
Δu disturbance
measurabl
of future
and
Thus,
y(t), Δv(t (9)
Δu
unconstrained control law (equation 9)control
is used. In law of
consequence, the GPC the different considers terms knowledge (9) are representedof future measurable
ion December and can
disturbances references. and
beDecember
expressed
can and references.
be Notice
expressed
as:can
references. and that references.
Notice
be as: an
expressedoptimization
Notice that
canthat and be Notice
an references.
optimization
as:expressed
ancan problem
thatand an
be expressed
optimization as: references.
Notice
optimization
mustproblem be
that
problemas: solved
Notice
must
an problem
optimization
be
using
must
6 of that solved
20 an
must
a quadratic
optimization
using
problem
be
be solved using a quadratic solveda cost
quadratic
must using
problem
function, be
6 of acost
solved
quadratic
20the mustfunction,using
be solved
cost a quadratic
function
using
where u22,
as polynomials 2017
p = Δu(tin−
Version
submittedthe vptou
1),operator
where = Energies
p Δv(t)= z22,that 2017
Δu(t −submitted
multiply
and y1),
p = vconsidered
py(t) =toΔv(t) Energies
for signals
and yp(w(t),
simplicity. =Hence, y(t) where
y(t), for the usimplicity.
Δu p solution
=and Δu(t Δv(t)). of− Hence,
the1),Thus, GPC
vpthe = (9)solution
Δv(t) and of yp = GPC y(t) for simplic
costwhen constraints
function, when when constraints
are taken
wheninto
constraints are
constraints
taken
account
are when
taken into are
since account
constraints
into taken there
when
account into
sinceis are constraints
account
nosincethere
taken
explicit there issinceinto
no are
solution
isexplicit
account
there
taken
no explicit is
[24].
into
solution
no
since account
explicit
For there
solution future
[24]. solution
is
since For
no
[23]. analysisexplicit
there
future
For [24]. is
the
future analysis
solution
no
Forexplicitfuture the
[24]. analysis
solution
For futu th
[2
can be
up = controller
Δu(t expressed
− 1), when vp no as:controller
= constrains
Δv(t) andare ypwhen considered
Δu(t)= y(t) for
no=constrains
Δu(t) issimplicity.
m(z)w(t) obtained
=law are
m(z)w(t)
Δu(t)
+(Equation
n(z)y(t) Hence,
considered
minimizing =+ m(z)w(t) the
n(z)y(t)
+
Δu(t) l(z)Δu(t) solution
isJobtainedwith= + controller
+m(z)w(t)
n(z)y(t)
Δu(t) respect
l(z)Δu(t) of(α=
+minimizing the fto
when
(z)
+ +GPCu,
m(z)w(t) +which
l(z)Δu(t)
+n(z)y(t) (α noαJfp(z) constrains
with leads
(z))Δv(t)
+ +the
+ respect
α
n(z)y(t) pto:
l(z)Δu(t)
(α areto+
(z))Δv(t)
f (z) + u,
considered+which
α
l(z)Δu(t) (αf (10)
(z))Δv(t)
pterms (z) leads
is++obtained
α(αto:
p(10)
(z))Δv(t)
f (z) + minimizi
αp (z))(10)
unconstrained
analysis the unconstrained
control law
unconstrained unconstrained
control
(equation
control law 9) unconstrained
(equation
control
is used. law 9)
Inunconstrained
(equation
isconsequence,
control
used. Inlaw
(9)) is 9)consequence,
used. iscontrol (equation
used.
the different
In In law
consequence, consequence,
9)the (equation
isterms different
used.inIn 9)(9) the
consequence,
terms
isare used.
differentdifferentrepresented
in In (9)consequence,
are
terms therepresented
different
inin (9) are theterms
represented
differentin (9) t
oller
e up when
= Δu(t no constrains
− 1), where vp are = Δv(t)uconsidered
p = and Δu(t yisp−obtained
=1),y(t) vp= minimizing
for simplicity.
Δv(t) and J ywith pHence,=respect y(t) thefor to
solution u, which
simplicity. of leads theHence, GPC to: the solution of the GPC
Equation(9)
as with:
polynomials
Δu(t) as
= with:
u m(z) are
=
K =(G−1 represented
polynomials
in
m(z)w(t)
m(z)
kG the operator
wwith:
 as
p−=z (z)+GkG as
polynomials
in the
n(z)y(t)
where
u
m(z) 
Sy= z polynomials
operator
that
−
Kwith:
=
ppzconsidered
(z)
−1
pGas+
kGmultiply
in
polynomials
zwhere
(G
(z) Gthe

l(z)Δu(t)
z
m(z)ppw that
uwith: p−
in
operator
z− G the
considered
multiply
as
where
dG 
+
+1 Sy Hvoperator
polynomials
in (α
z the
that
d − considered
(z)
z−
+2
d signals
operator
G +1
G multiply
+  α
.where

,.G z
z.H in
d that
(w(t),
+2
=p.ppp.u[zthe
(z))Δv(t)
zv
z,p,J)−
d multiply
signals
considered
that
+1
d operator
where
G y(t),
+N
.,.z.Hv
multiply
d (w(t),
+2 Δu d z
,−considered
zu
signals
dthat
+N
.G and
+1
= considered
y(t),
 multiply
,H
.],.K −kG
−1
,(8)
Δv(t)).
d (10)
(w(t),
Δu +2
d
v(G d
p )=

+Nsignals
and
+1
, .S.w signals
Thus,
considered
y(t),
],.,and
−kG Δv(t)).
z.− d
. .G +2 (w
Δu(9) (
d(w(t),
.Sy t
and
+N)
and ,
Thus,
signals
.p. −
.=−kG (8)
Δv(t)).
y(t),
G (9)
d (w(t),
Δu
Gand
+N Thus,
=p u],−kGand −y(t),
G (9
Δv
 
Jz= (z)
p =[z
(z) kG m(z) =, zp pp
u f[zu
= (z)(z)
u
kG,u, u
.p(z) zzleads
u
. (z)
.with
U
.],respect
u
=, n(z)zp[z U
.(z) u
.= .u,
=
zn(z) pz[z
u U u
n(z)u
l(z)
, zto:S,= U
],.l(z)
.n(z)
, z S= U
pl(z)
n(z) =S =
oller when no constrains y controller
( t
−1) , areand
∆u considered
when ∆v (

no
t ) ). is
constrains
obtained
Thus,  Equation are
minimizing (9)
 can with
is
be obtainedrespect
expressed

z minimizing
to as: which p
to: to which leads
u = Kcan−kG (G be expressed
 can
w
 G−kG
− G be
p . Moreover,
expressed
Sy as:Gp .−kG
p − can
G be
Moreover,G
αpGas:pexpressed
andup. α
p − −kG
α G can
Moreover, are,
andG as:
Hv
 be α+2 expressed
−
respectively,
pf−kG .αare,G can
Moreover,
p and
H
 be v as:
respectively, expressed
)
Gαpf.polynomials,
p p dare,
Moreover,
αp respectively,
andpolynomials, as:
αfαincluding are,
p and
(8)
respectively,
polynomials,
αfincluding the
are, coefficients
respectively,
polynomials,
the
including coefficients ofpolynomials,
the
theincluding past
coefficients
of the past including
the of coefficient
the the past
with:Km(z)
where = Qλ= −1+kG Gwherepz (z)
G. where
Finally,
K= according Qpλz (z) +G =
to[z
G. the
dfup+1
Finally, , z duaccording
receding , horizon
. . .. ..to U +N
, zstrategy,
the where
receding
], n(z) only K horizon == −kG
the Q first +strategy, Sand
value
G G. of l(z) Finally,
uonly is =theaccording first valuetoofthe u is receding horiz
u = K (G w  − G Sy u − = G K −1
G (G u − w G− 
G Hv
Sy − G− G H Gv ) u − G Hv − G λ
(8)
 H p vp ) (8)
(−kG H (−kG ) and ∆u H future
p )
(−kG and (−kG future
H p p ) and
(−kG
H) (−kG values
future H H) (−kG )of
(−kG
p values
and the 
future
Hdisturbance.
p H)of p
∆u)
p the
andvalues
p(−kG disturbance.
future of Depending
H) the (−kG values
disturbance. Depending
H) of
on))∆v values
thethe disturbance.
relation
Depending
on of the
the between
relation
disturbance.onDepending the between
the relation Depending
on the between
the relation
on the
th
p .Finally,
Moreover, αpifp kand are, respectively, polynomials, )pincluding the (, zcoefficients of the kpast
p p ( t ) + n ( z )−1 ypfirst
K=Q + G G.
λ−kG
computed, GΔu(t). Hence,according
computed, is tothe
Δu(t) (the
α tf)first
Δu(t). =
=receding
m row(z)of
Hence,
m(z)w(t)
Δu(t) w horizon
matrix
= +k is
if m(z)w(t) Δu(t) Kstrategy,
the
n(z)y(t) (,= )+
tΔu(t)
+ + row only
l (zΔu(t)
m(z)w(t)
n(z)y(t)
l(z)Δu(t) is
of the
given
matrix (+ + first
)l(z)Δu(t)
tcomputed,
= + by:
n(z)y(t)
+ Kvalue
(Δu(t)
m(z)w(t)
(α α−1 ff (z)
)of
Δu(t) + +
++u+
=
Δu(t). is
l(z)Δu(t)
α(α is
m(z)w(t) (fz(z) given
Hence,
(z))Δv(t)
ppn(z)y(t) + (+ α t+p)if(α
by: ,n(z)y(t)
(z))Δv(t)f is
l(z)Δu(t) (z)the + first (10)
α+p+(z))Δv(t) (α(10)
row
l(z)Δu(t) of matrix
f (z) +(10)
+ αp(α K −1
f (z),+
(z))Δv(t) Δu(t(10
αp (
disturbance-output disturbance-output disturbance-output
dead timedead and
timethe
−1dvdisturbance-output v process
ddead and
disturbance-output
time the input-output
dprocess
vdead and the time input-output
processddead
dead v and time,
time input-output
the dead dprocess
d and ,time,
such theinput-output
dead
dfactors
process, such time, are factors
input-output
dcalculated
dead
, such are
time, factors
calculated
deadd , such are
time, calculated
factors
du , such a
uted,
eK= Qλ(−kG
Δu(t). G G.
+Hence, Hpif )Finally,
and
wherek is the future first
Kaccording
= Q (−kGrow
λ +G of
to H) G.values
matrix
the receding K
Finally, of, Δu(t)the disturbance.
according
horizon is given
strategy,todthe by: Depending
receding
only the horizon firston value the strategy, relation
of udvisu+2 only between theufirst the value u
of u is d +Nu  d +N
with:inΔu(t) (m(z)) with:
mdifferent =in=kG different
ways.
kG 
m(z) w
pzp− (inzwith:
zFor
= ways.
(z)
)kG kG where
,different
the
Δu(t) 
m(z)
Sy configuration
pFor
pz= pways.
(z)
− in
=zthe
kG with:
where
((z)different
zkG)configuration

G
wFor

=m(z)
= p− puin
analyzed
[z zthe
[(z)kG d−with:
ways.
uu+1
(z) +where
different
= configuration
1 d
kG ,kG
,Syz= zanalyzed
d For
in
uu+2
pm(z)
[z
Hv+
−pways.
this
p−1 ,the
d2u +1  du +2
,zkGz.(z)
.− ..=
(z) .,.kGwork, .where
z.configuration
in .analyzed
kGFor
=..p,this

