Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: The paper provides an extensive experimental analysis of offset-free multivariable robust model predic-
Received 11 April 2020 tive control (MPC) of a neutralisation plant. The behaviour of the neutralisation plant was nonlinear and
Revised 30 May 2020
asymmetric. The controlled variable was a pH value of the outlet stream, and the manipulated variables
Accepted 30 July 2020
were the inlet volumetric flow rates of the acid and base. The observer was designed to estimate the
Available online xxx
system states. Operating conditions make the tuning of the robust controller a challenging task. First, the
Recommended by Prof. T Parisini conventional robust MPC was designed. The control setup was tuned and experimentally analysed in vari-
ous operating conditions. Based on the measured data, the novel approach of the real-time compensation
Keywords:
Neutralisation plant of the asymmetric behavior was investigated. Implementation of the proposed control strategy leads to
Robust model predictive control improved control performance. The real-time compensation of asymmetric behaviour can be considered
Real-time compensation of asymmetric for the control of other systems with nonlinear behaviour.
behaviour
© 2020 European Control Association. Published by Elsevier Ltd. All rights reserved.
Linear matrix inequalities
https://doi.org/10.1016/j.ejcon.2020.07.012
0947-3580/© 2020 European Control Association. Published by Elsevier Ltd. All rights reserved.
Please cite this article as: J. Oravec, M. Horváthová and M. Bakošová, Multivariable robust MPC design for neutralisation plant: Experi-
mental analysis, European Journal of Control, https://doi.org/10.1016/j.ejcon.2020.07.012
JID: EJCON
ARTICLE IN PRESS [m5G;August 21, 2020;11:47]
2 J. Oravec, M. Horváthová and M. Bakošová / European Journal of Control xxx (xxxx) xxx
Please cite this article as: J. Oravec, M. Horváthová and M. Bakošová, Multivariable robust MPC design for neutralisation plant: Experi-
mental analysis, European Journal of Control, https://doi.org/10.1016/j.ejcon.2020.07.012
JID: EJCON
ARTICLE IN PRESS [m5G;August 21, 2020;11:47]
J. Oravec, M. Horváthová and M. Bakošová / European Journal of Control xxx (xxxx) xxx 3
Please cite this article as: J. Oravec, M. Horváthová and M. Bakošová, Multivariable robust MPC design for neutralisation plant: Experi-
mental analysis, European Journal of Control, https://doi.org/10.1016/j.ejcon.2020.07.012
JID: EJCON
ARTICLE IN PRESS [m5G;August 21, 2020;11:47]
4 J. Oravec, M. Horváthová and M. Bakošová / European Journal of Control xxx (xxxx) xxx
Table 1 weak base leads to the equivalence point placed below the neutral
Identified parameters of the neutralisation plant.
pH value. Therefore, the presence of different types of reactants
Value ZA [s/mL] ZB [s/mL] TA [s] TB [s] DA [s] DB [s] would change the parameters of the mathematical model ZA , ZB ,
Minimum −2.7 0.7 92.9 94.6 2.9 1.8 TA , TB . Further technical details of identification are described in
Nominal −1.5 1.5 137.9 134.2 4.1 3.3 [27].
Maximum −0.6 2.3 182.8 174.0 5.3 4.7 The mathematical model was obtained in the form of 2 trans-
fer functions having the form as described in (3). These transfer
functions were transformed into the discrete-time domain uncer-
where [H+ ] is the concentration of hydrogen cations. The pH value tain state-space system
(pH) is a number within interval [1,14]. If pH value is equal to 7, x(k + 1 ) = A(v ) x(k ) + B(v ) u(k ), x ( 0 ) = x0 , (4a)
then the solution is neutral, pH value below 7 indicates an acidic
solution, and solution with pH value greater than 7 has alkaline
properties. The real-time values of pH in a reaction vessel were y(k ) = C x(k ), (4b)
measured by a pH probe, see Fig. 1, (II).
The manipulated variables were inlet flow rates of the acid so-
lution qA , and the base solution qB . The peristaltic Pump A (Fig. 1, A(v ) , B(v ) ∈ convhull A ( v ) , B ( v ) , ∀v , (4c)
(III)) was used to feed the acid solution to the reaction vessel.
