Control Engineering Practice 44 (2015) 89103

Contents lists available at ScienceDirect

Control Engineering Practice

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

A gradient optimization scheme for solution purication process

Bei Sun a,c, Weihua Gui a,b,n, Yalin Wang a,b, Chunhua Yang a,b, Mingfang He a,d
a

School of Information Science and Engineering, Central South University, 410083, China
Institute of Control Engineering, Central South University, 410083, China
Department of Electrical and Computer Engineering, Polytechnic School of Engineering, New York University, 11201, United States
d
Department of Computer Science and Engineering, Polytechnic School of Engineering, New York University, 11201, United States
b
c

art ic l e i nf o

a b s t r a c t

Article history:
Received 1 June 2014
Received in revised form
12 July 2015
Accepted 13 July 2015
Available online 12 August 2015

This paper presents a two-layer control scheme to address the difculties in the modeling and control of
solution purication process. Two concepts are extracted from the characteristics of solution purication
process: additive utilization efciency (AUE) and impurity removal ratio (IRR). The idea of gradient
optimization of solution purication process, which transforms the economical optimization problem of
solution purication process into nding an optimal decline gradient of the impurity ion concentration
along the reactors, is proposed. A robust adaptive controller is designed to track the optimized impurity
ion concentration in the presence of process uncertainties, disturbance and saturation. Oxidation reduction potential (ORP), which is a signicant parameter of solution purication process, is also used in
the scheme. The ability of the gradient optimization scheme is illustrated with a simulated case study of a
cobalt removal process.
& 2015 Published by Elsevier Ltd.

Keywords:
Solution purication
Gradient optimization
Additive utilization efciency
Impurity removal ratio
Robust adaptive control
Oxidation reduction potential

1. Introduction
Solution purication, which belongs to the more general area of
separation science and technology, is a key step in hydrometallurgy (Flett, 1992). As a widely used approach to obtain pure
metals from their raw ores, hydrometallurgy involves phase and
status transforms of metal elements. Typically, a hydrometallurgy
process is composed of leaching, purication and electrowinning.
The raw ore is rst treated in the leaching process, in which the
valuable metal in the solid state ore is extracted and converted
into soluble salts in liquid solution. As leaching process is not
completely selective, pregnant leaching solution inevitably contains undesired impurity ions. The presence of these impurity ions
would decrease the current efciency in the subsequent electrowinning process in which pure metal is recovered, resulting in
energy waste and downgrade of product quality (Bckman &
stvold, 2000). Therefore, the pregnant leaching solution is required to be puried to a certain degree prior to nal metal
winning.
Owing to the heterogeneous property of raw ores, the types of
impurity ions in the leaching solution are not unique. Consequently, a solution purication process is composed of several substeps designed to remove different impurities. For example,
n
Corresponding author at: School of Information Science and Engineering,
Central South University, 410083, China. Fax: 86 731 88876677.
E-mail address: gwh@mail.csu.edu.cn (W. Gui).

http://dx.doi.org/10.1016/j.conengprac.2015.07.008
0967-0661/& 2015 Published by Elsevier Ltd.

solution purication of zinc hydrometallurgy consists of copper

removal, cobalt removal and cadmium removal (Fig. 1). These
impurity ions possess different physical and chemical properties.
The technologies and reaction conditions adopted in the sub-steps
are not the same. However, these different sub-steps do share
some common features. The reactions conducted to remove different impurities are essentially oxidation reduction reactions and
require the use of additive and in some occasions catalyst (Sun,
Gui, Wu, Wang, & Yang, 2013).
The dosage of additive is crucial to both purication performance and production cost. An excessive amount of additive is a
waste of costly material, while an insufcient amount fails to remove the impurity adequately (Kim, Kim, Park, Song, & Jung,
2007). However, due to the inherent complexity of purication
reaction, uctuation of previous leaching process, stochastic disturbance and interactions between the intermediate sub-steps
inside purication process, it is difcult for the human operators to
adjust additive dosage precisely in order to achieve economical
and stable operation. As a consequence, some operators prefer to
use an excessive amount of additive to achieve the required purication performance. More seriously, a large process uctuation
may even cause the failure to meet the desired purication degree.
The existence of these problems has attracted the attention of
researchers from both metallurgy and control community. To the
author's best knowledge, the study on solution purication process begins from 1871 (Bckman & stvold, 2000). After that, the
research on solution purication process has passed through two

90

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

Fig. 1. Solution purication process in zinc hydrometallurgy.