,zuHz,pthe
ddU
[z ρp dUu+N
.− vwork,
.zp++1
.(z)
> N.(z) ]0z,.where
configuration
],,.in ,nand
n(z) analyzed
this
u dU +Ndu +1 
zρΔu(t)
(= z)> ,=.work,
[z
(ndp
−.0.],z−kGand
(z)
.−+ .in
n(z) analyzed
,ρd,zzvthis
.kG 
(9) dd
=
(nd)>uU+2
S = S> and
v[z work,
0−kG
p,and
+ − .1,and
],.in
+Ndu +1  du +2
dkG vthe
l(z).),this
.n(z) l.(nd
z(ρS
.>
z,=α
)> and
= work,
zp,f
1,
+ 0−.the
p,−kG .and
dl(z) )(9)
v.kG
],.ρα.>
n(z)
=.(nd
S,G
> zand
1, =
0p+the
uandd−kG
pl(z)
−v], n(
α)(nd >=
d.G pzkG Hv =
kG H pw Sy kG
U U
with
disturbance-output zdead = time where z pz = =
uted, Δu(t). if k isthe
Hence, computed, first
Δu(t). rowdHence,
 v
ofand matrix the
if

kKprocess
is−1 the , Δu(t) first input-output
row

ispgiven of by:
matrix dead 
Ktime, , Δu(t) u , such ispgiven factors by:are calculated p,f p,f
Δu(t) = −−kGkG polynomials
w
GG −
pp . .
−kG
kG Moreover,
polynomials
Moreover, are
Sy G ppas − . α Moreover,
polynomials
follows:
−kG
kG αare
p p and and
as
G G α u
α
p fppf p.
follows: α−
are, Moreover,
are, and
polynomials
are −kG
kG as respectively,
follows:
Hv
α
respectively, G
f pp are, .
polynomials
−
α are Moreover,
−kG
kGand
respectively,
as polynomials,
follows:
H
α
polynomials,G are,
are
v
fp pp p . α Moreover,
as respectively,
polynomials,
and
follows: including
α
including f are,
α (9) polynomials,
and respectively,
including
the the α coefficients
are,
coefficients the
respectively,
including
polynomials,
coefficients
of of the the the
past polynomials,
pastcoefficients
including
of the past theincluding
of coefficie
the pas
in different
This unconstrained ways. For
Thisthe
control configuration
unconstrained
law
  of the GPC
analyzed
control
 considers lawin of
this thework,
knowledge GPCρconsiders of>future 0 andunconstrained
This (nd
measurable
knowledge + pdv ) of > f
disturbances1, the measurable
control
future α
law p,f of the disturbances GPC considers knowledg
Δu(t) (= kG w
−(−kG H
Hp− ) and
)p(−kG
kG
and Sy future
Hp p−)(−kG
Δu(t)
future and
kG (−kG
(= −kG 
future
GH 
pu wpp)(−kG
H)
H) −and values
kGfuture
(−kG
values 
Sy
Hv
H)of of H p the
values
(−kG
−
p )(−kG
the kGand
disturbance.of
future
G
disturbance.H
H) Hthe
p uvp values
p ) disturbance.
−(−kG
and kGDepending 
future
of 
Hv
Depending the
H) (−kG
− (9)
disturbance.
values Depending
kG on the
of
onH
H) pthe
relation
thevvalues
p on Depending
disturbance.
the
relation between
ofrelation
thebetween on (9)
disturbance.
Depending
the
the
between relation theDepending
onbetweenthe relatio tho
unconstrained polynomials
control law
and references. are as
Notice follows:
ofand thethat GPC
references.
an considers
optimization Notice α pknowledge
(z)problem
that= αα an
pp(z) + of
optimization
must = α future
p α be
zα
p
−1
p +
(z) + measurable
solved α .
=
p problem
. z. α −1
+ using
p αand
+α+ p p .(z)
α must
. a disturbances
references.
.
p +z= −1−(dv +nd−1)
quadratic
zα αbe
p α+
p p solved
(z)
.+ . .α cost
+Notice
= zp
−(d
α
z α, −1
pp
v +nd−1) −(d
using
function,
+ that .
α a. pz. , +
z −1
quadratic
an αv +nd−1)
optimization
p + . . .cost
,+ z −(dv +nd−1)−(dv +nd−1)
αare pfunction,
problem z , must be, so
the disturbance-output
disturbance–output disturbance-output dead disturbance-output
time dead
dead dv and
time time d the
v dprocess
disturbance-output
and
1
vdeadand time
the 2 1
the input-output
process process
dv and
disturbance-output 2
dead input-output
the
input–output
1
time dead
(nd+d
process
2 v)
ddeadtime,
v and
1
input-output
(nd+d
dead
deaddthe
time u , such
v process
) 2 time,
time,
dv and 1
factors
(nd+d
ddead
uthe
dvinput-output
,usuch process
are
time,
) ,2 such(nd+d
factors
calculated
duinput-output
factors ,dead
v )
such
are calculated
time,
factors
(nd+d v )
ddead
uare
, such time,
calculated
factors
du ,
eferences.
unconstrained
when Noticecontrolthat
constraints This an
law
calculated are optimization
unconstrained
of
taken
when thedifferent
in GPC
into
constraints problem
accountconsiders
control ways. are must
law
since For ofbe
knowledge
taken there
−1
the thesolved
into is GPC
account
no
configuration using
ofexplicit future
considerssincea solution
quadratic
measurable knowledge
there
−(d
analyzed when
v +nd−1) is
[24]. cost
noin function,
disturbancesof future
constraints
explicit
For
this future
work, solution are measurable
analysis taken
> [24].
0 the
andinto For disturbances
( account
nd future + d since
analysis
) > 1, there theis no explicit
in anddifferent in
the future ways.
different
and the α (z)
p part For
future= hasin
ways.
α the different
andpfollowing
part + configuration
the future
1 αFor
has z the
ways.
p2 following
and in
form:
part+ configuration
different
.
thehas. For
.analyzed
+
future form:
and α the
following inways.
the configuration
different
partfuture
p(nd+d inanalyzed
hasz this
) form: For following work,
ways.
the in
part has following configuration
analyzed
this
,
ρ For
form:> work, 0the andin configuration
ρ thisanalyzed
(nd
>
form: v
ρ work,
0 + and d ) in
ρ
(nd > analyzed
>this1,
+ 0 dthe
work,
andv ) v α in
> (nd
p,fρ this
1, >
+ the work,
d 0 α
)and
v p,f > ρ 1,
(nd > the+0 and
αvp,)
d
constraints
eferences. are taken
Notice
unconstrained that
theandinto
control account
anreferences.
optimization
law
unconstrained
polynomials since
(equation Notice there
problem
are 9) that
control is used.
is
as noan
must
follows: explicit
lawoptimization
In be
(equation solution
solved
consequence, 9)using
problem
is[24].
used.
the For
av different
quadratic
must
In future beterms
consequence,
unconstrained costanalysis
solved function,
in using
(9)the the
control
are arepresented
different quadratic
law terms
(equation cost in (9)function,
9) are is represented
used. In consequence, th
polynomials
α p, f polynomials
are as follows: polynomials
are as follows: are polynomials
as follows: polynomials
are as follows: are as follows:
strained
constraints and
ascontrolthelaw
are
polynomials future
taken when inpart
(equation
into the has
constraints
account
operator following
as9)polynomials
is used.
since zare that In form:
there
takenconsequence,
αinf is
multiply the
(z)into
no= explicit
account
operator
αconsidered
αff(z) 1 zthe = different
(duzsolution
since
−d that
α v ) signals
α
f1f + z there
multiply
(d
(z) α u −d
f1
terms
[24].
= z )isu −d
v(w(t),
(d
α + no
For
f1 ff1
in
considered
α
z (d)+1
v (9)
explicit
future
y(t),
(z)
u
z as
−d (dare
+ vu polynomials
= )−d Δu
.
+ α
.represented
vanalysis
solution
α
. signals
)+1
α + and
z
f1ff1 fN(z) (du(d−d
αz+ u −d
Δv(t)).
=
. .
z vthe
[24].
(w(t),
(d
.
α in
) vu)+1
+ +
f1 ffN1the
−d
z
α α(d For
v Thus,
uoperator
)+N
+ −d
y(t),
z . (d
(d
.vfuture
.
u) −d
u −d
Δu
+ (9) vv)+N
α )+1and
z
ffN
analysis
1
that
zz+ (d
(d uu−d
Δv(t)).
. multiply
−d
. . vv)+1
+ the
)+N
α Thus,
+
fN zconsidered
. (du(9)
. . +−dv )+N (d
α fN z signals
u −dv )+N

ynomials inbethe operator z that multiply considered signals (w(t), −1 Δu and −( Thus, +nd(9)
dvv+nd−1) −1)−(dv +nd−1)
nstrained cancontrol lawunconstrained
expressed (equation
as: can 9) iscontrol
be used.
expressed Inlaw consequence,
α(equation
as:
α p
u −dv )
(z)
( z ) = αα α9)
p
p ptheis
(z)
+ used.
different
αα
(du −dv )+1 =p zα −1
y(t),
pIn α +
+ consequence,
p terms
(z)
.α
. . p. =
+z in
−1
α
α p(9)
+ can
Δv(t)).+.
(d+ud−dαare
. the
αbe
.p (z)
+p zrepresented
zdifferent
−(d
α −1
expressed =p +
α p .α . +
p .
zterms
(z) as:
+,α, pα = z
p in
−1
α p(9)
+ + ,.zare
. α . +
p represented
−(dv +nd−1)
z α −1
p + .
, . .z +−(dv +nd−1) −(dv +nd−1)
α p ,z
αwhere (z) =118 αfrefers z (ddegree +ofαdegree 1
1 z to
1 22 1
+ . to
.D . the
+ 2
αdegree 1
(nd+d
(nd
zto v v )+N
) 2 (nd+d 1 v ) 2 (nd+d 1 v ) 2 (nd+d v ) (nd+d v )

as where
f refers to the
where tond118
the
zrefers where
fthe polynomial
the
of the where
refers
degree polynomial of from
refers
the polynomial
equation fromthe of(w(t), the
degree
equation
(5). polynomial
from
Moreover,
of
nd in nd nd nd D D (9)the (5).equation
polynomial
Moreover,
Dpolynomials
from (5).equation Moreover,
polynomials
DThus, from referring (5).
equation
polynomials
Moreover,
referring (5). Moreover,polynom
referring
fN
elynomials
expressedinas:the operator 118
polynomials
z 118that multiply the considered
1
operator signals
that multiply
(w(t),
118
considered
y(t), Δu and signals Δv(t)). Thus, y(t), Δu and Δv(t)). (9)
and the future and the
part future
has and
following
part the has
future following
form: and
part the
has form:
future
followingand part the has
form:
future following part has form:following form: (10) + α(10)
canto past αtopm(z)w(t)
and
past future to+ past ffuture
values
α119p =and to fof past
future
values
disturbances
119α to αof and
fpast disturbances
values
future
and Δu(t)
the future = part αhas Δu(t)
n(z)y(t)
α
following α +m(z)w(t)
form: l(z)Δu(t) + n(z)y(t)
+ (ααfcan p(z) and
of+ αbe
+disturbances
future
α grouped
values
l(z)Δu(t)pcan (z))Δv(t) be αof fgrouped
into
+ values
disturbances
can
(αα f (z) be=
Δu(t) of
intogrouped
+ disturbances
α αp= αp+ can =αfbe
(z))Δv(t)
m(z)w(t) into. p grouped
α Additionally,
+αcan α+= .beαAdditionally,
f n(z)y(t) into
pgrouped +f .= into
Additionally,
l(z)Δu(t)
αp + α α=f+. α (αA
e expressed
118 whereas: nd 119 refers beto expressed
119
the degree as:
119
ofpthe polynomial D fromp equation (5). fMoreover, polynomials referring
Δu(t) =120m(z)w(t) theprediction
120 + n(z)y(t)
the prediction mechanism
120 +
the  l(z)Δu(t)
prediction
mechanismof120 the the+ (α
(dintrinsic
mechanism f (z)
prediction
of120 the the + α
intrinsic
compensatorp (z))Δv(t)
prediction
of
mechanism
(dvu)the −ddcompensator
intrinsic
(dcan mechanism
of be the
compensator
also can
intrinsic used
ofbe (10)
(duuuthe also to compensator
can
intrinsic
include
used be(duuto also compensator
future
include
) usedcan disturbance
beto future
also
include can used
disturbance
be future
also
to uinclude used
disturbance
)+N to future
incl
to past
119with: m(z)αp=and kGfuture pzwith:
(z)αwhere f m(z)values p= α(z)of ((z)
fkG disturbances
 ααduz(z)
=pzα[z(z)
)m(z)w(t)
=
zl(z)Δu(t) f f (where
+1 du,u− z=ddcan
−d
)) p
v +2
vuα f+
+ be
1,zα z.(z)(du
ααf+ .grouped
.ff(z)
−d
.z(.d=.=u,− αdvvfUdα
z+
[z
)+1
+N
1z
)+u into
+1
f11α z+ ],+
(d−d
,uwith:
fz(z)α.d. u−d
u
n(z) .. )..=
v +2 v )+1
+
+ m(z)
,= .α α.ff.p+ α
−kG
N.z
z+fz
(d
.= .d,α
((z) .u−d −d