To deliver the base solution, the peristaltic Pump B (Fig. 1, (IV)) where k ≥ 0 is an instance of the discrete-time domain, x(k ) ∈ Rnx
was used. The input voltage of the peristaltic pumps was within is the vector of system states, u(k ) ∈ Rnu are the control inputs,
[0, 5] V. The input concentrations of acid and base were cA = y(k ) ∈ Rny are the system outputs. Parameter x0 represents the
0.01 mol dm−3 , cB = 0.01 mol dm−3 , respectively. measured or estimated vector of the system initial conditions, A ∈
Considered CSTR has a significant nonlinear and asymmetric be- Rnx ×nx is a system-state matrix, B ∈ Rnx ×nu represents a matrix of
haviour, see [26]. This behaviour is a consequence of the nonlinear system inputs, C ∈ Rny ×nx is a matrix of the system outputs. Param-
shape of the titration curve. Titration curve depicts the dependence eter (v ) represents vth vertex of the system, v = 1, . . . , nv , where
of pH on the volume of reagent added to the solution. nv is the total number of uncertain system vertices.
Because of the complex behaviour of the controlled system, the The robust controller design aimed to remove the steady-state
parameters of the mathematical model were identified introduc- control error. In order to reach this goal, the vector of system states
ing the interval uncertainties. The boundary values of the uncertain x(k ) was extended by an integral action in the following way:
parameters were experimentally determined by generating the set x (k )
of multiple negative and positive step changes either of the flow x (k ) =
k , (5a)
j=0 e ( j )
rate qA or the flow rate qB . Step responses were measured in var-
ious operating conditions. The step-response-based identification e ( k ) = w ( k ) − y ( k ), (5b)
method [23] was applied to identify the parameters of the transfer
function in the form: where e(k ) ∈ is the control error and w(k ) ∈
Rny is the refer- R ny
ence.
Z
G (s ) = e−Ds , (3) Then, the augmented system is defined as follows:
Ts + 1
x (k + 1 ) =
x(k ) +
A (v ) B(v ) u(k ), x (0 ) =
x0 , (6a)
where s is a complex number frequency parameter of Laplace
transform, Z is the gain, T represents the time constant, and D
stands for the time delay. y(k ) =
Cx(k ), (6b)
Uncertain parameters were expressed using intervals bounded
by minimum and maximum values of each system parameter, i.e., satisfying
the gains ZA , ZB , the time constants TA , TB , and the time delays
A (v ) 0 (v )
DA , DB . Throughout the paper, the subscript A corresponds to the (v ) =
A , (v ) = B
B , = C
C 0 , (7)
−tsC (v ) I 0
acid solution and the subscript B represents the base solution. The
minimal and maximal transfer function parameters including the where A, B,
C are the matrices of the system augmented with re-
nominal, i.e., mean parameters, are summarised in Table 1. Param- spect to (w.r.t.) the integral action.
eters ZA , TA , and DA were determined measuring and identifying First, the steady-state operating point of the CSTR was experi-
step responses on the step changes of the input flow rate qA . If mentally determined as follows: qsA = qsB = 5 mL s−1 and pHs = 7.
the amount of acid in the reaction vessel increases, then the pH The superscript s represents the steady-state value.
value of the solution decreases. This behaviour represents an in- Then, the steady-state operating point of the closed-loop sys-
versely proportional dependence leading to the negative values of tem was normalised, i.e., it was shifted to the origin. Particularly,
the gains ZA , see Table 1. The parameters listed in Table 1 were in (6) we became:
identified subject to the controlled variable and manipulated vari-
xA ( k )
able in the deviation form.
x (k ) =
x (k ) , (8a)
Identified system parameters are strongly dependent on the
k B
specific type of acid and base considered in the reaction, as well j=0 e ( j )
as on their input concentrations. Considering the different type of
acid or base would change the shape of the titration curve and uA ( k ) q (k ) − qsA
u (k ) = = A , (8b)
would move the equivalence point of an acid-base reaction. There- uB ( k ) qB(k ) − qsB
fore, it would changes the volume of the reagents necessary for
the process of neutralisation. In this case study, the considered re-
y(k ) = pH(k ) − pHs . (8c)
action of the weak acid and strong base leads to the equivalence
point located above the neutral pH value. Considering a reaction As the states of the augmented system in (8a) were not mea-
of the strong acid and strong base, the equivalence point could be surable, the state observer was introduced to estimate the system
placed at pH = 7. Finally, considering a pair of the strong acid and states.