stages of rapid development. The rst stage of rapid development

is driven by the mechanism study of solution purication process
(Boyanov, Konareva, & Kolev, 2004; Bckman & stvold, 2000;
Brve & stvold, 1994; Dib & Makhlou, 2007, 2006; Dreher,
Nelson, Demopoulos, & Filippou, 2001; Fugleberg & Jarvinen, 1993;
Lew, 1994; Nelson, Wang, Demopoulos, & Houlachi, 2000; Polcaro,
Palmas, & Dernini, 1995; Sun et al., 2013; Tozawa, Nishimura,
Akahori, & Malaga, 1992; van der Pas & Dreisinger, 1996; Yang, Xie,
Zeng, Wang, & Li, 2006). It is found that the impurity removal
reaction is rst order solid-liquid phase kinetics/electrode reaction
with its reaction rate affected by dosage and particle size of additive, temperature, pH and type of catalyst, etc. The second stage
of rapid development is caused by the application of control theory to solution purication process which is usually conducted in
CSTRs (Continuous Stirred Tank Reactors). Wu (2001) studied the
use of a LMI (Linear Matrix Inequality) based robust model predictive controller to an industrial CSTR (Continuous Stirred Tank
Reactor) problem with explicit input and output constraints.
Knapp, Budman, and Broderick (2001) applied to a CSTR process
an adaptive algorithm which uses a neural network representation
to learn the process on-line. Antonelli and Astol (2003) studied
the output feedback regulation of endothermic and exothermic
chemical reactors in the presence of control bounds. Yu, Chang,
and Yu (2007) proposed a stable self-turning PID (Proportional
Integral Derivative) control scheme for multivariable nonlinear
systems with unknown dynamics and applied the scheme to a
simulated CSTR process. Di Ciccio, Bottini, Pepe, and Foscolo (2011)
built a nonlinear feedback control law for a CSTR with recycle by
using tools of differential geometry and observer theory. Hoang,
Couenne, Jallut, and Le-Gorrec (2012) developed nonlinear control
laws for isothermal CSTR based on the Lyapunov method. The
above results mainly focus on the control of a single reactor. Wang,
Gui, Teo, Loxton, and Yang (2012) and Li, Gui, Teo, Zhu, and Chai
(2012) studied the use of controlparametrization method to
minimize the zinc dust consumption of a zinc sulphate electrolyte
purication process composed of multiple reactors.
Different from the above research, this paper develops a control
scheme for solution purication process based on its characteristics. The concepts of Additive Utilization Efciency (AUE) and
Impurity Removal Ratio (IRR) are proposed based on an analysis of
solution purication process. By using these two concepts, the
control of solution purication process is decomposed into two
problems, i.e., estimated additive dosage optimization and robust
adaptive tracking control of the optimized operating point. Correspondingly, the proposed control scheme is composed of two
layers. The upper layer works on a slow time scale. The additive
dosage optimization, which has an economic objective function
subject to constraints on purication performance and process
stability, is transformed into nding an optimal decline gradient of
impurity ion concentration along the reactors. On contrast, the
lower layer works on a fast time scale. A robust adaptive controller,
in which Oxidation Reduction Potential (ORP) plays a central role,
is designed to track the optimized impurity ion concentrations in
the presence of model uncertainties, disturbance and saturation.
The rest of this paper is organized as following. In Section 2, an
analysis of solution purication process is conducted. The problems arising in the control of solution purication are pointed
out. The two-layer control scheme is introduced in detail in Section 3. The ability of the scheme is tested and discussed in Section
4. The concluding remarks are given in Section 5.

Fig. 2. Solution purication process.

2. Process analysis
Solution purication process is a continuous process composed
of N ( N 1, N Z ) consecutive reactors and a thickener in which
the liquidsolid separation takes place (Fig. 2 shows a solution
purication process composed of four reactors and a thickener).
Consider the reaction described by Eq. (1), along the reactors,
impurity ion B is gradually reduced by reaction with additive A
under specic reaction conditions and the assistance of catalyst.
The overow of the thickener is delivered to subsequent process.
The underow which contains crystal nucleus benecial to impurity removal is recycled to promote cementation

mB n + + nA = nAm + + mB

(1)

The technical index and the economical index of solution purication process are the impurity ion concentration of the puried
solution which reects the purication performance, and additive
consumption which relates to the production cost, respectively.
2.1. Process model
The reaction kinetics of Eq. (1) can be described by a rst-order
kinetic equation

dc
= kA s c
dt

(2)

in which, c is the concentration of impurity ion, k is the reaction

rate, As is the reaction surface area available for impurity removal
in unit volume of the reactor.
Consider the process described in Fig. 2, assume that the uid
in each reactor is perfectly mixed, and the contents are uniform
throughout the reactor volume. Then according to the mass balance principle, the dynamics of the process can be described by
following equations:

dc1
Fin
Fin + Fu
=
c0
c1 k1A s1c1
dt
V
V
dci
Fin + Fu
Fin + Fu
=
ci 1
ci ki A si ci
dt
V
V
i = 2, 3, 4

(3)

in which V is the volume of the reactor, Fin is the ow rate of the

impure input solution from previous stage, Fu is the ow rate of
the recycled underow solution, c0 is the impurity ion concentration of the input solution, cj is the efuent impurity ion
concentration of the jth reactor ( j = 1, 2, 3, 4 ), kj and A s j are the
reaction rate and reaction surface area in unit volume of the jth
reactor ( j = 1, 2, 3, 4 ), respectively.
2.2. Role of ORP
Purication process is essentially an oxidationreduction reaction and also an electrode reaction composed of many parallel
electrode reactions. According to the independence principle of
parallel electrode reactions (Antropov & Beknazarov, 1972), the
electrode reactions are independent of each other. Their unique
shared characteristic is the electrode potential, which is also called