+
kG
z−f=
d.dvvUvα
Sv
))+1
)+N
α
)++Additionally,
and
+N
fp fN z+
α (d
zz(z)
],,fl(z). z.−d
n(z)
−d
.(d
where
+ u −d
vv)+N
==α+ v )+1 (du
f−kG
Npz
αzf1+ (z)z−d (d
. S.uv.−d )+N
+v )+1
and
= α du+
[zfl(z) z (d.,=
+1 .z−d
.d+ uv+2 α,f.N.z.(d. u.−d .,
Δu(t) = 121 m(z)w(t)
estimations,
121 + n(z)y(t)
estimations,
v, Δu(t)
121through
zα
+f=
estimations,
v, the
through 121
f 11zestimations,
polynomial
v, +
the + (α
through f
121 (z)
n(z)y(t)
polynomialα
1z
estimations,
. thev,
1 α
In+ p (z))Δv(t)
l(z)Δu(t)
polynomial
through
such
α . aIn v,scheme,
the
such +
through α (α
polynomial
a. α(z)
scheme,
fthe
In f1N
the
1
such +
implicit (10)
α
polynomial αathe
p (z))Δv(t)
scheme,
.
N 1
feedfroward
implicit
In such
1
α the .a
feedfroward
In
scheme,
implicitcompensation
such a (10)
feedfroward
the scheme,
compensation
implicit
N
the feedfroward
compensation
implicit feed c
m(z)120=−kG the prediction
kG 
Gpzp(z) where
. Moreover, mechanism
pz−kG αp and
(z)  =of
G[z αpthe
d u +1
.f Moreover,
are,intrinsic d u +2
, zrespectively, , compensator
.α. p. .and d
. . ,polynomials,
zαf are,
U +N f
can be also
], respectively,
n(z) f
including
= −kG used −kG  to
polynomials,S include
f
the andGcoefficients
p .l(z) Moreover,future f
including
= ofdisturbance
the αpthe
f
past
andcoefficients αf are, respectively, of the pastpolynomials,
118 122where 122 refers
where in to
is included
nd
118 is included nd122 the
control
118
dreferswhere
is+1degree
inincluded
law, to
control the
nd of118
122 refers
however the
degree
in
law, where
polynomial
is control to
included ofthe
however
the122 the isdegree
daUdisturbance
nd law,
118
drefers
in polynomial
where
included
thecontrol
Dhoweverfromof
to
disturbance ndthe
the
cannot
in equation
law, refers
thepolynomial
degree
control
D from
however tocompletely
disturbance
becannot (5).
ofthe
equation
law, the Moreover,
degree
the
D be polynomial
however fromcannot (5).
completely
disturbance ofequation
Moreover,
attenuated
thepolynomials
be polynomial
disturbance
Dcompletely from
(5).
attenuated
cannot as was polynomials
Moreover,
equation
referring
be shown
D from
attenuated
cannot
as
completely was (5).polynomials
in referring
equation
be
shown Moreover,
completely in(5).
asattenuated
was referring
shown polyno
Moreo
attenu
as in w

m(z) = estimations,
Gp .121Moreover,
(−kGkG H α)where
pzp(z) and
with:
pandwhere through
v, future
αndm(z)
fp are,
(−kG
refers
z (z)(−kG
= thekGto
=H polynomial
respectively,
the
p )pand
[z
H) (z), zpolynomials,
zvalues
udegree
where
du +2
future αoff,.the
.the.In
(−kG
p such
. . ,including
disturbance.
.polynomial
.z (z) zH)= scheme,
+N
[zvalues ],u +1Dthe
n(z) ofdthe
Depending
, zfrom u +2implicit
coefficients
=the,−kG .disturbance.
Equation . .on
.(−kG 
. .S,thefeedfroward
of
zandd(5).
H Uthe +N
relation
p )l(z) ],past
Depending
and
Moreover, n(z) =future compensation
between = −kG (−kG
on the
polynomials the 
andvalues
Srelation
H) referring
l(z) between
=of the the disturbance. D
119 123to [9,13].
past αpThis to[9,13].
and past future
isfuture
due This p to and
to isλpast
[9,13]. future
duevalues
parameter
This
to λαandof
is
[9,13]. to
parameter
f due values
disturbances
that future
past
to
isThis used of
[9,13]. fdisturbances
αparameter
that
αpis and
to
invalues
iscan
due past
the
This
usedfuture be
tointernalαof
thatis
in grouped
pparameter
dueand
disturbances
iscan
the f used values
future
be
structure
to
internalλ in into grouped
parameter
that of
the can
αstructure
of
fis disturbances
values
internal
the
used be
into
that
GPC.
αin pgrouped
of of disturbances
can. α.into Additionally,
be grouped .that
can Additionally,
be into grouped
αfα.GPC. Additionally
ofαinto .
G G Hp122 is future
included
.) disturbance-output
and
Moreover, (−kG
αto in
p−kGand
past 119
control
123
dead
H)
α α are, law,
values
.time
and Moreover,
123
however
respectively,
f pdisturbance-output
G p
α
119
dofv and the αααffpthe
disturbance.
the
αp119
123
disturbance
polynomials,
and
values
process
dead αftime ofare, Depending
123 λ
cannot
drespectively,
input-output vincluding
disturbances
and
119
thedead be
on the
process completely
the
polynomials,
can time,
λ
relation
coefficients
be α
dgrouped
input-output u , such
disturbance-output attenuated
between
including
offactors the into
dead
α
the
past as
=
the
αare
time, was
= αdstructure
+is
the
αthe
shown
coefficients
calculated
dead +
pu ,time
Itused
α =
such
internal
fwas
αGPC. fdin
inpshown
of + the
Itthe
of the
structure
α
Additionally,
vfactors
and
αwas internal
f GPC.
the
=
are
shown
past
αpfor
process
It
of
structure
+
calculated
wasthe
that shown
for=
input-output dea
Itthe
pthatwas
+GPC ααforfsh
=
120 124the λ prediction
> 120 0,
124 the
λ >prediction
feedforward mechanism
0, 124 feedforward
120 λ the mechanism
prediction
compensator
> of
0, the 120intrinsic
feedforward
124compensator
λ the
mechanism
of
will
> 0, the
prediction
never
124 compensator
intrinsic
feedforward
compensator
λ
120 will > theof
performnever
0, mechanism
prediction
thecompensator
feedforward intrinsic
can
compensator
will
perform
as a be
never
classical mechanism
of
also
ascompensator
the
can
compensator
perform
a willused intrinsic
be
classical
disturbancenever also
to of
as include
athe
can
used
willcompensator
disturbance
perform intrinsic
classical be to
compensation
never future
also
include
as compensator
used
disturbance
perform
a disturbance
can
compensation
classical future
tobe
structure.
as include
also
a disturbance
can
used
compensation
disturbance
classical future
structure. be to also include
disturbanc
disturbance used
compensati
structure. futu
to c
[9,13]. This the isprediction
ddue to λ parameter mechanism that ofisForthe
used intrinsic
thein the compensator
internal du ,structure can
0ofin be
inthe also
GPC. used Itdways. to include αfuture disturbance
bance-output
Hp123)inand dead
future time
(−kG
(−kG and the
values process
)different
and of
future the input-output
(−kG
disturbance. dead
values in time,
Depending of work, the onsuch
disturbance.
ρthe >factorsrelation are
Depending calculated
between vwas )ρthe on >>shown the relationthat for+ between dv ) > the