Please cite this article as: J. Oravec, M. Horváthová and M. Bakošová, Multivariable robust MPC design for neutralisation plant: Experi-
mental analysis, European Journal of Control, https://doi.org/10.1016/j.ejcon.2020.07.012
JID: EJCON
ARTICLE IN PRESS [m5G;August 21, 2020;11:47]
J. Oravec, M. Horváthová and M. Bakošová / European Journal of Control xxx (xxxx) xxx 5
3. Robust MPC
X
C
A (v ) X +
B (v ) Y y2sat I
0, (13e)
This section describes the theoretical principles of the consid-
ered robust MPC approach. The robust MPC method introduced in
[17] was considered. Although this well-known LMI-based method where the decision variables are X, Y, U, γ . The parameter X ∈
was refined by many later works reducing the conservativeness, we Rnx ×nx is the weighted inverse of the Lyapunov matrix P. Y ∈
use this control strategy to demonstrate the efficacy of the real- Rnu ×nx is the auxiliary matrix of controller design. U ∈ Rnu ×nu is
time compensation of the asymmetric behaviour. The benefit of the the auxiliary matrix of the controller design, which ensures taking
proposed real-time compensation method is that it can be imple- into account the constraints on manipulated variables in each con-
mented to all later works derived from [17], e.g., see [9,21,39,42], trol step. The parameter γ ∈ R is the weight parameter of X. The
and references therein. SDP in (13) was optimised in each control step, i.e., robust MPC
was implemented in receding-horizon-based framework. The sym-
3.1. Robust MPC formulation bol in (13) represents the symmetric structure of the LMIs, and
y2sat denotes element-wise power of 2.
The advantage of the robust MPC design is its ability to con- Similar to the approach in [17], the quadratic Lyapunov func-
sider symmetric constraints on the manipulated variables and the tion V(x(k ) ) = γ x(k ) X−1 x(k ) was used. The recursive feasibility of
controlled variables in the form: the optimisation problem is established by the robust positively in-
−usat u(k ) usat , −ysat y(k ) ysat , ∀k ≥ 0, (9) variant set in (13b). The objective function in (13a) maximises the
volume of robust positively invariant set, see [5]. The asymptotic
where usat ∈ Rnu ,
ysat ∈ Rny
are the limit values of the symmetric
stability in the sense of Lyapunov is ensured by LMIs in (13b) w.r.t.
constrains.
the penalty matrices in the quality criterion in (10). In (13c), I is an
The aim of robust MPC design was to ensure offset-free ref-
identity matrix and 0 is a zero matrix of appropriate dimensions.
erence tracking minimising well-known LQR-based control perfor-
LMI in (13c) ensures the control performance set by the penalty
mance criterion formulated w.r.t. the augmented system states in
matrices defined in (10). LMIs in (13d), (13e) take into account
(6a):
the constraints on manipulated variables and controlled variables
∞
in (9).
min x(k ) Q + u(k ) R ,
2 2
(10)
Based on the feasible solution of X and Y in (13), the gain ma-
k=0
trix is computed as follows:
where R ∈ Rnu ×nu
0 is the penalty matrix of manipulated vari-
F = Y X−1 .
ables.
Q ∈ R2 nx ×2 nx
0 is the penalty matrix of the extended vec- (14)
tor of states
x, which has the following form:
3.2. Robust MPC setup
Qp 0
Q= , (11)
0 Qi
From the controller design viewpoint, there are 3 challenges of
where penalty matrix Q consists of the proportional-like penalty robust MPC: (i) feasibility – to satisfy the constraints in (9), (ii)
matrix Qp of system states, and integration-like penalty matrix Qi optimality – to tune the penalty matrices Q , R to minimise the
of the integral action in (5a). The considered objective function quadratic criterion in (10), and (iii) implementation – to estimate
represents an LQR-like objective function for the uncertain system. the system states in (6). Therefore, the considered robust MPC had
Therefore, the infinite prediction horizon is considered. The robust the following setup.
