Alexandria Engineering Journal (2014) 53, 939948

H O S T E D BY
Alexandria University

Alexandria Engineering Journal

www.elsevier.com/locate/aej
www.sciencedirect.com

ORIGINAL ARTICLE

Coagulation process control in water

treatment plants using multiple model predictive
control
Oladipupo Bello *, Yskandar Hamam, Karim Djouani

Electrical Engineering Department/FSATI, Tshwane University of Technology, Pretoria X680, South Africa

Received 3 April 2014; revised 4 August 2014; accepted 7 August 2014

Available online 27 August 2014

KEYWORDS Abstract A multiple model predictive control (MMPC) strategy is proposed for coagulation
Water treatment plants; control in water treatment plants. The proposed control strategy is developed to work effectively
Multiple model predictive with different local operating regions of the chemical dosing unit where coagulation takes place.
control; The surface charge and pH values of the chemically treated water are controlled at the same time
MPC local controller; to reduce waste, operational costs and improve the quality of portable water distributed for public
Coagulation consumption by manipulating the ow rates of the chemical reagents. The simulation results show
that coagulation control can be effectively controlled by using the proposed control strategy.
2014 Production and hosting by Elsevier B.V. on behalf of Faculty of Engineering, Alexandria
University.

1. Introduction operating regime to another and thereby reduce wastage of

Coagulation is an important unit operation in the water
treatment plants. Coagulation is described by complex chemi-
cal reactions and coagulation chemical dosing unit exhibits
nonlinear behaviour [1]. Thus, coagulation control is quite
challenging. Furthermore, in the operations of chemical dosing
unit, irregular variations in the water quality and demands,
depletion of chemical reagents stock, system faults and plant
operators errors often lead to multiple operating regimes
resulting to either underdose, normal or overdose operating
condition. In order to reduce the transition time from one
* Corresponding author.
E-mail addresses: engroobello@gmail.com (O. Bello), hamama@tut.
ac.za (Y. Hamam), djouanik@tut.ac.za (K. Djouani).
Peer review under responsibility of Faculty of Engineering, Alexandria
University.
http://dx.doi.org/10.1016/j.aej.2014.08.002
1110-0168 2014 Production and hosting by Elsevier B.V. on behalf of Faculty of Engineering, Alexandria University.
chemical reagents, minimise production cost and meet the nal
water quality target, an effective control strategy must be
implemented at water treatment plants.
Model predictive control (MPC) is a widely accepted and
useful control strategy for industrial process applications as
a result of its ability to re-formulate the control problem as
an optimisation problem and demonstrate satisfactory perfor-
mance when satisfying operational and safety constraints.
However, MPC is an ineffective approach to design local con-
trollers for systems with strong nonlinearities and sudden
changes in the operating conditions [22]. In order to address
this problem, MMPC strategy has been identied as a viable
solution in the literature [6,13,18,20,26,12,10,24].
MMPC as an optimal control strategy is developed using sev-
eral linear models that represent separate local regions or
regimes of a nonlinear system. Each linear model effectively
describes its local region of the nonlinear system. Corresponding
940 O. Bello et al.