G different ways. v For
H) H inpthe configuration ways. analyzed H) configurationthis analyzed and different
this(nd work,
+ For
1, 0 the andthe configuration
(nd
p,f 1, the αp,fin this work,
analyzed
estimations,
The121
estimations,
121 125 ideal
125 estimations,
The
v,v, through
compensator
ideal
through 125 estimations,
121 compensator
The v, the
the through
ideal
can polynomial
polynomial
125 the
becompensator
121Theobtained
can
v, estimations,
through
polynomial
ideal
be α
125 ..only
ffobtained
The
121 InIn the
compensator
can such
estimations,
suchideal
when
v, bepolynomial
αfthrough
.aobtained
only
a In
scheme,
compensator
λ when
scheme, such
=canthe 0,be
v, α athrough
only
as
λthe
fpolynomial
the .scheme,In
obtained
= was implicit
when
can 0,
implicit such the
bethe
shown
as λonlyaobtained
polynomial
was feedfroward
α= scheme,
implicit
fin . 0, In
shown
feedforwardwhen assuch
[13]. only the
wasfeedfroward
λin αcompensation
afimplicit
.scheme,
Nevertheless,
= [13].
when
shown In as
0,
compensation such feedfroward
compensation
=the
Nevertheless,
λin was a the
[13]. scheme,
implicit
shown
0, asNevertheless,
was incompensation
thethe
feedfroward
shown
[13].implicit Neve
inthe [1
ferent ways.
rbance-output
124 For
polynomials
λ > 0, feedforward
the
dead time are configuration and
disturbance-output
asdvfollows: compensator
polynomials analyzed
the process dead arewill in
input-output
time never
this
as follows: perform
work,
dv anddead the process
ρ α as a
time, duinput-output
> 0 classical
and , such factors
(nd disturbance
+ d v dead
polynomials
) > are time,
1, compensation
the
calculated are
αdp,f u ,as such follows: structure.
factors are calculated
isisincluded
122 126 included
resulting 126 is
122 ininincluded
resulting
controlcontrol
control 126122scheme in
is
law,
control
resulting
law; included
control however
is
scheme
however, law,in is
characterized
126control
122 resultingthe
control
however
theincluded
disturbance
scheme
characterized
126 controllaw,
resulting
disturbance
122by theis
incharacterized
however
isimplicitincluded
disturbance
control
cannot
schemebycontrol co-design
implicit
cannot law,
thein
isbe scheme control
disturbance
cannot
completely
however
characterized
be by co-design
for
implicit
completely law,
becharacterized
isfeedback thecompletely
cannot
however
attenuated
for disturbance
co-design
by andbefeedforward
feedback
implicit
attenuated the attenuated
completely
asco-design
for
by disturbance
cannot
was
and
implicit
feedback shown
as feedforward
was as
be attenuated
part was
co-design
forand
showncompletely
cannot
inand shown
feedback
feedforward
part be
asand
for completely
in
was
attenuated
and
feedback shown
feedforw
part and as
andain
omials
ferent125are The
ways. ideal
as follows:
For the compensator
in configuration
different ways. can beFor
analyzed obtained theinconfiguration
thisonlywork, when ρ λanalyzed
>=0 0, andasin (nd was this + shown dv ) >in
work, ρ1,> [13].the 0 and αNevertheless,
p,f (nd + dv ) the > 1, the αp,f −1
in[9,13].
123 127 the high
[9,13]. 123 This
127 [9,13].
This the
α isp (z)
bandwidthduedue
high
is This
127= tobandwidth
123 α the
λin[9,13].
is
to parameter
+due
the
high
the α to
λThis
feedback
bandwidth
in
z
127 λ−1 parameter
that
theis
the
parameter
123 α + [9,13].
due
(z)
. is
feedback
high
loop,
. . =
+used
127to
in α This
λ that
bandwidth
what
the
that
123 parameter
in+ [9,13].
loop,the
is
feedback
high
is
α used
due
results internal
used
z what
bandwidth
−(d
z in−1 to
This
vin
that
the
+loop, the
results
+nd−1)
λ
in very
. parameter
is
structure
is
feedback
.the . internal
due
used
what
+
, in
aggressive
in to
the
internal
α in
very of
results
λ that
structure
the
loop, parameter
feedbackthe internal
aggressive
z is GPC.
changes
−(d
what
in
structure used
very of
loop,
v +nd−1) that
It
in
the
structure
results was
the
aggressive
changes
in αof
,whatGPC.
is
controlin
the
(z) internal
used
shown veryof
It
results
= in
GPC. in
α was
the
signal.
changes that
the
control structure
aggressive
+ inGPC.
shown
It internal
αfor
very
Due
was It
signal.
inz that
of
was
control structure
aggressive
changes
+ the for
shown
Due
. . GPC.
. signal.
+ inα p(nd+dv )G
of
that
It
changes
controthe
was
Due fo
nomials resulting
126 are as follows: control
polynomials schemeare is as characterized
follows:
−1
p1 p2 by implicit p
−(d
co-designp1(nd+dv )p2 for feedback and
+nd−1)
p(nd+d feedforward
v) partp and p1 p2
shown
124 128αλp (z)to 0,
> the feedforward
αimplicit
that,
=124
128 pλ1to >the
for
+ α0, λ
p2128 zfeedforward
> compensator
relation
implicit
124 to
0, +λthe the
.>.between
relation
. 0, feedforward
αcompensator
implicit
feedforward
+ 128 p124 will
to
between
feedback never
λrelation
the z 128
> compensator
implicit feedforward
will
0,compensator
v
124 perform
feedback
to and
between
theλnever
relation
feedforward
,implicit
> 0,as
and perform
feedforward
will
acompensator
feedback classical
between
willfeedforward never
relation as
actions,
never andafeedback
perform
disturbance
classical
compensator
between will
feedforward
performsuch
actions, neveras disturbance
structure
and asuch
feedback
as compensation
classical
perform
will
feedforward
aactions, never
anddisturbance
structure
classicalcannot compensation
as
such aperform
structure.
classical
feedforward
beactions,
cannot used as
structure
disturbance compensation
as
be disturbance
suchstructure.
aacannot
classical
actions,
used structure
asbe a disturban
such compens
structure
used cannot
structu
as a
127
andthethehigh
future bandwidth
part hasin and the
followingthe feedback
future form:
−1 part
loop, haswhat
(nd+d
following results
v) in+nd−1)
form: very aggressive and changes the−(d future in controlpart has signal.
following Due form:
compensation
125 129α The stand
p (z) ideal αThe
=129alone
125 stand structure.
pcompensatorideal
1 +disturbance
αpalone2125
129 z compensatorThe
stand can The
disturbance
+ ideal
.α.compensator
p.alone
(z)
+ ideal
be129αobtained
= pcompensator
125 can
disturbance
stand compensator
pThe
αcompensator
1 v+ be −(d
ideal
zwhatonly
alone
α
129 p2125
v −1
obtained stand
zlimits can
when
compensator
The
compensator
disturbance
what
+alone be can
only
. ideal
.,its .limits
λ be
obtained when
applicability.
+ =disturbance
α obtained
pcompensator
compensator
what
0, can
its asapplicability.
λonlywas
be
limits
z= This vonly
obtained
when
shown
compensator
0, its as
+nd−1)
can
what was
λbe
particular when
applicability.in
Thisonly
=shown
,limits [13].
obtained
0,whatwhen
particular
λ its
issue =
asNevertheless,
was
in 0, only
applicability.
limits
This can as
λ[13]. shown
issue
= was
itswhen
particular
be 0,addressed shown
Nevertheless,
asinThis
applicability.
can the
was
λ [13].
be issue shown
=addressed
particular
0,canNevertheless,
as the
This was
beinparticular
[13].
issue shown
addressed canNe
th
e future to the
128 part hasimplicit
following
in [13]. relation
form: between feedback
Nevertheless, the resulting and (nd+d
feedforward
control
)
scheme actions, is such
characterized
(nd+dv )
structure by cannot
implicit beco-design
used as a for feedback
126 130resulting
with126 control
resulting
predictive
with
αf (z)compensator= α f1 zscheme
predictive control
feedforward (d resulting
with u −d is )scheme
characterized
vfeedforward
predictive
+αα control
compensator
with is
(d resulting
u characterized
−d scheme
feedforward by
predictive
compensator
v )+1 that
(dwith implicit
control
−dis resulting
was characterized
vpredictive by
)feedforward
compensator scheme
thatco-design
recently implicit
control
was (d (dufeedforward
u−d −d is
recently
introduced co-design
for
by
characterized
compensator
vthat
v)+1
)+Nscheme implicit
feedback
was introduced is
recently
compensator
in for characterized
co-design
[12]
that and
feedback
by was
and implicit
feedforward
introduced
(d inu −d [12]
which
recently for
that
)+N andand by
co-design
feedback
was is feedforward
in implicit
which part
introduced
summarized
(d [12]
recently
u −dv and
andand
)for
is co-design
feedback
feedforward
part
summarized
introduced
which
in [12] (dandis−d for
and vand feedback
summarized
in
)+1 part
which feedfo
[12]
+ . . .is+
and
f f(z)1z = itsαapplicability.
f1 z + . . . + αThis f1Nzz particular + issue . . . +can αf be zαaddressed
f (z) = αf1 z + αf z
130 130126 130 126 130 126 u v u
he future
129 standhas
part alone and disturbance
following
and the form:
feedforwardfuture part part hasand followingwhat
the limits
high form: bandwidth invthe feedback loop, whichNresults in very aggressive 1
131the =inhigh 131bandwidth
following the high
in following
section. bandwidth
fin inthe high
feedback
section.
following inbandwidth
the in .loop,
section.the
feedback
following high
αwhat
fin zthe
in bandwidth
loop,
results
the
feedback
following
section. high
what in in bandwidth
very
results
loop,
section. theaggressive feedback
what
in very inresults the aggressive
loop,
changes feedback inwhat very inchanges
results
loop,
aggressive
control what inin
signal.
very
control
results
changes aggressive
Due signal.
ininvery control changes
Due
aggressive signal. in contchan
Du
(d −d ) (d −d )+1 (d −d )+N
130 withnd α127f (z)
predictive α 1z
f127
feedforward
u v
+compensator
131α
127
1z
u v 131
that+
127 .was . +recently
131
N
127 u
introduced in [12]equation and which isthe summarized
where refers to
to=theα128 the where
degree nd of refers
the polynomial
to the degree D from
of the equation polynomial (5). 118Moreover,
D where
from nd polynomials
refers (5). to Moreover,
referring
degree of polynomials
the polynomial referring D from equation
fimplicit to therelation implicit
αf1αzto fbetween
the
relation
=implicit feedback
+zbetween
. to. relation
theα+ and
ffeedback
implicitαzbetween
feedforward
to
z the relation
and implicit
feedback
feedforward
+actions,
between
. . relation and αsuch ffeedback
feedforward
actions,
zbetween structure such
and feedback
actions,
feedforward
cannot
structure and
such
be cannot used feedforward
structure
actions, asbea used such
cannot actions,
as structure
abe used suchcanno as
stru
118 118 (d −d ) (d −d )+1 (d −d ) (d −d
(d −d)+N )+1 (d −d )+N
α128f (z) 1z
u v
+ 128 (z) u v
αf128 u
.+ v 128 u u v v
.+ u v

to intofollowing section. N f1
nd119refers theαpdegree of119thetoαpolynomial from equation (5).ofMoreover, polynomials αreferring
131 1 N
past 129
and standfuture alone
129 fpast
stand values
disturbance αp129and
alone ofD disturbances
stand future
disturbance
compensator alone values
α129f compensatorcan
disturbancewhatbe
stand alone
limits grouped
disturbances
129 compensator
what
stand
disturbance into
its applicability.
limits
alone119 α can
to
its
what
disturbance=past
compensator belimits
αThis
applicability. pgrouped+ and
α
p particular
its .applicability.
compensator
fwhat future
into
ThisAdditionally,
limitsααf=what
issue
particular its values
can αThis +
applicability.
p be limits
issue of
α f .disturbances
particular
addressed itsAdditionally,
can applicability.
beThis issue
addressed canbe
particular
can beaddressed
This groupe
issue
particu ca
et nd
α120 and
the future
p refers to the118
predictionαdegree values
f where of
mechanism
120 nd ofthe
the disturbances
refers
polynomial
prediction
of to the
the intrinsicD can
degree from
mechanism be grouped
ofequation
theofpolynomial
compensator the into
(5). intrinsic
can Moreover,
αbeD =also from
compensator
αp used polynomials
+ equation
αthe fto . include Additionally,
(5).referring
prediction
can be Moreover,
future
also mechanism used
disturbancepolynomials
to include
of the referring
intrinsic
future disturbance
compensator can be al
130 with predictive
130 with predictive
feedforward
130 withfeedforwardpredictive
compensator
130 with feedforward
compensator
predictive
that 130 was with recently
compensator
that
feedforward
predictive
120
wasintroduced recently feedforward
that
compensator wasintroduced
in recently[12]compensator and
thatinintroduced
which
was [12]recently andisthat summarized
which
inwas [12]
introduced recently
isand summarized
which inintroduced
[12] is summarized
and in which [12]
ediction
st α andmechanism
future
121p estimations, fv, through
α of
to the
values
past intrinsic
αof disturbances
and
estimations,
p the polynomial compensator
future α can values
through
f αf . Inthe
v,insection. can
be be
grouped
of
such also
disturbances
polynomial used
into
a scheme, α to =include
can
the. In α be
implicit
p121such
+ future
grouped
α .
estimations,
f afeedfroward disturbance
Additionally,
scheme, v, into α thethrough = compensation
implicit
α + the α .
feedfroward
f polynomial Additionally, compensation
αf . In such a scheme, th
inα
119
131 in following
121
131 in following
section. 131 following 131 section. in following 131 section.
ffollowing section. p
Energies 2018, 11, 123 6 of 17

changes in control signals. Due to the implicit relation between feedback and feedforward actions,
such structure cannot be used as a stand alone disturbance compensator, which limits its applicability.
This particular issue can be addressed with a predictive feedforward compensator that was recently
introduced in [12] and which is summarized in the following section.