MPC designed for an infinite prediction horizon outperforms the The robust MPC took considered constraints on the manipulated
standard MPC designed for a finite prediction horizon. The optimi- variables and the controlled variables as defined in (9). The flow
sation problem handles the infinite prediction horizon by formu- rates of the acid qA and base qB were constrained within interval
lating the tractable optimisation problem using the LMIs. [0, 10] mL s−1 . The controlled variable, pH, was constrained within
The robust MPC computes a gain matrix F ∈ Rnu ×(nx +ny ) of the interval [0, 14] by definition. Therefore, the considered constrains
state-feedback control law: were in the following form:
u(k ) = [Fp(k ) Fi(k )]
x(k ) =
F(k )
x(k ), (12) −5 u(k ) 5, −7 y(k ) 7, ∀k ≥ 0. (15)
where Fp(k ) and Fi(k ) are the proportional and the integral parts Penalty matrices in (10) had the form:
of the state feedback linear control law. The gain matrix F is con-
structed by solving the following optimisation problem of SDP, see Qp,A 0 0
RA 0
[17]: Qp = 0 Qp,B 0 , R= , (16)
0 RB
0 0 Qi
minγ ,X,Y,U γ (13a)
where Qp,A , Qp,B are the proportional-like penalty matrices, Qi is
the penalty of integral action, RA , RB are penalties of the manipu-
1 lated variables. The set of experiments was performed to investi-
s.t. :
0, (13b)
x(k ) X gate various setups of the penalty matrices, see Table 2.
⎡ X
⎤ The subject of the case study was a neutralisation plant with
a single controlled variable, i.e., the pH of the reaction mixture.
⎢
A (v ) X +
B (v ) Y X ⎥ However, two manipulated variables were available, i.e., the flow
⎣ 1/2 ⎦
0, (13c)
Q X 0 γI rate of acid qA and the flow rate of base qB . Both manipulated vari-
1/2
R Y 0 0 γI ables have a significant influence on the controlled variable, but
Please cite this article as: J. Oravec, M. Horváthová and M. Bakošová, Multivariable robust MPC design for neutralisation plant: Experi-
mental analysis, European Journal of Control, https://doi.org/10.1016/j.ejcon.2020.07.012
JID: EJCON
ARTICLE IN PRESS [m5G;August 21, 2020;11:47]
6 J. Oravec, M. Horváthová and M. Bakošová / European Journal of Control xxx (xxxx) xxx
RMPC setup Qp,A Qp,B Qi RA RB Switching between the pair of penalty matrices can cause sta-
RMPC1 1 1 1 10 10 bility issues. This subsection investigates the closed-loop system
RMPC2 1 1 10 1 1 stability based on the quadratic Lyapunov function. The closed-
RMPC3 1 1 10 10 10 loop system Lyapunov function candidate VCL (x) has the form of
RMPC4 10 10 1 10 10 a quadratic function:
VCL (x ) = x(k + 1 ) PCL x(k + 1 ), (21)
their influences represent internal states of the system. These in- where PCL is the Lyapunov matrix of the closed-loop system satis-
ternal states were not measurable directly. As the robust MPC de- fying
signs the state-feedback control law in (12) considering the values
of the states, the state observer is required. The Luenberger ob- PCL = PCL
0. (22)
server was designed to estimate the system states xA , xB of the Substituting the control law (12) into (6), we obtain the following
augmented state vector x in (8a). Considering only the nominal robust stability condition:
model of CSTR, the LQR-based optimal pole placement strategy
[23] was used to design the observer L ∈ Rnx ×ny . The associated (v ) + B
A (v ) F(i ) (v ) + B
PCL A (v ) F(i ) − PCL 0, v = 1, . . . , nv ,
penalty matrices were experimentally tuned as follows:
(23)
0.5 0
Qo = , Ro = 0.1, (17) where F(i ) is the ith gain matrix of nF robust MPC setups consid-
0 1
ered to compensate the asymmetric behaviour, i = 1, . . . , nF .
where Qo ∈ Rnx ×nx and Ro ∈ Rnu ×nu are the penalty matrices of the Given is uncertain system in (6) and the set of nF stabilising
observer design. The Luenberger observer L was computed: feedback gains F(i ) , i = 1, . . . , nF . If there exists such Lyapunov ma-
trix PCL in (22) satisfying (23) for all v = 1, . . . , nv , i = 1, . . . , nF ,
L = −1.1646 2.3917 . (18)
then the closed loop system is stable. Therefore, the stability of
4. Real-time compensation of asymmetric behaviour switching control law is certified in advance by solving the set of
(nv × nF ) LMIs in (23).