MPC local controller is designed for each linear model. All the The simulation studies on the MMPC performance on the
models are computed in parallel and the best model/controller chemical dosing unit show promising results.
pair is selected based on the identication error at each instant
[22]. The switching scheme engaged for the model/controller 2. Materials and methods
pairs operates when another pair yields smaller identication
error than the current pair. The global control input is thus 2.1. Description of Rietvlei water treatment plant
generated using the parameters of the selected pair.
In Yu et al. [25], MMPC control strategy was proposed for
Rietvlei water treatment plant, City of Tshwane, South Africa
the simultaneous regulation of mean arterial pressure and car-
was built between 1932 and 1934. From inception, the plant
diac output in heart failure patients by adjusting the infusion
has undergone a number of upgrade projects to ensure that
rates of inotropic agent and vasodilator. However, the compu-
safe and portable drinking waters are supplied for public con-
tation using Bayes theorem was immense and involved several
sumption in Pretoria. The typical production capacity of the
stages. Zhang and Morris [27] proposed MMPC control strat-
plant is about 40 ml per day. The plant draws raw waters from
egy using a recurrent neuro-fuzzy network to identify the local
Rietvlei Dam. There are three different abstraction points at
linear models of a pH neutralisation process.
different levels on the dam tower, namely: top; middle and bot-
Porrio et al. [20] examined the application of a multi-
tom draw-off points. Only one point is used at a time depend-
model predictive control to an industrial distillation system
ing on water level in the dam. Raw water quality is typically
of a renery (C3/C4 splitter). The proposed MMPC performed
low in turbidity but with high algal content. However, the tur-
signicantly better than the conventional MPC controller. In
bidity level may be high due to discharges from oods and
Li et al. [13], the authors proposed MMPC strategy based on
rainfall. These changes in water turbidity often lead to difcul-
TakagiSugeno modelling for a nonlinear process control.
ties in coagulation. A pumping station and transport system
The proposed controller was applied successively to regulate
are available to lift water from the dam and to convey it to
a MIMO pH neutralisation process.
the treatment plant. Raw water ows through the intake pipe
A study by Zeng et al. [26] on the application of MMPC
into the chemical dosing unit. The ow of water through the
strategy to control dissolved oxygen concentration in wastewa-
intake pipe is measured using owmeter. The amount of water
ter treatment process was performed. The authors of Mazinan
abstracted is also used to determine the amount of chemicals
and Sadati [15] proposed a multiple model predictive control
that should be applied during the subsequent process. Raw
based on fuzzy decision mechanism. The validity of the pro-
water inlet valve is placed after the owmeters to control the
posed control strategy was veried using a water tubular heat
quantity of water owing into the plant.
exchanger system. Guolian et al. [10] proposed MMPC based
Fig. 1 illustrates the coagulation chemical dosing unit of
on fuzzy adaptive weighted algorithm to control the steam
Rietvlei water treatment plant. It is made up of a concrete mix-
temperature of a thermal power plant under a wider range of
ing tank with inlet and outlet channels. There are four metring
load changes.
or dosing pumps that are used for dosing chemicals into the
Bagheri et al. [2] proposed MMPC based on adaptive
inuent raw waters owing into the treatment plant. Two of
weighting matrices to improve the performance of a general
the pumps are on active position and the remaining two pumps
predictive control (GPC) applied to a multivariable pH neu-
are on redundant position. One of the active pumps feeds poly-
tralisation process. Shamasaddinlor et al. [21] proposed
electrolyte solution while the other feeds ferric chloride solution
MMPC algorithm using the fuzzy weighting decision making
into the concrete mixing tank. The polyelectrolyte solution is
scheme for single-input and single-output pH neutralisation
referred to as sudoc 3835, a blend of epichlorohydrin/dimeth-
process control. However, the application of their proposed
ylamine (polyamine) and aluminium chlorohydrate [17]. The
scheme to multiple-input and multiple-output (MIMO) sys-
pumps are controlled by programmable logic controllers
tems was recommended. In all these previous studies, research-
(PLCs) that are connected to the plants supervisory control
ers have demonstrated the capabilities of MMPC to control
nonlinear systems with interesting results.
In this paper, the design, application and performance of
MMPC strategy to coagulation chemical dosing unit for water
treatment plants are examined and studied. The motivation of
this study is to demonstrate that MMPC is suitable for effec-
tive coagulation control and optimisation of chemical reagents
that are used for water treatment process. The control
approach includes the development of MPC local controller
for each distinct operating region of a chemical dosing unit.
A switching scheme is developed to select the appropriate
model/controller pair that could adequately represent the cur-
rent operating condition of the unit from the model/controller
bank. The surface charge and pH level of the chemically trea-
ted water are regulated to meet the desired target quality levels
by manipulating the ow rates of the primary coagulant, co-
coagulant and hydrated lime ow rates. Rietvlei water treat-
ment plant, South Africa is selected for this study. The process Figure 1 Coagulation chemical dosing unit of Rietvlei water
modelling of the chemical dosing unit of the plant is presented. treatment plant.
Coagulation process control in water treatment plants using multiple model predictive control 941

and data acquisition (SCADA) system. The quantities of the HCO

3   C5 H12 ON n   Al
3
 Fe3   Ca2
coagulation chemicals dosed depend on the ow rate of the H   OH  6
raw water owing into the plant. Calcium hydroxide (hydrated
lime) in slurry form is also added to the mixing tank using a dia- Let the difference of the ionic concentrations be expressed as:
phragm pump to stabilise the water and adjust the pH value to      3   3   2 
X HCO
3  C5 H12 ON n  Al  Fe  Ca
the desired level.
7
The chemically treated water ows out slowly and evenly
through a series of bafed or occulation channels, to grow where
the ocs. The water from the bafed channels ows into the
X H   OH  8
Dissolved Air Floatation/Filtration (DAFF) unit. The unit
removes all the ocs that have just been formed in the occu- Assuming that there is perfect mixing in the tank reactor,
lation channels. The occulated water is passed over the super- the relationship between the ionic concentrations of the efu-
saturated water. This leads to a oatation process where the ent and input concentrations using the material balance equa-
ocs rise as a result of the water bubbles attached to them. tions can be expressed as:
 