Predictive Feedforward Compensator

Version December
Version22,December
Version
2017 submitted
December
22, 2017 tosubmitted
Energies
22, 2017 submitted
to Energies to Energies 6 of 20 6o
From the previously presented approaches, it can be deduced that the analyzed solutions provide
some advantages and also drawbacks. In [12], it was found that advantages of previously introduced
where up =where where
Δu(t u−p 1),
= Δu(t
vp = up − Δv(t)
=1), vand
Δu(tp = −y1),
Δv(t)
p =vpy(t) and for
yp simplicity.
= Δv(t) = and
y(t)y for=Hence,
simplicity.
y(t) for
thesimplicity.
solution
Hence, theofHence,
the
solution
GPC theofsolutio
the G
schemes can be merged using a predictive feedforward compensator that isp derived from the GPC
algorithm withcontroller
implicit when
controller
no constrains
feedforward controller
when
actionnoareconstrains
when
considered
using ano constrains
areisconsidered
specific obtainedare considered
algorithm minimizing
isconfiguration.
obtainedisJminimizing
obtained
withItrespect
was to u, respect
minimizing
with
Jshown which
J with
that u, which
leads
to respect u, w
to: toleads
an improved disturbance compensator u = K−1 can(Gbe obtained
=− −1 by  setting
−
K− −1  the error
upp− −G in the
GSy −− feedback
  structure
GHvHp v−p )G H(8)
u w KG (G
uSy=pw GG(GGSypw GHv pu p− GG GG vpp )−
HpHv
pu − G p vp )
equal to zero. This assumption can be confirmed by the GPC control law provided in Equation (8).
where Kthat
Taking into account =Q where
the
λ +set 
where
Kpoint
G =G.QFinally,
λis+K according
G=to
set Qw
G. λ (Finally,
+j)G=to y(according
G. tthe
)Finally,
receding
for j =according
tot +
the
horizon
dreceding
+ 1,tostrategy,
. the
. . , horizon
treceding
+ donly
+ strategy,
N,the
horizon
firstthe
and only
value
strategy,
theoffirst
uonly
isvalue
the firs
of
system is in a computed,
steady state,computed,
then
Δu(t). Hence,
u = ∆u
if k isHence,
p computed,
Δu(t). ( t − the
1 ) =
Δu(t). if k
first0.row
Hence,
is the
of matrix
Equation if k is
first(8) K offirst
row −1
the row K
, matrix
representing
Δu(t) is −1
ofgiven, Δu(t)
matrix
control K is ,given
by: −1 results in:
law Δu(t)by:is given by:
 −1  
uΔu(t)
= (Qλ=+kG G)−
GΔu(t)
w =
kGkG
(−Δu(t)
GSywp−
−−
=
Hv GG
kG
kG Syp−
w
H v−
pu ppkG
 
)−.kG
kG
SyG −upkG
pHv
p −  
Gpu
− kG
kG H
Hvpp−
vp−
kG 
kG Hv
(11) 
Hp−vpkG(9)H p vp

This
Following thisunconstrained
Thisaunconstrained
equation, control
Thisfeedforward
typical unconstrained
law control
of the compensator
GPC
lawcontrol
of
considers
thelaw GPC ofknowledge
considers
the GPC
structure can of
considers
knowledge
be future
obtainedmeasurable
knowledge
ofonly
future
when of
measurable
disturbances
future measurable
disturban
andpreviously)
λ = 0 (as shown references.
and and,
references.
Notice and
as that
references.
an
Notice
optimization
a consequence, thatNotice
anλ optimization
Q =problem
that an optimization
0 results must problem
in: be solved
must
problem
using
be solved
must
a quadratic
be
using
solved
cost
a quadratic
using
function,
a quadratic
cost funct
when constraints
whenareconstraints
taken
wheninto
constraints
areaccount
taken into
are
since
taken
account
thereintoissince
account
no explicit
theresince
issolution
nothere
explicit
is
[24].
nosolution
explicit
For future
[24].
solution
analysis
For [24].
future
theFor
analysis
futur
unconstrainedunconstrained
control law = −G−law
u (equation
unconstrained
control
1
Hv
9) is−used.
control
−1
Glaw
(equation H9)
In visp ,used.9)
(equation
pconsequence, Inisconsequence,
used.
the different
In consequence,
the
terms
different
in (9) (12)
theare
terms
different
represented
in (9)
terms
are represen
in (9) a
as polynomials
as polynomials
in the operator
as polynomials
in the
z that
operator
multiply
in thezoperator
that
considered
multiply
z thatsignals
considered
multiply
(w(t),
considered
signals (w(t),
y(t), Δu signals
andy(t),(w(t),
Δu and
Δv(t)). Thus, (9) andThus,
y(t),Δv(t)).
Δu Δv(
where H and Hp refer to future and past disturbances, and G is the process dynamics. Then, it results
in a commoncan be expressed
feedforward canstructure
be
as:expressed
can where
be expressed
as: the disturbance
as: dynamics are divided by the process
dynamics with a reversed sign. Δu(t) In =the case
m(z)w(t)
Δu(t) when = future
+m(z)w(t)
n(z)y(t)
Δu(t) =+ disturbance
m(z)w(t)
+l(z)Δu(t)
n(z)y(t)+++ are (αknown,
n(z)y(t)
l(z)Δu(t) +pthe (αffirst +term in (10)
f (z) + +l(z)Δu(t)
α (z))Δv(t)
(z) + α(αp (z))Δv(t)
f (z) + αp (z))Δv(t)
Equation (12) is also considered, v(t + j) for j = t + d + 1, . . . , t + d + N. In such a case, it is possible
with: m(z)
to use the predictive mechanismwith:
= kGm(z)
p
forzwith:
(z) = where
kG m(z)
disturbance 
ppz=z(z)
(z)kG where

=pz[z(z)
compensation pduz+1 where
(z) , du +2 du +1
zfor =p a[z . . . ,. range
,zwide
.(z) z. ,duz[z
= +2
dUdu
,+N+1 du +2 dU +N
. . ],.,systems,
of z.n(z)
. . , z,= .−kG
. . including
. ],. .n(z) and
, zdSU +N
= ],−kG n(z)=S
l(z) = and
−kG l(z
p . −kG
−kG Gproblems. Moreover, .αMoreover,
p and f. are,
Moreover,
p and
respectively,
αfαare, and respectively,
polynomials,
αf are, can respectively,
be polynomials,
including polynomials,
the including
coefficients including
the ofcoefficients
the the
pastcoefficien
of the
those with inversion  Moreover, 
Gp−kG Gαpthe αdeveloped p scheme easily used with different
feedback controllers,
(−kG H since
p ) it
and
(−kG only focuseson
future
H p )
(−kG(−kG
and  the disturbance
future
H H)
p ) and
values
(−kG futureof

H) the
(−kG compensation.
values
disturbance.

H) of values
the disturbance.
Depending
of the disturbance.
on
Depending
the relationDepending
on the
between
relation
on the
the between
relation
disturbance-output
disturbance-output
dead disturbance-output
time dvdead and time the process
dv and
dead timeinput-output
thedprocess
v and the input-output
deadprocess time,input-output
ddead
u , such time,
factors
dead
du , such
are
time, calculated
factors
du , such arefactors
calcul
Future Disturbance Estimation
in different ways.
in different
Forinthe ways.
different
configuration
Forways.the configuration
For
analyzedthe configuration
in analyzed
this work,in analyzed
ρthis> 0work, andin this ρ >work,
(nd +0 dand v )ρ> >1,0+
(nd the
anddvα)(nd> +
p,f 1, dthe
v)
In previous works, [8,9,24],
polynomialspolynomials a study
are as follows:
polynomials of different
are as follows: forecasting
are as follows: techniques was performed in order
to extend the functionality of the MPC approach. The availability of the future signal values are
used to improve the prediction mechanism αp (z) accuracy,
= αp1 α+p (z) αwhichz −1ααp+results
p2 = 1p (z)
+
. . .α=+p2 z in−1
α + .α
1better
p(nd+d
−(d
z −1αv +nd−1)
. z. performance
v )p2
+ p+ ... +
(nd+dv )
, z −(d
αof v +nd−1)−(dv +nd−1)
thev )MPC
p(nd+d z, ,
controller [9]. The previous studies show that relatively lowest prediction errors were reached using
and the future
the Double Exponential andpart
thehas
Smoothing future
and
following
the
partfuture
technique hasform:
following
and part has following
therefore form: form:
this forecasting method was also used in
the study performed in this work. u −dv ) −d
u −d
v ) v )+1(du −d
αf (z) = αf1 α z (d
f (z) = +
(du(d
ααfα1fzf(z)
1 z = αf +1 zαf+1 z.(d.v u). −d
+vα)+1 fN1 zz+ (duu−d
(d
. .−d v)+1
. v+ )+N
αf+N z. (d. .u+ −dv )+N (du −dv )+N
α fN z
The Double Exponential Smoothing (DES) technique is characterized by two equations:
118 where nd 118refers
whereto118nd
the where
refers
degreeto ofthe
nd refers
thedegree
polynomial
to the of the
degree Dpolynomial
from
of the equation
polynomial
D from (5).equation
Moreover,
D from(5). equation
polynomials
Moreover, (5). Moreover,
polynomials
referring polynom refer
to past α S = + ( 1 − )( S + b ) , (13)
p and to past
future
119αpto αandfpast
values
future
αp and
ofαfdisturbances
future
valuesαof f values
disturbances
can beofgrouped disturbances
can be intogrouped
αcan= beαinto pgrouped
+ α f .= into Additionally,
α α=f . αAdditiona
p + αf .
119 119 k φx k φ k − 1 k − 1 αp +
120 the prediction
120 themechanism
prediction
120 the prediction
of
mechanism
the intrinsicmechanism
of the compensator
intrinsic
of thecompensator
intrinsic
can be also compensator
used
can be to also
include
canused befuturealso
to include
used
disturbance
tofuture
include disturba
futur
bk = γ ( S k − S k − 1 ) + ( 1 − γ ) bk − 1 , (14)
121 estimations,
121 estimations,
v, through
121 estimations,
the
v, through
polynomial v,the through
polynomial
α f . In thesuch polynomial
αa f scheme,
. In such α the
f .a In
scheme,
implicit
such athe
feedfroward
scheme,
implicit the feedfroward
compensation
implicit feedfroward
compensa
where xk 122
is actual signal
is included
122 in isvalue,
control
included
122
Sis
law,
k inis the unadjusted
included
control
however in
law,the
control
however forecast,
disturbance
law, the
however bk isthe
cannot
disturbance the
be estimated
disturbance
completely
cannot be cannot trend,beφ completely
attenuated
completely isattenuated
as the shown
was attenuated
as inwas show as
smoothing123parameter for
[9,13]. This
123
data,
[9,13]. and
is due123toThisγ is
λ[9,13]. the
parameter
is dueThis smoothing
to λthat
isparameter
dueis used parameter
to λ parameter
inthat for
theisinternal
usedthat trend.
in structure
is
theused The
internal one
inofthe the sampling
structure
internal
GPC. of instance
Itstructure
was
the GPC. shown of Itthe
that
wasGPC.
for
shownIt was
thats
ahead prediction is obtained by x̂k+1 = Sk + bk and the m− sampling periods forecast is obtained
124 λ > 0, feedforward
124 λ > 0,124feedforward
compensator
λ > 0, feedforward compensator
will nevercompensatorperform
will never as awillperform
classical
neverasperform disturbance
a classical as acompensation
disturbance
classical disturbance compensation
structure. compensa
struct
usign x̂k+m = Sk + mbk . The initial values for Sk and bk were set to Sk = x1 and bk = ( x1 + x2 + x3 )/3
125 The ideal125 compensator
The ideal
125 The
compensator
canideal be obtained
compensator
can be only obtained
can
whenbeλonly obtained
= 0, whenas only was
λ =when shown0, asλin was
=[13].shown
0, asNevertheless,
was in shown
[13]. Nevertheless,
inthe [13]. Nev
as recommended in [8,9]. Moreover, the values for φ and γ ∈ (0, 1) can be obtained via optimization
126 resulting126control
techniques as described in [8].
resulting
scheme
126 control
resulting
is characterized
scheme
controlis schemecharacterized
by implicit is characterized
co-design
by implicitfor byco-design
implicit
feedbackco-design forandfeedback
feedforwardfor andfeedback
feedforward
part and and feedforpart
127 the high 127
bandwidth
the high127inbandwidth
the highfeedback bandwidth
in the loop, feedback
what
in theresultsloop,
feedback what
in very loop,
results
aggressive
what in very
results changes
aggressive
in very in control
aggressive
changes signal.
inchanges
control
Due in signal.
contr
128 to the implicit
128 to the
relation
implicit
128 tobetween
therelation
implicit
feedback
between
relation andfeedback
between
feedforward and
feedbackfeedforward
actions, andsuch feedforward
actions,
structure such
actions,
cannot structure
be
suchusedcannot
structure
as abe canno
used
129 stand alone
129 disturbance
stand alone
129 stand
compensator
disturbance
alone disturbance
compensator
what limits compensator
its
whatapplicability.
limitswhat its applicability.
limits
This particular
its applicability.
This issue
particular
canThis be particular
issue
addressed can be issue
addres
can
130 with predictive
130 withfeedforward
predictive
130 with predictive
feedforward
compensator feedforward
compensator
that was recently compensator
that was introduced
recently
that was inintroduced
recently
[12] and introduced
which
in [12]isand summarized
inwhich
[12] and is summar
which i
131 in following
131 in
section.
following
131 in following
section. section.
Energies 2018, 11, 123 7 of 17