This section describes theoretical principles of real-time com-
pensation of asymmetric behaviour of CSTR. From the implemen- 5. Results and discussion
tation point of view, it is not limited by the particular structure of
the SDP optimisation problem. Stability of the closed-loop system This section discusses the results of the experimental imple-
considering the switching control law is also investigated. mentation of 4 LMI-based robust MPC setups (Table 2) and in-
vestigates the control performance of the robust MPC with real-
4.1. Implementation of switching control law time compensation of asymmetric behaviour. Each control setup
was analysed w.r.t. the set of quality criteria.
MPC is usually designed w.r.t. the linear time-invariant (LTI)
prediction model. The influence of the nonlinear and asymmet- 5.1. Experiment setup
ric behaviour is minimised by considering a set of LTI systems.
However, so-called plant-model mismatch decreases the control Laboratory CSTR used in the experimental case study was man-
performance. The idea of proposed real-time compensation of ufactured by Armfield, see Fig. 1. The closed-loop control was
asymmetric behaviour is based on the observation that imple- designed in MATLAB/Simulink R2018a environment using CPU i7
menting robust MPC with a certain penalty matrix setup has a 3.4 GHz and 8 GB RAM. Real-Time Windows Target v4.1 toolbox
satisfactory control performance when applied to a negative step ensured the communication with CSTR, and the communication
changes of the reference. On the other hand, the same control time was tc = 0.2 s. Sampling time of the control was ts = 10 s.
setup causes a decrease of control performance when applied to a MUP toolbox [29] was used to formulate robust MPC, SDPs were
positive step changes in the reference. The tuning of penalty ma- formulated using YALMIP [18], and solved by MOSEK [1]. The aver-
trices to obtain a single reliable setup is difficult and conservative. age time necessary to solve the optimisation problem in (13) was
Therefore, we introduce a real-time compensation, that toptim = 0.2 s. Therefore, the value of sampling time ts was signif-
switches between two reliable control setups. In order to deter- icantly greater than toptim that provides sufficient amount of time
mine the switching rule, multiple robust MPC setups were anal- to process the data and to delegate the optimal control action to
ysed. Finally, based on the evaluated control performance, the con- the plant.
trol setups denoted as RMPC2 and RMPC3 were determined as the The control performance criteria were evaluated considering
most suitable, see Table 2. The switching function determined the tF = 500 s, i.e., initial nk = 50 control steps. In order to obtain the
control setup based on the sign of the control error as follows: value of non-measurable states, state observer in (18) was imple-
mented. The overall experiment setup is illustrated in Fig. 2.
1 : 10 : 10, e(k ) > 0
Qp : Qi : R = (19) The offset-free reference tracking problem was investigated.
1 : 10 : 1, e(k ) ≤ 0.
The control performance was investigated considering 4 operating
If the control error at a given time instant e(k) had a non positive points represented by the following sequence of the step changes
value, then RMPC2 control setup was implemented. Otherwise, if of the reference pH value: (i) 6 → 7, (ii) 7 → 8, (iii) 8 → 7, and (iv)
the value of e(k) was positive, then RMPC3 control setup was con- 7 → 6.
sidered, see Table 2. It is obvious, that only penalty matrix R varies
in (19). Then, the switching rule in (19) was formulated in a com- 5.2. Experimental results
pact form:
Figs. 3–6 show the experimental results for 4 robust MPC se-
R = 1 + 4.5(1 + sign(e(k ))) sign(e(k )). (20)
tups summarised in Table 2. The trajectories of the controlled vari-
Further discussion on the results is in Section 5. able and the estimated values of pH are plotted in Figs. 3(a)–6(a).
Please cite this article as: J. Oravec, M. Horváthová and M. Bakošová, Multivariable robust MPC design for neutralisation plant: Experi-
mental analysis, European Journal of Control, https://doi.org/10.1016/j.ejcon.2020.07.012
JID: EJCON
ARTICLE IN PRESS [m5G;August 21, 2020;11:47]
J. Oravec, M. Horváthová and M. Bakošová / European Journal of Control xxx (xxxx) xxx 7
Please cite this article as: J. Oravec, M. Horváthová and M. Bakošová, Multivariable robust MPC design for neutralisation plant: Experi-
mental analysis, European Journal of Control, https://doi.org/10.1016/j.ejcon.2020.07.012
JID: EJCON
ARTICLE IN PRESS [m5G;August 21, 2020;11:47]
8 J. Oravec, M. Horváthová and M. Bakošová / European Journal of Control xxx (xxxx) xxx
nk −1
ISE ( pH ) = ts ( pH (k ) − pH s )2 , (24)
k=0
Step change of reference pH value 6 → 7 is depicted in Fig. 3(a).