The oating particles form a scummy brown layer or froth d C5 H12 ON n   
on the surface of the lter bed while separation from it takes V C5 H12 ON n in qa
dt
place through the lter beds. The DAFF lters are cleaned reg-  
 C5 H12 ON n qout 9
ularly by manual backwashing operation. The backwash water
 3 
is kept in settling tanks for couple of hours. Afterwards, the d Al    3 
settled sludge is pumped into a sewerage system and the rela- V Al3 in qa  Al qout 10
dt
tively clean water is pumped back into the dam. The ltered  
water from the DAFF ows into Granular Activated Carbon d Fe3    3 
V Fe3 in qb  Fe qout 11
(GAC) ltration unit to eliminate foul odour, taste and colour dt
caused by the dissolved organic matter. The water from GAC  
ows into the chlorination chamber. Here, the chlorine gas is d Ca2    2 
V Ca2 in qc  Ca qout 12
added to the water to disinfect the clean water before it is dt
pumped to the storage reservoirs and distributed to nal con-  
sumers [5]. d HCO    
V 3
HCO 
3in qin  HCO3 qout 13
dt
 
2.2. Nonlinear model of the chemical dosing unit where C5 H12 ON n is the polyamine ionic  concentration at
the mixing tank outlet, C5 H12 ON n in the polyamine ionic
 
The coagulation chemical dosing unit involves a nonlinear pro- concentration at the mixing tank inlet, Al3 is the aluminium
 
cess. The dynamics of the process model are expressed as fol- ionic concentration at the mixing tank outlet, Al3 in alumin-
 
lows [3]: ium ionic concentration at the mixing tank inlet, Fe3 is
2C5 H12 ONCln 2Al2 ClOH5 2CaHCO3 2 ! the ferric ionic concentration at the mixing tank outlet,
 3 
Fe ferric ionic concentration at the mixing tank inlet,
2C5 H12 ON n 4AlOH3 OH 2CaCl2 4CO2 1  in2 
Ca calcium ionic concentration at the mixing tank outlet,
 2 
2FeCl 2CaHCO3 2 !2FeOH3 3CaCl2 6CO2 2 Ca calcium ionic concentration at the mixing tank inlet,
 in  
HCO3 bicarbonate ionic concentration of the efuent
 
CaHCO3 2 CaOH2 !2CaCO3 2H2 O 3 stream, HCO 3in bicarbonate ionic concentration of inuent
stream (raw water), qa ow rate of sudoc 3835 solution, qb
Chemical reactions in (1)(3) take place simultaneously.
Therefore, the overall chemical equation becomes: ow rate of ferric chloride solution, qc ow rate of hydrated
lime, qout ow rate of the efuent stream (chemically dosed
2C5 H12 ONCln 2Al2 ClOH5 2FeCl3 CaOH2 water), qin ow rate of the inuent stream (raw water), V vol-
ume of mixing tank, H  hydrogen ions concentration, OH 
6CaHCO3 2 !2C5 H12 ON n 2OH hydroxide ions concentration and kw dissociation constant of
2FeOH3 4AlOH3 5CaCl2 10CO2 water.
Adding (9)(12) and then subtract the sum from (13) gives:
2CaCO3 2H2 O 4          
    d HCO 3 d C5 H12 ON n d Al3 d Fe3 d Ca2
V V V V V
The reaction invariants of (4) are C5 H12 ON n ; Al3 ; 
dt
 
dt
 
dt
 
dt
 3 
dt
 3   2        HCO C5 H12 ON n in qa  Al3
3in qin  in qa  Fein qb

Fe ; Ca ; H ; HCO  2
3 ; OH ; SO4 ; Cl  and CO3 .
2
      
     Ca2 C5 H12 ON n in qout Al3 qout
in qc
However, Cl ; SO24 and CO2 3 ions present in the system  3   2   

Fe qout Ca qout  HCO3 qout 14
do not take part in the neutralisation reactions [3]. Therefore,
the electroneutrality equation of (4) is expressed as: The material balance expression of the mixing tank reactor
        yields:
C5 H12 ON n Al3 Fe3 Ca2 H 
  dX       
HCO  V HCO3in qin  C5 H12 ON n in qa Al3
in qa
3 OH  5 dt
   2 
 Fe3
in qb  Cain qc  Xqout 15
Rewriting (5) gives:
942 O. Bello et al.