Control Signal Saturation Constraints

Due to its operation principles, the predictive feedforward compensator is able to handle the
constraints into the optimization procedure. As shown perviously, the analyzed compensator is
derived from GPC structure and keeps the ability to handle the process constraints in an independent
way. This characteristic is important from a practical point of view since, in a classical feedforward
compensator, this issue is not considered. Nevertheless, due to the fact that feedforward compensator
works as an independent block from the main feedback control loop, it is necessary to take into account
the value of the control signal computed by the feedback controller. In such a way, the available change
in the control action is determined and considered in the optimization procedure.
Due to the physical limitations of control signal (actuator saturation limits, umin ≤ u(t) ≤ umax ),
it is necessary to include this data into an optimization procedure, where control signal u(t − 1) =
u FB (t − 1) + u FF (t − 1) considers the part that corresponds to the feedback controller u FB and u FF from
the feedforward compensator. In a general way, these constraints can be expressed as R∆u 6 c where:
" # " #
T lumax − lu(t − 1)
R= ;c = ,
−T −lumin + lu(t − 1)

and T is N × N lower triangular matrix of ones, l is a vector, 1 × N of ones. In such a way, saturation
limits can be expressed as a function of inequalities on control signal increments:

lumin ≤ T∆u + u(t − 1) ≤ lumax .

3. Microalgae Production in a Raceway Reactor

The experimental raceway reactor used for modelling purposes is located at experimental station
“Las Palmerillas” (Almería, Spain—property of CAJAMAR Foundation). The modelled reactor has
100 m2 surface area and is composed of two 50-m long (1 m wide) channels forming U-shape bends
(see Figure 2) [25]. The reactor operates at a constant depth (0.2 m) to provide desired performance
and considering power consumption issues. Resulting reactor volume is 20 m3 . The raceway
photobioreactor facilities can be set up to use different carbon dioxide sources in order to provide
a flexible platform to evaluate different systems and techniques. Nevertheless, for the modelling
purposes, the plant was configured to use flue gases.
The flue gas used for pH control was taken from industrial heating boiled (diesel-fueled with the
average gas composition was: 10.6% CO2 , 18.1 ppm CO, 38.3 ppm NOx, and 0.0 ppm SO2 ). In this
operation scheme, exhaust gas was refrigerated to the environment temperature and stored in a 1.5 m3
pressure tank (compressed to 2 bar) with automatic regulation of pressure. The compressed flue gas was
supplied to the reactor through a 150 m pipeline of 40 mm diameter. Finally, the injection system uses
the solenoid on/off valve and input flow was measured with a digital flow meter (detailed infirmation
for raceway setup can be found in [17,25]). The injection instant as well as aperture time is provided
by the control algorithm used for pH control.
Taking into account the limitation of the actuation system that is used in the pilot-scale raceway
reactor, it is necessary to convert the continuous control signal into the on/off actions of solenoid
control valve. To this end, the PWM (Pulse Width Modulation) technique is used to translate the signal
provided by the controller into a train of pulses with modulated width. The value of the pulse width
is determined by the control system and can vary between 0 and 100%. Moreover, the modulation
frequency was set to 0.1 Hz being optimal for the main controlled variable. The next section summarizes
the raceway reactor modelling that is presented based on the development from [21].
Energies 2018, 11, 123 8 of 17

50 m

0.65 m

1m
<<<<

Sump Paddlewheel

0.3 m

1m
(a)

(b)

Figure 2. Raceway reactor. (a) scheme (top and side view); (b) modelled experimental facilities.

3.1. Raceway Reactor Dynamics and Models

The culture growth can be modelled as a function of the photosynthesis rate. The main parameter
that determines the photosynthesis rate is the available light, based on external irradiance, culture
characteristics and reactor geometry [26,27]:

I0 (t)
Iav (t, x ) = (1 − exp(−Ka Cb (t, x )h), (15)
Ka Cb (t, x )h

where t is the time, x the space, I0 is the solar irradiance on an obstacle-free horizontal surface, Ka is
the extinction coefficient, Cb is the biomass concentration, and h is the liquid height on the channels.
The available average irradiance is correlated with the photosynthesis rate by a hyperbolical
function as proposed in [28]. This function is completed by adding the rest of the factors that limit
the microalgal growth (under sufficient conditions of nutrients). Thus, the influence of the pH culture
value and dissolved oxygen of the culture have been modeled as described in [15]:

PO2,max Iav (t, x )n

PO2 (t, x ) = (1 − αs )
Ki exp( Iav (t, x )m) + Iav (t, x )n
  z 
[O2 ](t, x )
1− (16)
KO2
−C1 −C2
    
B1 exp − B2 exp − αs RO2 ,
pH (t, x ) pH (t, x )

where PO2 is the photosynthesis rate (oxygen production rate per biomass mass unit), PO2,max is the
maximum photosynthesis rate for microorganisms under the culture conditions, n is the form exponent,
and the term in the denominator is the irradiance constant, which increases as an exponential function
Energies 2018, 11, 123 9 of 17

of average irradiance, Ki and m being form parameters of this relationship, KO2 is the oxygen inhibition
constant and z is a form parameter. For the pH influence on the photosynthesis rate, B1 , B2 are
the preexponential factors and C1 , C2 the activation energies of the Arrhenius model. Furthermore,
a constant respiration rate RO2 was included in order to represent the respiration phenomenon,
and a solar distributed factor αs as the shadow projection on the perpendicular axis of the reactor walls.
The pH value of the culture is related to other species such as dissolved carbon dioxide, [CO2 ],
carbonate, [ HCO3− ], or bicarbonate, [CO32− ] by several equilibrium equations, as can be seen in [29],
the balance of one of them being necessary to obtain predictions of pH along time and space. In this
work, the total inorganic carbon concentration, [CT ], is modelled taking into account the photosynthesis
process performed by the microalgae culture, and the transport phenomena due to the recirculation
of the culture along the raceway. Assuming constant velocity, v, and constant cross-sectional area
obtained by the multiplication between the liquid height, h, and the channel width, w:

∂[CT ](t, x ) ∂[CT ](t, x ) PCO2 (t, x )Cb(t, x )

wh = −whv + wh
∂t ∂x MCO2 (17)
+ whKlaCO2c ([CO2∗ ](t, x ) − [CO2 ](t, x )),

where PCO2 is the carbon consumption rate, KlaCO2c is the mass transfer coefficient for CO2 , and MCO2
is the molecular weight of carbon dioxide. The total inorganic carbon, [CT ], is related to the carbon
dioxide concentration in the liquid phase [CO2 ] and the equilibrium concentration in the gas phase
[CO2∗ ]. The equilibrium concentration can be calculated, according to Henry’s law, taking into account
the CO2 properties in the air.
Regarding dissolved oxygen concentration, a homologous balance can be established as follows:

∂[O2 ](t, x ) ∂[O2 ](t, x ) PO (t, x )Cb(t, x )

wh = −whv + wh 2
∂t ∂x MO2 (18)
+ whKlaO2c ([O2∗ ](t, x ) − [O2 ](t, x )),

where PO2 is the photosynthesis rate (oxygen production rate per biomass mass unit), MO2 is the
molecular weight of oxygen, KlaO2c is the volumetric gas-liquid mass transfer coefficient for oxygen at
the channels, and ([O2∗ ] − [O2 ]) is the driving force. The equilibrium concentration in gas phase [O2∗ ] is
calculated as a function of the oxygen concentration in the gas phase based on Henry’s law.
Analogous mass balances are applied at the paddle-wheel and the sump of the reactor, bearing in
mind that these parts can be represented by ODE (Ordinary Differential Equation) expression.
In addition to the liquid phase, the sump is designed to incorporate carbon dioxide by means of
CO2 injections in gaseous form, and remove dissolved oxygen accumulated into the channels by air
injections. Therefore, mass balances on the gas phases are needed to include these phenomena in the
model. Since the nitrogen molar flow can be considered constant because its solubility is approximately
zero, the balances presented in this paper are formulated by relations from the rest of gases to nitrogen
molar ratio. Owing to oxygen, the next balance (19) can be established:

dYO2,out (t) Q gas

=− (YO2,out (t) − YO2,in (t))
dt Vs (1 − ε s (t))
(19)
V (1 − ε s (t))
− KlaO2s mol ([O2∗ ](t) − [O2 ](t))lm ,
y N2 ε s (t)

where Vmol is the molar volume under reactor conditions (pressure and temperature), YO2 is the oxygen
to nitrogen molar ratio in the gas phase, which is defined at the inlet and outlet of the sump, and y N2 is
Energies 2018, 11, 123 10 of 17

the nitrogen molar fraction. For the carbon dioxide, an analogous mass balance can be defined (20),
where YCO2 is the carbon dioxide to nitrogen molar ratio in the gas phase:

dYCO2,out (t) Q gas

=− (YCO2,out (t) − YCO2,in (t))
dt Vs (1 − ε s (t))
(20)
V (1 − ε s (t))
− KlaCO2s mol ([CO2∗ ](t) − [CO2 ](t))lm .
y N2 ε s (t)

3.2. The pH Control System—Process Model, Control Design and Implementation

For the GPC-based controllers design, a linear model for the controlled process is required. In the
analyzed system, the pH value is affected mainly by CO2 injections and CO2 uptake as a function
of photosynthesis. The injected CO2 makes a possible formation of carbonic acid, which decreases
the pH value. On the contrary, the gradual increase in pH is provoked by the photosynthesis process
where CO2 is consumed and O2 is produced. As a consequence, the changes in solar irradiance result
in photosynthesis rate variation, which affects the medium pH value. In the presented approach, it is
considered that pH culture is the process output, the aperture of injection valve is the process input
and solar irradiance is the measurable system disturbance. Such a relationship can be characterized by
simplified linear models [14,30].
The raceway reactor pH process was modelled using a linear reduced-order model. The model
structure is developed considering distribution of the actuators and sensor, the reactor architecture
and dominant dynamics in measured data. The developed model relates the carbon dioxide injections
(input variable) with pH (output variable) and is given by following structure [14,30]:

k1 kr
pH (s) = e − tr s u ( s ) + v ( s ), (21)
1 + τ1 s 1 + τr s
| {z } | {z }
Pu Pv

where pH is the culture medium pH, u is the control variable, and v is the solar radiation. A first
term, Pu , relates the pH level to CO2 injections, tr , refers to a time delay between the CO2 injection
point and the pH measure point. The Pv term captures the solar radiation effect on pH value.
Through process identification and validation procedures, the following parameters were obtained:
k1 = −0.005 pH%−1 , τ1 = 16.5 min, tr = 7 min, kr = 0.0007 pH m2 W−1 , and τr = 118 min. Due to a
low complexity of the model, there exist modelling errors, which makes the presented approach more
realistic from a practical point of view [14,30]. Once the simplified models for pH control have been
obtained, the corresponding control schemes are built.
The simulations presented in this work have been performed using the raceway
photobioreactor nonliner model developed by [21] and the control system implemented in
Matlab/Simulink—Mathworks c (version 2013R, Mathworks, Natick, MA, USA) environments.
The balances based on partial and ordinary differential equations, which establish the base of the
model, have been implemented by code in Matlab. The control system has been implemented also in
Matlab and coupled with the nonlinear model in the Simulink environment.