The other positive step change 7 → 8 is depicted in Fig. 4(a). Nega-
tive step changes 7 → 6 and 8 → 7 are shown in Figs. 5(a) and 6(a). evaluated the precision of the reference tracking of the controlled
The associated manipulated variables are depicted in Figs. 3(b)– variable.
6(b), where the dashed curve represents the flow rate of acid and The integral value of squared input, ISI(q), determined by:
the solid curve shows the flow rate of the base. As can be seen in
Figs. 3–6, the estimated values of pH tracked the measured values
with satisfying precision. Based on the estimated values of system
nk −1
ISI (q ) = ts (qA (k ) − qsA )2 + (qB (k ) − qsB )2 , (25)
states, the robust MPC ensured offset-free reference tracking for all
k=0
operating points (Figs. 3(a)–6(a)), and the evaluated manipulated
variables did not violate the constraints (Figs. 3(b)–6(b)).
Control performance was analysed using the set of sum-of- evaluated the consumption-like value of acid and base solutions.
square-based quality criteria [23]. First, integral value of squared er- Finally, the integral value of squared difference of input, ISDI(q),
ror, ISE(pH), given by was evaluated by
Please cite this article as: J. Oravec, M. Horváthová and M. Bakošová, Multivariable robust MPC design for neutralisation plant: Experi-
mental analysis, European Journal of Control, https://doi.org/10.1016/j.ejcon.2020.07.012
JID: EJCON
ARTICLE IN PRESS [m5G;August 21, 2020;11:47]
J. Oravec, M. Horváthová and M. Bakošová / European Journal of Control xxx (xxxx) xxx 9
Table 3
Closed-loop control performance criterion ISE(pH).
Table 4
Closed-loop control performance criterion ISI(q) × 104 .
Table 5
Closed-loop control performance criterion ISDI(q).
Table 6
Closed-loop control performance criterion ISE(pH).
nk −1
ISDI (q ) = ts (qA (k ) − qsA )2 + (qB (k ) − qsB )2 (26)
k=0
and its value penalises the oscillation of flow rate. As the actuators
were the peristaltic pumps, minimisation of this criterion max-
imises the life-cycle of actuators. Fig. 7. RMPC with real-time compensation of asymmetric behaviour for reference
The control performance was simultaneously judged by the step change 6 → 7. (a) Controlled variable – measured pH (dark red); estimated pH
quality criteria in (24)–(26). The evaluated values of these criteria (light red dashed); reference (black dotted). (b) Manipulated variables – volumetric
flow rate of acid solution (red), volumetric flow rate of alkaline solution (blue);
are summarised inTables 3–5.
constraints (black dash-dotted).
It is possible to observe, that each operating condition had a
particular RMPC setup ensuring the best control performance, as Table 7
the consequence of nonlinear and asymmetric behaviour of the Closed-loop control performance criterion ISI(q) × 104 .
plant. In order to minimise the influence of these properties, real- RMPC setup pH: 6 → 7 pH: 7 → 8 pH: 8 → 7 pH: 7 → 6
time compensation of asymmetric behaviour was considered. The
RMPC2 0.24 1.48 1.41 1.28
theoretical backgrounds of the compensation were introduced in
RMPC3 1.26 1.67 1.33 1.53
Section 4. Based on the data summarised in Tables 3–5, the RMPC RMPC2&3 0.23 0.24 0.36 0.24
setups with minimum values of ISE(pH), ISI(q), and ISDI(q) were IF(ISI) 3.26 6.56 3.80 5.85
determined for the implementation of real-time compensation of
asymmetric behaviour. More specifically, setups RMPC2 and RMPC3
were selected as suitable setups for the compensation. Therefore, factor (IF) was evaluated. The relative improvement factor for ISE
the real-time compensation-based control setup in (20) is denoted criterion is given by:
by RMPC2&3.