Thus, (15) can be expressed as: chemicals used for the process. This would be achieved in this
dX      3  study by manipulating the ow rates of the primary coagulants
V HCO
3in qin  C5 H12 ON n in Alin qa qa , co-coagulant qb and hydrated lime reagents qc in
dt
   2  order to keep the SC and pH values of the chemically treated
 Fe3
in qb  Cain qc  Xqout 16 water at the desired set point levels determined by the plants
The dissociation equation for water is operators. The main challenge in this process is the presence
of severe uctuations in the raw water quality caused by daily
H OH  kw 1014

17 and seasonal changes in weather conditions, variations in
Substituting (8) into (17) gives a quadratic expression: water dam level and demands by consumers, heavy rainfall/
2
oods, industrial efuents and agro-allied activities. The pres-
H   XH   kw 0 18 ence of nonlinearity in the system dynamics and varying oper-
The solutions of the equation (18) are written as: ating conditions could also affect the performance of the
r ! chemical dosing unit. These identied operational challenges
X 4kw must be handled by the control scheme to meet the water qual-
if X < 0; thenH  1 2 1
2 X ity target at the minimum operational cost. Control system for

p adequate coagulation control should meet these highlighted
if X 0; thenH  kw 19 control objectives and also keep the input and the output vari-
r !
X 4kw ables within the safe and acceptable operational limits.
if X > 0; thenH   1 2 1
2 X
2.4. Multiple model predictive control (MMPC)
The concentration of the hydrogen ions H  may be
expressed in the logarithmic function as: The deployment of model predictive control (MPC) in several
pH  log H 
20 industrial applications has been very successful as a result of its
ability to handle multivariable control problems with con-
The expression for the surface charge (SC) of the raw water straints on the system variables[16]. The principle of operation
is obtained as [3]: of MPC involves establishing the dynamic model of the system
 
12 to be controlled and specifying the physical constraints on the
2
r njT sinh 1:15pH0  pH 21 system variables. The objective of the MPC is to determine the
p
control input sequence that minimises a given performance
where r surface charge (SC), j Boltzman constant, T temper- index or cost function J, based on a desired output trajectory
ature,  relative dielectric permittivity, pH0 pH at point of zero over the prediction horizon P. From the calculated input
charge and n ionic strength. sequence, only the rst sample is applied to the process. The
From (16), (20) and (21), a system of differential and alge- entire procedure is repeated at the next sampling instant,
braic equations (DAE) that describes the dynamics of the according to the receding horizon strategy [7,11,23].
coagulant dosage system is formed. The surface charge and For a system with varying operating regions, the MPC is
pH are the controlled output variables while the coagulant ineffective to minimise the transition time and reduce the
ow rate qa , co-coagulant ow rate qb and pH adjustment amount of substandard products produced during transition.
chemical ow rate qc are the control input variables of the A well established approach is to decompose the full range
model. of the systems operation into a number of operating regions
or regimes. A local model and/or MPC local controller is
2.3. Control problem description developed for each region. These local model/controller pairs
are integrated together or synthesised to produce a global
The control objective of an effective control system for coagu- model and controller generally referred to as multiple model
lation in water treatment plants is to optimise the quantity of predictive control strategy [18].

Figure 2 Schematic of multiple model predictive control scheme.

Coagulation process control in water treatment plants using multiple model predictive control 943

Figure 3 Flowchart for the MMPC strategy.

944 O. Bello et al.

Table 1 Operating regions of the dosing unit.

Operating region Surface charge (leq/mg) pH Values (pH unit)
Underdose SC < 0 pH > 8
Normal 0 6 SC 6 1 105 7 6 pH 6 8
Overdose SC > 1 105 pH < 7

2.4.1. Plant description

Table 2 Process modelling parameters. In this study, it is assumed that the plant in Fig. 2 is described
by the following state space representation [19]:
Variable and symbols Values and units
Sudoc 3835 ions 0.00001 mol/L xk 1 Ap xk Bp up k Dp d
    
concentration, C5 H12 ON n Al3 yk Cp xk Pp d 22
 
Ferric ions concentration, Fe3 0.00001 mol/L where p represents plant, x 2 R ; u 2 R and y 2 Rp . d
n m
 2 
Calcium ions concentration, Ca 0.00001 mol/L denoted the plant disturbance vector. Dp and Pp represent
 