4. Results
This section shows simulation experiments using the proposed predictive feedforward
compensator coupled with different feedback controllers and applied to pH control in raceway
photobioreactors. The predictive feedforward compensator was tested in simulation, where the
nonlinear model described in Section 3 was used as virtual plant. All data used in the simulation
study were collected from the real photobioreactor operating in continuous mode at real conditions
in Almería (Spain) during spring 2017. The simulations were performed for a period of seven days,
where different profiles of solar irradiance have been considered, since it is the main disturbance
that affects the pH control process during diurnal conditions (through the photosynthesis process).
Energies 2018, 11, 123 11 of 17

Moreover, in all considered control systems, the continuous signal obtained from the controller must
be translated into a discontinuous signal used to drive the valve. For this purpose, a PWM technique is
used, with a frequency of 0.1 Hz. Notice that microalgae growth requires that the operation variables
must be maintained at optimum values, where for the microalga used in this work an optimal value of
w(t) = 8 is required.
In this study, two different feedback control strategies, namely PI and GPC controllers,
have used the influence of feedforward action. Notice that the objective of this study is not a direct
comparison of the control schemes, but the analysis of feedforward predictive compensator and its
influence on control performance in different control approaches. The effectiveness of the predictive
feedforward compensator is analyzed through a simulation study where different approaches
for measurable disturbances compensation are considered. The PI controller is tested in three
configurations: self-contained (with no specific disturbance compensation technique), coupled with
classical feedforward (PI + CFF) and predictive feedforward compensator (PI + PFF). The GPC-based
approaches include: stand-alone configuration (with no specific disturbance compensation technique),
GPC with intrinsic feedforward action (GPC + IFF) and coupled with predictive feedforward
compensator (GPC + PFF).
The PI feedback controller was tuned according to the SIMC (Skogestad Internal Model Control)
design rule [31], and results in K = −235 and Ti = 990. Moreover, the CFF compensator was
obtained using the methodology presented in Section 2 obtaining the following transfer function:
in CFF = (0.0007 + 0.01155s)/(−0.005 − 0.59s), where a non-realizable part was omitted (due to a
dead time inversion problem). The GPC-based feedback controllers were set up to satisfy required
performance resulting in: N = 10, Nu = 5, δ = 1, λ = 0.05. For GPC + IFF configuration,
all parameters are the same as for the stand-alone variant and only measurable disturbances are
considered during the development of the algorithm (see Section 2 for details). The PFF compensator
parameters were set to N = 7, Nu = 1, δ = 1, λ = 0 (these parameters were kept unchanged
for all configurations). Additionally, the future disturbance estimations were obtained using a
DES technique and its configuration parameters were set to φ = 0.99 and γ = 0.1 following
recommendations from [13]. All controllers have been implemented with sampling time Ts = 1
min. Moreover, due to the physical limitation of the actuator, the control signal has been saturated
between 0–100% and considered in an optimization procedure for all GPC-based approaches including
the PFF compensator. For the PI-based configuration, the antiwindup technique was exploited to deal
with control signal saturation.
The control performance for each analyzed configuration was determined using the
following indexes: Z ∞
I AE = e(t)dt, (22)
0
Z ∞
I AC = u(t)dt, (23)
0

CSE = ||∆u(t)||, (24)

where I AE represents Integrated Absolute Error between reference and controlled variable, the I AC
refers to Integrated Absolute Control signal and CSE stands for Control System Effort.
Moreover, three complementary indexes are included for biomass concentration Cb ,
photosynthesis rate PO2 , and total amount of flue gases Gas supplied to the raceway photobioreactor.
The first two are obtained from the nonlinear process model and are used to demonstrate the influence
of the evaluated algorithm on microalgae growth. The former index shows the usage usage of the
control medium (flue gases), which can be directly related with energy consumption. The bigger value
means that more energy was consumed for gas compression, being less effective from a economic
point of view.
Since a graphical result for seven days and all analyzed control techniques will not allow one to
see the results properly, one representative day has been selected for PI and GPC feedback controllers,
Energies 2018, 11, 123 12 of 17

which are shown in Figures 3 and 4, respectively. These figures show the results of different feedforward
approaches coupled with two feedback controllers in order to check the efficiency of disturbance
compensation for pH process in a raceway photobioreactor. For each feedback control technique,
different disturbance compensation approaches are tested and all other parameters and variables are
kept unchanged. In such a configuration, the eventual changes in control performance are due to the
efficacy of the disturbance attenuation, which permits us to compare different feedforward schemes.

8.2
setpoint PI PI+CFF PI+PFF
pH

7.8
8 10 12 14 16 18

Control Signals [%]

20

10
u

0
8 10 12 14 16 18
20
uFB

10

0
8 10 12 14 16 18

10
uFF

0
8 10 12 14 16 18
Solar Irradiance
real DES
[W/m2]

8 10 12 14 16 18
t [hours]

Figure 3. Control results for PI (Proportional-Integral) as feedback controller and different

feedforward approaches.

Figure 3 shows the control results for the configuration where the PI controller is used in feedback
loop. For this case, three configurations have been evaluated where disturbance compensation
was performed using: feedback controller (PI), classical feedforward compensator PI + CFF and a
predictive feedforward compensator PI + PFF. From the first plot in Figure 3, it can be seen that all
tested schemes are able to maintain the pH level close to its set point. Nevertheless, the configurations
using the dedicated disturbance compensator, PI + CFF and PI + PFF, obtain better precision.
In these two configurations, the control signal is composed of two parts, u FF and u FB , obtained from
feedforward and feedback controllers, respectively. The u FF component provided depends on the
solar irradiance, and proper structure of the compensator. Moreover, from the u FF plot, it can be
observed that significant differences between PI + CFF and PI + PFF are obtained. This is due to a
predictive approach used in PFF that considers dead time inversion and exploits the future disturbance
information estimated using the DES technique. In such a case, the estimated future values of the
solar irradiance (see red plot on bottom graph in Figure 3) has an offset in comparison to the real solar
irradiance signal. Notice that, in the schemes with feedforward compensators, the resulting control
signal from the compensator is applied only if the pH level is bigger than 7.9; otherwise, it is set to
zero. This condition is set to prevent CO2 injections when pH is below the set point.
From the obtained results, it can be seen that none of the analyzed compensators are able
to completely attenuate the disturbance effect. Even the PFF approach with future disturbance
compensation does not achieve the complete attenuation. This is mainly due to two reasons. The first
one is related with plant-model mismatch and the second can be linked with future disturbance
Energies 2018, 11, 123 13 of 17

estimation errors. Both factors are making the predictive disturbance compensator less effective in
comparison to the nominal case where no modelling and estimation errors are considered. However,
the PFF performance is better in comparison to the classical approach, which is confirmed by the
indexes related to control accuracy and process performance.
The average performance indexes for the previously analyzed configurations are summarized in
Table 1. The I AE index shows that the best accuracy, for the analyzed seven-day period, is obtained
for PFF. On the other hand, the CFF approach improves the I AE measure in comparison to the
configuration where only the feedback controller is used to compensate the disturbances. Additionally,
the PFF configuration obtains the lowest I AU measure, being the most effective in control resource
utilization. The improved accuracy of the approaches with feedforward compensators is obtained at
the expense of higher control signal variability, which is shown by the CSE index where PFF obtains
the biggest value. Besides that, the predictive feedforward compensator uses the lowest amount of the
flue gases (see Gas index) from all tested configurations, being the most energy efficient.

Table 1. Average performance indexes for PI as feedback controller and different configurations of
feedforward control schemes for the seven-day period.

Feedforward I AE I AU CSE Gas Cb PO2

Scheme [−]103 [−]104 [−] [m3 ] [kg/m3 ] [kgO2 /m3 s]
- 4.983 3.773 147 6.28 0.499 0.0278
CFF 4.872 3.706 193 6.17 0.514 0.0285
PFF 4.713 3.421 212 5.96 0.529 0.0301

From the process performance point of view, application of feedforward compensators improves
the overall productivity of the biomass, which is confirmed through Cb and PO2 measures. The PFF
configuration provides the best growth conditions, keeping the pH value close to its reference for
a longer time. The photosynthesis rate increments were shown using PO2 measure provided by
the nonlinear model. This index reflects the oxygen concentration and it is proportional to the
photosynthesis rate. The PO2 increased 5.6% for the PFF configuration when compared to the CFF
technique. Similar improvements are obtained in terms of energy savings. The changes in energy
required for control purposes are calculated using the Gas index. This measure expresses the volume of
the flue gases used during the seven-day time period. The energy usage for the microalgae production
in the raceway reactor is mainly related to the energy consumed by the flue gas compressor. Based on
this issue, a direct relation between Gas and energy used can be established. The reduced amount of
gas usage by the control scheme has a proportional impact on the energy savings. In this case, the PFF
based scheme reduced the energy usage about 4.5% regarding the CFF. These facts are crucially
important for biomass production from microalgae since they increase the process rentability.
The results for the second control scenario, where the GPC algorithm is used as a feedback
controller are shown in Figure 4. As in the previously analyzed scheme, the PFF compensator combined
with the GPC controller provides the most accurate control of the pH. In such a configuration, the pH
value is controlled close to its reference and the signal is characterized by reduced oscillations. The GPC
standard intrinsic disturbance compensator, IFF, is able to improve the performance of the stand-alone
GPC, but, due to design compromises (co-design trade-off between feedback and feedforward), it is
less effective than the proposed predictive feedforward compensator. This fact is even more visible
when the stand-alone GPC configuration is analyzed. In this case, the disturbance compensation is
performed through a feedback loop, since no additional compensator is included. The disturbance
compensation oriented design (allowing quick change in control signal) results in overshoot in the
initial stage of the day [20]. Notice that, in all analyzed configurations, the feedback controller has the
same tuning, and the disturbance compensation technique consideration improves the performance.
In such a case, the best accuracy is obtained for the GPC + PFF configuration. When the control signal
is analyzed for all tested configurations (plots two, three and four in Figure 4), it can be observed
Energies 2018, 11, 123 14 of 17

that the predictive feedforward compensator is starting before other configurations. This is mainly
due to the predictive mechanism that is able to handle future disturbance providing the disturbance
attenuation before its effect appears on the controlled variable. Additionally, the possibility for optimal
adjustment of the control signal weighing factor λ is another element that has influence on control
signal changes. It should be highlighted that PFF configuration in nominal cases (no modelling and
disturbance estimation errors) should be able to attenuate completely the effect of disturbance [12].
In industrial practice, it will be impossible to meet this condition and consequently one can not expect
perfect compensation of the measurable disturbances. Nevertheless, significant improvements in
control performance regarding classical approach can be obtained as shown in the analyzed case.

8.2
setpoint GPC GPC+IFF GPC+PFF
pH

7.8
8 10 12 14 16 18

Control Signals [%]

20

10
u

0
8 10 12 14 16 18
20
FB

10
u

0
8 10 12 14 16 18

10
FF
u

0
8 10 12 14 16 18
Solar Irradiance
real DES
[W/m2]

8 10 12 14 16 18
t [hours]

Figure 4. Control results for GPC as the feedback controller and different feedforward approaches.