1
Control performance of robust MPC with real-time compen- (ISE (RMPC2 ) + ISE (RMPC3 ) )
sation of asymmetric behaviour is depicted in Figs. 7–10, where IF (ISE ) = 2 . (27)
Figs. 7(a)–10(a) show the controlled variable (solid curves) and the ISE (RMPC2&3 )
estimated pH values (dashed curves). The associated manipulated The evaluation of IF(ISI) and IF(ISDI) was analogous to (27). If the
variables are in Figs. 7(b)–10(b), where the dashed curves repre- improvement was not achieved, the improvement factor is denoted
sent acid solution and the solid curves denote alkaline solution. by symbol †, see Table 6.
The quality criteria in (24)–(26) were also evaluated, see Implementing the real-time compensation of asymmetric be-
Tables 6–8. The generated data were analysed w.r.t. control setups haviour improved the majority of the analysed quality criteria.
RMPC2 and RMPC3 that were considered to construct the real-time However, implementation of this method increased the value of
compensation. For each quality criterion, the relative improvement ISE(pH) criterion for two of four operating points, see Table 6.
Please cite this article as: J. Oravec, M. Horváthová and M. Bakošová, Multivariable robust MPC design for neutralisation plant: Experi-
mental analysis, European Journal of Control, https://doi.org/10.1016/j.ejcon.2020.07.012
JID: EJCON
ARTICLE IN PRESS [m5G;August 21, 2020;11:47]
10 J. Oravec, M. Horváthová and M. Bakošová / European Journal of Control xxx (xxxx) xxx
Table 8 high, then the actuator suffers and its life-cycle is shortened. The
Closed-loop control performance criterion ISDI(q). decrease in this criterion minimises the maintenance costs. The
RMPC setup pH: 6 → 7 pH: 7 → 8 pH: 8 → 7 pH: 7 → 6 value of this criterion was on average reduced by 85.6 % compared
to original RMPC setup.
RMPC2 1.69 21.38 9.71 21.09
RMPC3 21.51 22.23 20.83 20.98
RMPC2&3 1.68 2.81 1.52 4.96 6. Conclusions
IF(ISDI) 7.41 7.76 10.04 4.24
Please cite this article as: J. Oravec, M. Horváthová and M. Bakošová, Multivariable robust MPC design for neutralisation plant: Experi-
mental analysis, European Journal of Control, https://doi.org/10.1016/j.ejcon.2020.07.012
JID: EJCON
ARTICLE IN PRESS [m5G;August 21, 2020;11:47]
J. Oravec, M. Horváthová and M. Bakošová / European Journal of Control xxx (xxxx) xxx 11
Acknowledgement
References
Please cite this article as: J. Oravec, M. Horváthová and M. Bakošová, Multivariable robust MPC design for neutralisation plant: Experi-
mental analysis, European Journal of Control, https://doi.org/10.1016/j.ejcon.2020.07.012
JID: EJCON
ARTICLE IN PRESS [m5G;August 21, 2020;11:47]
12 J. Oravec, M. Horváthová and M. Bakošová / European Journal of Control xxx (xxxx) xxx
[23] J. Mikleš, M. Fikar, Process Modelling, Identification, and Control, Springer Ver- [33] P. Scokaert, D. Mayne, MIN-MAX feedback model predictive control for con-
lag, Berlin Heidelberg, 2007. strained linear systems, IEEE Trans. Autom. Control 43 (1998) 1136–1142,
[24] C.M. Nguyen, P.N. Pathirana, H. Trinh, Robust observer-based control designs doi:10.1109/9.704989.
for discrete nonlinear systems with disturbances, Eur. J. Control 44 (2018) 65– [34] L. Vandenberghe, S. Boyd, Semidefinite programming, SIAM Rev. 38 (1996)
72, doi:10.1016/j.ejcon.2018.09.002. 49–95.
[25] I. Nodozi, M. Rahmani, LMI-based model predictive control for switched non- [35] Z. Wan, M.V. Kothare, Efficient robust constrained model predictive control
linear systems, J. Process Control 59 (2017) 49–58, doi:10.1016/j.jprocont.2017. with a time varying terminal constraint set, Automatica 48 (2003) 375–383,
09.001. doi:10.1016/S0167-6911(02)00291-8.