Bicarbonate ion concentration, HCO 0.00001 mol/L 
the dynamic matrices of the disturbance, d.
3
Hydrogen ion concentration, H 
107
Tank volume, V 29,000 L 2.4.2. Model description and state estimation
Dissociation constant of water, kw 1014
Temperature, T 298 K The model bank (comprising the N linearised models) repre-
Ionic strength, n 50 106 mol/L sents the full range of operation of the dosing unit. It can be
Relative dielectric permittivity,  80 stated as:
Boltzman constant, j 1:38x1023 J K1
xk 1 Ai xk Bi uk Di dk
23
yk Ci xk Pi dk
where i 1; . . . ; N; uk is the control input vector, xk is the
state vector and yk is the controlled output vector at the sam-
Table 3 Parameters specication of the model predictive pling instant k. Ai ; Bi and Ci are the model matrices. Di and Pi
control. are the model disturbance matrices. dk is the model
disturbance.
Parameters Values
Assume that noise is present on both the process and out-
Prediction horizon, p 15 put, the model representation is given as [24,19]:
Control horizon, m 4
Constraint: sudoc 3835 ow 0 < qa < 2 xk 1 Ai xk Bi uk Di dk wx k
Constraint: ferric chloride ow 0 < qb < 2 dk 1 dk wd k 24
Constraint: hydrated lime 0 < qc < 2
Weight: surface charge 1 yk Ci xk Pi dk vk
Weight: pH 1
where wx k; wd k and vk are white noise and normal dis-
Rate weight for input variables 0.2
tributed, with mean zero and covariance Qx ; Qd and Rv
respectively.
Thus, the one-step prediction model equations are:
x^k 1jk Ai x^kjk Bi ui k Di d^kjk
The proposed MMPC-based control system for coagulation
chemical dosing unit is shown in Fig. 2. The whole full range of d^k 1jk d^kjk 25
operation of the dosing unit is divided into N operating regions y^k Ci xk Pi ui k
with N linearised models that adequately describe the behav-
iour of the each local region. MPC local controller is designed The states are estimated from the plant measurement yk by
for each local model. A switching mechanism based on a using the linear quadratic ltering techniques as:
dened criterion through which the MPC local controller tran- x^kjk x^kjk  1 Lx yk  y^kjk  1
sit from one to another as operating condition changes is 26
incorporated into the control system [24]. The selector is d^kjk d^kjk  1 Ld yk  y^kjk  1
designed to select the model that is nearest to the dynamics where Lx and Ld are the gain matrices of the steady-state Kal-
of the current state of the system. The key components of man lter.
the proposed MMPC strategy include the model identication;
and control parts. The identication part is made up of prede- 2.4.3. Control optimisation problem
ned models that cover different regions of the process and the
The MPC local controllers in the controller bank solve the
decision mechanism that selects which model has the least
quadratic optimisation problem stated as [4]:
error[15,21]. The owchart in Fig. 3 shows the step-by-step pro-
cedure for the design and simulation of the MMPC algorithm Ju min
|{z} Sy k Su k SDu k 27
proposed for the coagulation chemical dosing unit. Duk;...;DukM1
Coagulation process control in water treatment plants using multiple model predictive control 945

Figure 5 Control moves of the ow rates or input variables.