Table 2 summarizes the control performance indexes for a scheme where the GPC controller
is used in the feedback loop. In this case, the lowest value of I AE is obtained by the GPC + PFF.
The IFF scheme improves stand-alone GPC, but it is less accurate in pH regulation than the predictive
feedforward compensator. Moreover, it can be observed that configurations with disturbance
compensation mechanism reduce the control effort, I AU, which has positive impact on the CO2
usage. For the CSE index, similar behavior as the previously analyzed scheme was observed.
The configurations with disturbance compensation technique provide the control signal that is
characterized by increased variability, whereas the control signal from PFF obtains the biggest CSE
index value. From all tested configurations, the best growth conditions were obtained for PFF,
which improves PO2 index about 6.4% in comparison to the IFF. Moreover, stand-alone GPC was
characterized by the biggest consumption of the control resources (Gas). The smallest usage of flue
gases was obtained by PFF configuration. As a consequence, the predictive feedforward compensator
is the most energy efficient approach, reducing it about 3.8% in comparison to the IFF scheme. This fact
is of great importance in raceway photobioreactor biomass production, since it allows the possibility
to reduce the process maintenance costs.
Energies 2018, 11, 123 15 of 17

Table 2. Average performance indexes for GPC as the feedback controller and different configurations
of feedforward control schemes for a seven-day period.

Feedforward I AE I AU CSE Gas Cb PO2

Scheme [−]103 [−]104 [−] [m3 ] [kg/m3 ] [kgO2 /m3 s]
- 4.741 3.522 25.4 5.89 0.521 0.0297
IFF 4.672 3.485 33.7 5.80 0.547 0.0309
PFF 4.538 3.397 58.5 5.64 0.571 0.0329

On the other hand, application of measurable disturbance compensation techniques has a positive
influence on the photobioreactor productivity, which is represented by the Cb measure. From this
index, it can be observed that the highest productivity is achieved for PFF, improving the stand-alone
GPC algorithm as well as IFF configuration. In such a case, Cb has increased about 9%. This improved
productivity is obtained with lower control resources utilization having a positive impact on the
process performance.
From the obtained results, for both feedback control techniques, it can be deduced that the
predictive feedforward compensator can provide better performance than the classical approach.
Nevertheless, its final performance will depend on plant model mismatch and disturbance estimation
accuracy. As shown in the previous analysis, the predictive feedforward compensator can handle
measurable disturbance in a more efficient way in comparison to the classical feedforward compensator
as well as GPC with intrinsic compensation. This realistic evaluation considered aforementioned
issues for the pH control problem in the raceway photobioreactor, where important benefits for
this particular process were obtained. Independently on the feedback controller, the application of
predictive feedforward compensator results in an improved photosythesies rate, providing an average
increase of 6% , and where the biomass concentration was increased about 4%. At the same time,
the average energy usage was reduced about 4%.

5. Conclusions
In this paper, the evaluation of a predictive feedforward compensator and its combination
with the PID and the MPC techniques were presented. The analysis was performed through a
simulation study, where the pH control problem in a raceway photobioreactor was considered. To show
the properties of the analyzed technique, several feedforward compensators have been considered.
Through those configurations, it was demonstrated that significant improvements were obtained in
control accuracy when a predictive feedforward compensator is used. In such a case, the effectiveness
of the analyzed feedforward scheme was verified, obtaining significant improvements in control system
performance indexes when compared to classical compensators. The positive impact of predictive
feedforward compensator was also confirmed through microalgae growth performance indicators.
For this particular process, it was possible to improve the photosynthesis rate about 6% and biomass
concentration was increased about 4%. Moreover, the predictive compensator provides a possibility
to reduce the energy required for process operation while simultaneously providing a more efficient
way for the flue gas usage (used as CO2 source). The average energy demand was decreased about 4%
when a predictive compensator was used.
Based on the obtained results, it was demonstrated that application of a predictive feedforward
compensator has a positive impact on the analyzed process performance, even in the presence of
modelling and disturbance estimation errors.

Acknowledgments: This work has been partially funded by the following projects: DPI2014-55932-C2-1-R,
DPI2014-55932-C2-2-R, DPI2014-56364-C2-1-R and DPI2012-31303 financed by the Spanish Ministry of Economy
and Competitiveness and EU-ERDF (European Regional Development Fund) funds); and the UNED (Universidad
Nacional de Educación a Distancia) through a postdoctoral scholarship.
Energies 2018, 11, 123 16 of 17

Author Contributions: All authors have participated in the research summarized in this manuscript,
providing development ideas, establishing design methodology, as well as evaluation strategy and analysis
of obtained results. All authors have been involved in preparation of the manuscript.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Åström, K.; Hägglund, T. Advanced PID Control; ISA Press: Research Triangle Park, NC, USA, 2006.
2. Guzmán, J.L.; Hägglund, T. Simple tuning rules for feedforward compensators. J. Process Control 2011,
21, 92–102.
3. Veronesi, M.; Visioli, A. Automatic tuning of feedforward controllers for disturbance rejection. Ind. Eng.
Chem. Res. 2014, 53, 2764–2770.
4. Hast, M.; Hägglund, T. Design of optimal low-order feedforward controllers. In Proceedings of the 2nd
IFAC Conference on Advances in PID Control, Brescia, Italy, 28–30 March 2012.
5. Rodríguez, C.; Normey-Rico, J.E.; Guzmán, J.L.; Berenguel, M. Robust design methodology for simultaneous
feedforward and feedback tuning. IET Control Theory Appl. 2016, 10, 84–94.
6. Li, X.; Liu, S.; Tan, K.K.; Wang, Q.G.; Cai, W.J. Predictive feedforward control. In Proceedings of the 12th
IEEE International Conference on Control & Automation (ICCA), Kathmandu, Nepal, 1–3 June 2016.
7. Mercorelli, P. A switching Kalman Filter for sensorless control of a hybrid hydraulic piezo actuator using
MPC for camless internal combustion engines. In Proceedings of the IEEE International Conference on
Control Applications, 3–5 October 2012; pp. 980–985.
8. Pawlowski, A.; Guzmán, J.L.; Rodríguez, F.; Berenguel, M.; Sánchez, J. Application of time-series methods to
disturbance estimation in predictive control problems. In Proceedings of the IEEE Symposium on Industrial
Electronics, Bari, Italy, 4–7 July 2010.
9. Pawlowski, A.; Guzmán, J.L.; Rodríguez, F.; Berenguel, M.; Normey-Rico, J.E. Predictive control with
disturbance forecasting for greenhouse diurnal temperature control. In Proceedings of the 18th World
Congress of IFAC, Milan, Italy, 28 August–2 September 2011.
10. Elso, J.; Gil-Martínez, M.; García-Sanz, M. Quantitative feedback-feedforward control for model matching
and disturbance rejection. IET Control Theory Appl. 2013, 7, 894–900.
11. Shridhar, R.; Cooper, D.J. A tuning strategy for unconstrained SISO Model Predictive Control. Ind. Eng.
Chem. Res. 1997, 36, 729–746.
12. Pawlowski, A.; Rodríguez, C.; Guzmán, J.L.; Berenguel, M.; Dormido, S. Predictive feedforward
compensator for dead-time processes. In Proceedings of the 20th World Congress of IFAC, Toulouse,
France, 20 October 2017.
13. Pawlowski, A.; Guzmán, J.L.; Normey-Rico, J.E.; Berenguel, M. Improving Feedforward Disturbance
Compensation Capabilities in Generalized Predictive Control. J. Process Control 2012, 22, 527–539.
14. Pawlowski, A.; Mendoza, J.L.; Guzmán, J.L.; Berenguel, M.; Acién, F.G.; Dormido, S. Selective pH and
dissolved oxygen control strategy for a raceway reactor within an event-based approach. Control Eng. Pract.
2015, 44, 209–218.
15. Costache, T.A.; Acién, F.G.; Morales, M.M.; Fernández-Sevilla, J.M.; Stamatin, I.; Molina, E. Comprehensive
model of microalgae photosynthesis rate as a function of culture conditions in photobioreactor.
Appl. Microbiol. Biotechnol. 2013, 97, 7627–7637.
16. Pawlowski, A.; Fernández, I.; Guzmán, J.L.; Berenguel, M.; Acién, F.G.; Dormido, S. Event-based Selective
Control Strategy for Raceway Reactor: A simulation study. In Proceedings of the 11th IFAC Symposium on
Dynamics and Control of Process Systems, including Biosystems—DYCOPS-CAB 2016, Trondheim, Norway,
6–8 June 2016.
17. De Godos, I.; Mendoza, J.L.; Acién, F.G.; Molina, E.; Banks, C.J.; Heaven, S.; Rogalla, F. Evaluation of carbon
dioxide mass transfer in raceway reactors for microalgae culture using flue gases. Bioresour. Technol. 2014,
153, 307–314.
18. Pawlowski, A.; Mendoza, J.L.; Guzmán, J.L.; Berenguel, M.; Acién, F.G.; Dormido, S. Effective utilization of
flue gases in raceway reactor with event-based pH control for microalgae culture. Bioresour. Technol. 2014,
170, 1–9.
Energies 2018, 11, 123 17 of 17

19. Bernard, O. Hurdles and challenges for modelling an control of microalgae for CO2 mitigation and biofuel
production. J. Process Control 2011, 21, 1378–1389.
20. Pawlowski, A.; Rodríguez, C.; Guzmán, J.L.; Berenguel, M.; Dormido, S. Measurable disturbances
compensation: Analysis and tuning of feedforward techniques for dead-time processes. Processes 2016, 4, 12.
21. Fernández, I.; Acién, F.G.; Guzmán, J.L.; Berenguel, M.; Mendoza, J.L. Dynamic Model of an Industrial
Raceway Reactor for Microalgae Production. Algal Res. 2016, 17, 67–78.
22. Guzmán, J.L.; Hägglund, T.; Åström, K.J.; Dormido, S.; Berenguel, M.; Piguet, Y. Feedforward control
concepts through interactive tools. In Proceedings of the 18th IFAC World Congress, Milano, Italy,
28 August–2 September 2011.
23. Camacho, E.F.; Bordóns, C. Model Predictive Control; Springer: London, UK, 2007.
24. Hernández-Hernández, C.; Rodríguez, F.; Moreno, J.C.; da Costa Mendes, P.R.; Normey-Rico, J.E.;
Guzmán, J.L. The Comparison Study of Short-Term Prediction Methods to Enhance the Model Predictive
Controller Applied to Microgrid Energy Management. Energies 2017, 10, 884.
25. Mendoza, J.L.; Granados, M.R.; de Godos, I.; Acién, F.G.; Molina, E.; Banks, C.J.; Heaven, S. Fluid-dynamic
characterization of real-scale raceway reactors for microalgae production. Biomass Bioenergy 2013, 54, 267–275.
26. Acién, F.G.; García, F.; Chisti, Y. Photobioreactors: Light regime, mass transfer, and scaleup. Prog. Ind. Microbiol.
1999, 35, 231–247.
27. Acién, F.G.; Fernández, J.M.; Molina, E. Photobioreactors for the production of microalgae. Rev. Environ.
Sci. BioTechnol. 2013, 12, 131–151.
28. Molina, E.; Sevilla, J.M.; Pérez, J.A.; Camacho, F.G. A study on simultaneous photolimitation and
photoinhibition in dense microalgal cultures taking into account incident and averaged irradiances.
J. Biotechnol. 1996, 45, 59–69.
29. Camacho, F.; Acién, F.; Sánchez, J.; García, F.; Molina, E. Prediction of dissolved oxygen and carbon dioxide
concentration profiles in tubular photobioreactors for microalgal culture. Biotechnol. Bioeng. 1999, 62, 71–86.
30. Berenguel, M.; Rodríguez, F.; Acién, F.G.; García, J.L. Model predictive control of pH in tubular
photobioreactors. J. Process Control 2004, 14, 377–387.
31. Skogestad, S. Simple analytic tuning rules for model reduction and PID controller tuning. J. Process Control
2003, 13, 291–309.

c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).