[26] J. Oravec, M. Bakošová, L. Hanulová, M. Horváthová, Design of robust MPC with [36] Z. Wan, M.V. Kothare, An efficient off-line formulation of robust model predic-
integral action for a laboratory continuous stirred-tank reactor, in: Proceedings tive control using linear matrix inequalities, Automatica 39 (2003) 837–846,
of the 21st International Conference on Process Control, Štrbské Pleso, Slovakia, doi:10.1016/S0 0 05-1098(02)0 0174-7.
2017, pp. 459–464, doi:10.1109/PC.2017.7976257. [37] R. Yadbantung, P. Bumroongsri, Tube-based robust output feedback MPC for
[27] J. Oravec, M. Bakošová, L. Hanulová, Experimental investigation of robust MPC constrained LTV systems with applications in chemical processes, Eur. J. Con-
design with integral action for a continuous stirred tank reactor, in: Proceed- trol 47 (2019) 11–19, doi:10.1016/j.ejcon.2018.07.002.
ings of the 57th IEEE Conference on Decision and Control, Miami, Florida, USA, [38] Y. Yang, S. Dubljevic, Linear matrix inequalities (LMIs) observer and controller
57, 2018a, pp. 2611–2616, doi:10.1109/CDC.2018.8619737. design synthesis for parabolic PDE, Eur. J. Control 20 (2014) 227–236, doi:10.
[28] J. Oravec, M. Bakošová, L. Hanulová, A. Mészáros, Multivariable robust model 1016/j.ejcon.2014.05.002.
predictive control of a laboratory chemical reactor, in: Proceedings of the 28th [39] W. Yang, G. Feng, T.J. Zhang, Robust model predictive control of uncertain lin-
European Symposium on Computer Aided Process Engineering, 28, Elsevier, ear systems with persistent disturbances and input constraints, in: Proceedings
2018b, pp. 961–966, doi:10.1016/B978- 0- 444- 64235- 6.50169- 8. of the European Control Conference, Zürich, Switzerland, 2013, pp. 542–547,
[29] J. Oravec, M. Bakošová, Software for efficient LMI-based robust MPC design, in: doi:10.23919/ECC.2013.6669673.
M. Fikar, M. Kvasnica (Eds.), Proceedings of the 20th International Conference [40] M.N. Zeilinger, D.M. Raimondoe, A. Domahidid, M. Morari, C.N. Jones, On real-
on Process Control, Slovak University of Technology in Bratislava, Štrbské Pleso, time robust model predictive control, Automatica 50 (2014) 683–694, doi:10.
Slovakia, 2015, pp. 272–277, doi:10.1109/PC.2015.7169975. 1016/j.automatica.2013.11.019.
[30] S. Rastegar, R. Arajo, J. Sadati, J. Mendes, A novel robust control scheme for [41] S. Zhang, D. Zhao, S. Spurgeon, Robust distributed model predictive control
LTV systems using output integral discrete-time synergetic control theory, Eur. for systems of parallelstructure within process networks, J. Process Control 82
J. Control 34 (2017) 39–48, doi:10.1016/j.ejcon.2016.12.006. (2019) 70–90, doi:10.1016/j.jprocont.2019.06.005.
[31] S. Salehi, M. Shahrokhi, A. Nejati, Adaptive nonlinear control of pH neutraliza- [42] L. Zhang, J. Wang, K. Li, Min-max MPC for LPV systems subject to actuator
tion processes using fuzzy approximators, Control Eng. Pract. 17 (2009) 1329– saturation by a saturation-dependent Lyapunov function, in: Proceedings of the
1337, doi:10.1016/j.conengprac.20 09.06.0 07. Chinese Control Conference, Xi’an, China, 2013, pp. 4087–4092.
[32] M. Saltik, L. Özkan, J. Ludlage, S. Weiland, P.V. den Hof, An outlook on robust
model predictive control algorithms: Reflections on performance and computa-
tional aspects, J. Process Control 77 (2018) 77–102, doi:10.1016/j.jprocont.2017.
10.006.
Please cite this article as: J. Oravec, M. Horváthová and M. Bakošová, Multivariable robust MPC design for neutralisation plant: Experi-
mental analysis, European Journal of Control, https://doi.org/10.1016/j.ejcon.2020.07.012