Figure 4 Set point tracking performance of the MMPC strategy To account for the plant-model mismatch, the actual refer-
on the chemical dosing unit.
ence trajectory, rk is corrected to avoid optimal control in
open-loop sense [11]. The MPC therefore is presented with
set points that use the current plant output as feedback, an
where equivalent of integral action. This estimates the plant-model
mismatch or disturbance estimate as:
X
P X
ny
 y 2
S y k wj rc;j k i  yj k 1 28 ek yk  ym k 34
i1 j1
where yk and ym k are the plant and model outputs at
X
M nu n
X o2 instant k, respectively. The disturbance estimate is assumed
S u k wuj uj k i  1  uj  29 constant throughout the predicted horizon, P. The actual ref-
i1 j1 erence trajectory for the MPC is corrected with the disturbance
estimated to give the desired reference trajectory as:
X nDu n
M X o2
SDu k wDu rc k rk  ek 35
j Duk i  1 30
i1 j1
For the chemical dosing unit in this study that has two out-
subject to: put variables and three input variables, the weights matrix of
each MPC local controller are dened as follows.
umin 6 uk i  1 6 umax ; for i 1 to M 31
wDuj diagRDu1 ; RDu2 ; RDu3 36
Dumin 6 Duki1Duki2 6 Dumax ; for i 1 to M 32
wuj diagR1 ; R2 ; R3 37
ymin 6 yk i 6 ymax ; for i 1 to P 33
wyj diagQ1 ; Q2 38
where rc k is the vector holding the corrected reference trajec-
tory, yk jjk is the j-step ahead predicted output given the where RDu1 ; RDu2 ; RDu3 ; R1 ; R2 ; R3 ; Q1 and Q2 are the eight
present output measurements, wyj is the positive denite output weight coefcients in [0, 1] that can be adjusted by the designer
error weighting matrix, wDu is the positive semi denite input to minimise deviation of the two output variables from their
j
weighting matrix. uj is the nominal value of input j. set point values or track their set points satisfactorily.
The weighting matrices and the prediction horizon param-
eters, P, and the control horizon M are the tuning parameters 2.4.4. Switching mechanism
which can be used to shape the closed-loop response of the sys- The switching algorithm for the proposed MMPC strategy is
tem [25]. The choice of prediction horizon P especially when realised as follows [8,9]:
it is large improves the nominal stability of the closed loop sys-
tem. However, smaller values may easily lead to unstable sys-  Compute the performance index of each local model l
tem [23,6,14]. using:
946
Il c1 e2l c2 Rkik0 qik0 e2l k l 2 1; N 39
where c1 > 0; c2 > 0; q 2 0; 1 are tuning parameters, k0 is
the sampling instant when the change happens and
el k yk  yl k l 1; 2; . . . ; N 40
 Calculate and compare the performance index in (40) every
sampling instant.
 Generate the switching variables S l k from
0 1
Sl k H@ |{z}
min Il k  Il kA 41
l21;N
 The switching scheme is implemented by calculating and
comparing the performance indices every sampling instant,
generating the switching variables S l k from (42) where
H x is the Heaviside unit step function given by:
1 xP0
Hx 42
0 x<0
3. Simulation results and discussion
In this section, simulation studies were carried out to evaluate
the performance of the MMPC on the chemical dosing unit.
The whole operating region of the system was dened and par-
titioned into three local operating regions as shown in Table 1.
For each of these local regions, a linearised model was devel-
oped. Using these linearised models, three MPC local control-
lers were designed to control the dosing unit. The parameters
Figure 6 Disturbance rejection performance of the MMPC
scheme.
O. Bello et al.
in Tables 2 and 3 were used for the development and simula-
tion of the local models and MPC local controller respectively.
3.1. Set point tracking performance
Fig. 4 shows the performance of the MMPC scheme when set
points changed abruptly at time, t = 6 h from
9  106 leq=mg, 7 pH) to 1:8  106 leq=mg, 9.5 pH)
and remained there for 6 h. At t = 12 h, the set point changed
and moved to 5:5  105 leq=mg, 5.5 pH) and at t = 18 h,
the set point changed nally to 5  106 leq=mg, 7.5 pH). It
is observed that the MMPC scheme tracks the changes in set
points satisfactorily. The control moves of the manipulated
variables are shown in Fig. 5.
3.2. Disturbance rejection performance
In the next simulation study to examine the performance of the
proposed control scheme to reject disturbances, a sinusoidal
noise (frequency of 100 rad s1) was added to the output of
the chemical dosing unit. In addition, disturbances represented
by step change in the ow rates of the hydrated lime and ferric
chloride were injected to the dosing unit at t = 6 h and
t = 12 h respectively. The set points were set for SC and pH
as 0.000005 leq/mg and 7.4 respectively. Fig. 6 illustrates the
MMPC performance when perturbations were introduced to
the system. Fig. 7 shows the trajectory of the control inputs
to the chemical dosing unit. The simulation result shows that
MMPC is robust against the added noise and step input sig-
nals. It demonstrates very good ability to reject disturbances
in form of sensors measurement error, faults, change in daily
water demand and water quality variations that may occur
during the water treatment process.
Figure 7 Control moves of the ow rates or input variables.
Coagulation process control in water treatment plants using multiple model predictive control 947

4. Conclusions II. Normal operating region

Operating point = 1:2 106 leq=mg 7.5 pH unit]
State space matrix:
In this study, the application of MMPC control strategy to
2 3
achieve effective coagulation control in water treatment plants 0:8923 1:523  106 3:963  107 1:982  107 5:946  107
6 7
has been demonstrated. The proposed control strategy handles 6 287:4 0:9873 0:01797 0:008983 0:0269 7
6 7
explicitly the constraints on the control inputs resulting in sat- A6
6 0 0 0 0 0 7
7 :4
6 7
isfactory performance in the presence of sudden changes in 4 0 0 0 0 0 5
chemical reagent ow rates. The main contribution of this 0 0 0 0 0
study is the control of challenging coagulation using MMPC 2 3
design. The simulation studies performed on the chemical dos- 0 0 0
6 7
ing unit show the feasibility of the MMPC to achieve better 60 0 07
6 7
performance where other conventional control strategies might B6
61 0 077 :5
be inadequate. 6 7
40 1 05
In order to extend this study that has been based on com-
0 0 1
puter simulation only. The use of a pilot plant or laboratory
scale water treatment plant would be considered to evaluate, " #
0:8923 1:523  106 3:963  107 1:982  107 5:946  107
compare and improve the performance of the proposed control C :6
287:4 0:9873 0:01797 0:008983 0:0269
scheme.
Furthermore, the local models for MMPC were obtained III. Overdose operating region
by linearisation around the operating points. When the operat- Operating point = 4:76  105 leq=mg 10 pH unit]
ing points change, the linearised models will be inaccurate to State space matrix:
predict the responses of the dosing unit. In this regard, studies
2 3
will be performed to develop and use nonlinear models which 0:9243 1:602  106 6:808  108 3:404  108 1:021  107
6 7
can give better representation of the dosing unit. 6 6897
6
0:8626 0:0849 0:004245 0:1274 7
7
The switching mechanism used in this study could be fur- A6
6 0 0 0 0 0 7
7
6 7
ther improved to produce smooth transition or continuity 4 0 0 0 0 0 5
when moving from one model to another. Future work will 0 0 0 0 0
focus on using different switching techniques for the MPC :7
local controllers to improve the close-loop performance and 2 3
0 0 0
stability of the control system. 6 7
60 0 07
6 7
Acknowledgements B6
61 0 077 :8
6 7
40 1 05
The authors express their gratitude to Tshwane University of
0 0 1
Technology, Pretoria for their support. Special thanks to Prof.
" #
Guoyuan Qi and anonymous reviewers for their useful com- 0:9243 1:602  106 6:808  108 3:404  108 1:021  107
ments, and Rietvlei water treatment plant, City of Tshwane C :9
6897 0:8626 0:0849 0:004245 0:1274
for granting us permission to use their facility for this study.

Appendix A. Appendix
References
Operating regions of the coagulation chemical dosing unit
[1] A. Adgar, C.S. Cox, C.A. Jones, Enhancement of coagulation
control using the streaming current detector, Bioprocess Biosyst.
I. Underdose operating region Operating point = Eng. (2005) 349357.
5:73  105 leq=mg 5.9 pH unit] [2] P. Bagheri, V. Mardanlor, A. Fatehi, Multiple model predictive
State space matrix: control of multivariable pH process using adaptive weighting
2 3
1:078 1:588  106 1:887  106 9:385  107 2:815  106 matrices, in: Proceedings of the 18th IFAC World Congress,
6 7 2011 Milano, 2011, pp. 1236612371.
6 4:026 0:174 0:1157 0:05787 0:1736 7
6 7
A6
6 0 0 0 0 0 7
7 :1 [3] O. Bello, Y. Hamam, K. Djouani, Dynamic modelling and
6 7 system identication of a coagulant dosage system for water
4 0 0 0 0 0 5
0 0 0 0 0 treatment plants, in: Proceedings of the 3rd International
Conference on Systems and Control, 2013, Algiers, 2013.
2 3 [4] A. Bemporad, A. Morari, M. Morari, N.L. Ricker, 2013, Model
0 0 0
6 7 Predictive Control Toolbox: Users Guide, Natick, The
60 0 07 MathsWorks, 2013.
6 7
B6
61 0 077 :2 [5] City of Tshwane. S.a., Rietvlei Water Treatment Plant. <http://
6 7 www.tshwane.gov.za> (accessed 20.01.13).
40 1 05
[6] D. Dougherty, D. Cooper, A practical multiple model adaptive
0 0 1 strategy for single-loop MPC, Control Eng. Pract. 11 (2003)
" # 141159.
1:078 1:588  106 1:887  106 9:385  107 2:815  106 [7] M. Galvez-Carrillo, R. De Keyser, C. Ionescu, Nonlinear
C :3
4:026 0:174 0:1157 0:05787 0:1736 predictive control with dead-time compensator: application to
a solar power plant, Sol. Energy 83 (2009) 743752.
948
[8] L. Giovanini, M. Grimble, Predictive feedback control, in:
Proceedings of the 42nd IEEE Conference on Decision and
Control, vol. 1, 2003, pp. 923928.
[9] M. Gopinathan, J.D. Boskovic, R.K. Mehra, C. Rago, A
multiple model predictive scheme for fault-tolerant ight control
design, in: Proceedings of the 37th IEEE Conference on
Decision and Control, vol. 2, 1998, pp. 13761381.
[10] H. Guolian, Z. Jinfang, L. Junjun, Z. Jianhua, Multiple model
predictive control based on fuzzy adaptive weights and its
application to main steam temperature in power plant, in:
Proceedings of 5th IEEE Conference on Industrial Electronics
and Applications, 2010.
[11] B. Grosman, D.R. Lewin, MPC using nonlinear models
generated by genetic programing, in: G. Raqul, J. Sten Bay
(Eds.), Computer Aided Chemical Engineering, Elsevier, 2001.
[12] A.W. Hermansson, S. Syae, S.B. Mohd Noor, Multiple model
predictive control of nonlinear pH neutralisation system, in:
IEEE International Conference on Industrial Engineering and
Engineering Management, Macau, China, 2010, pp. 301304.
[13] N. Li, S. Li, Y. Xi, Multi-model predictive control based on the
TakagiSugeno fuzzy models: a case study, Inform. Sci. 165
(2004) 247263.
[14] F. Manenti, Considerations on nonlinear model predictive
control techniques, Comput. Chem. Eng. 35 (2011) 24912509.
[15] A.H. Mazinan, N. Sadati, Fuzzy multiple model predictive
control of tubular heat exchanger, in: IEEE World Congress on
Computational Intelligence, Hong Kong, 2008, pp. 18451852.
[16] M. Morari, J.H. Lee, Model predictive control: past, present and
future, Comput. Chem. Eng. 23 (1999) 667682.
[17] NSF International, NSF Product and Service Listings, 2013.
<http://info.nsf.org/> (accessed 19.08.13).
[18] L. Ozkan, M.V. Kothare, C. Georgakis, Control of a solution
copolymerisation reactor using multi-model predictive control,
Chem. Eng. Sci. 58 (2003) 12071221.
O. Bello et al.
[19] G. Pannocchia, Robust disturbance modelling for model
predictive control with application to multivariable ill-
conditioned processes, J. Process Control 13 (2003) 693701.
[20] C.R. Porrio, E. Almeida Neto, D. Odloak, Multi-model
predictive control of an industrial C3/C4 splitter, Control Eng.
Pract. 11 (2003) 765779.
[21] A. Shamasaddinlor, A. Fatehi, A.K. Sedigh, M.M. Karimi,
Study of multiple predictive control on a pH neutralisation
plant, in: Proceedings of the 9th Asian Control Conference,
Istanbul, 2013.
[22] X. Shenggang, Z. Lifang, M. Ailiang, Z. Luwen, A new
switching scheme for multi-model predictive control using
clustering modelling, in: Proceedings of the Fifth International
Conference on Fuzzy Systems and Knowledge Discovery, 2008,
pp. 484488.
[23] R. Shridhar, D.J. Cooper, A novel tuning strategy for
multivariable model predictive control, ISA Trans. 36 (1997)
273280.
[24] M. Soliman, O.P. Malik, D.T. Westwick, Multiple model
predictive control for wind turbines with doubly fed induction
generators, IEEE Trans. Sustain. Energy 2 (2011) 215225.
[25] C. Yu, R.J. Roy, H. Kaufman, B.W. Bequette, Multiple-model
adaptive predictive control of mean arterial pressure and cardiac
output, IEEE Trans. Biomed. Eng. 39 (1992) 765778.
[26] J. Zeng, D. Xue, D. Yuan, The study of multiple model
predictive control in wastewater treatment processes, in:
Proceedings of the Seventh International Conference on
Machine Learning and Cybernectics, Kunming, 2008, pp.
21662169.
[27] J. Zhang, J. Morris, Nonlinear model predictive control based
on multiple local linear models, in: Proceedings of the American
Control Conference, Arlington, VA, 2001, pp. 35033508.