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

91

mixed potential and determines the rate of all the electrode reactions by affecting the electron transfer rate between oxidant and
reductant.
ORP, which represents the comparative oxidability or reducibility of the solution and the extent of oxidation reduction
reaction, is highly related to the mixed potential of electrode reactions. The relation between ORP and reaction rate has been
studied in Sun et al. (2013). It is found that in a certain range, a
more negative ORP represents a faster reaction rate and vice versa

Ee + 2F (emix eeq )
k = A f exp

RTc

(4)
Fig. 4. Internal and external environments of purication process.

emix = peorp + q

(5)
1.6

2.3. Control difculties arising in solution purication process

Similar with most industry processes, purication process interacts with both internal and external environments (Fig. 4). Eq.
(3), which describes only the ideal reaction kinetics, is not a sufcient description of the overall dynamics.
Specically, the dynamics of solution purication is inuenced
by external and internal environments mainly from the following
aspects:
1. The raw ore contains intricately mixed minerals with randomly
varying properties. And the operations taken in the preceding
processes of solution purication may not always leading to
satisfying results. As a result, the characteristics of the input
solution, which include the impurity ion concentration and
types of elements in the solution, are time varying.
2. Under some circumstances, the usually xed conguration of
the reaction conditions needs to be adjusted, such as the ow
rate of underow, type of catalyst, stirring rate, and reaction
temperature.
3. In a certain range, a more negative ORP will result in a faster
reaction rate. However, there exists a point beyond which the
effect of ORP on the reaction rate is saturated or even reversed.
Because if ORP is very negative, it is likely to generate more
metal subsulfate which will attach to the surface of additive and
reduce its activity;
Correspondingly, the control difculties arising in solution
purication include:

Fig. 3. ORP-additive dosage controller.

Reaction rate and ORP under different conditions

Condition 1
Condition 2
Condition 3

1.4
1.2
Reaction rate (s )

where k is the reaction rate, emix is the mixed potential, eorp is ORP
of the solution, Af is the frequency factor, Ee is the standard activation energy, F is the Faraday constant, R is the ideal gas constant,
Tc is the reaction temperature, eeq is the equilibrium potential, is
the inuence factor of electrode potential variation to cathode
activation energy, p and q are the linear approximation parameters
to be identied.
Due to its signicance, in some plants (Fugleberg & Jarvinen,
1993; Sun et al., 2014), the reaction rate is controlled by adjusting
the setting value of ORP which can be monitored online continuously and controlled by changing the dosage of additive
(Fig. 3). Thus, from the above discussion, ORP is an external representation of the mixed potential, and an intermediate variable
between reaction rate and additive dosage.

x 10

Same ORP result in different reaction rate

0.8
0.6
0.4
0.2
0
500

510

520

530
540
ORP (mv)

550

560

570

Fig. 5. Relation between ORP and reaction rate under different conditions.

1. Model uncertainties: The impurity ion concentration and type of

elements in the input solution are time varying. Also the reaction conditions are not consistent. Thus, the parameters in the
process model are not constant. As illustrated by Fig. 5, under
different conditions, the same setting value of ORP would result
in different reaction rates.
2. Disturbance: Equipment failure, external excitation and unreasonable operations would bring disturbance to solution
purication process.
3. Saturation: The relation between ORP and the reaction rate exhibits a saturation-like phenomenon. In addition, subjects to
the physically constraints, for each reactor, the setting value of
ORP is bounded in an appropriate range (Fig. 6).
The control objective of solution purication process is to reduce the impurity ion concentration to a predened acceptable
range using the smallest amount of additive dosage and keep the
process stable. In order to get rid of the control difculties and
achieve the control objective, a two-layer gradient optimization
scheme is proposed in next section.

3. Gradient optimization scheme

For almost every industrial process, the realization of control
objective relies on a subtly designed control scheme. In this section, based on two intuitional concept derived from the characteristics of solution purication process, a two-layer gradient
optimization scheme is developed. The control problem is formulated and transformed into a constrained optimization

92

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

required purication performance.

Bounded input and input saturation

x 10

Reaction rate (s1)

3.2. Basic idea of gradient optimization

Using the two concepts, for a block of leaching solution with
volume V and impurity ion concentration c0, assume that AUE of
each reactor is constant, then during its retention in a solution
purication process with N consecutive reactors, the total additive
dosage can be formulated as

1.5

zw =

0.5

zi =
i=1

where z w0 =

0
500

520

540

560
580
ORP (mv)

600

620

Fig. 6. Bounded input and input saturation.

problem, or in other words, nding an optimal decline gradient of

impurity ion concentration along the reactors. The control difculties, which include model uncertainties, disturbance and saturation, are handled by a robust adaptive controller.
3.1. Concept design
3.1.1. Additive utilization efciency
The amount of additive fed into the reactor affects the purication effect and production cost. A major problem in deciding
the additive dosage is that not all the additive added involve in
impurity removal (Sun, Gui, Wang, & Yang, 2014). According to the
mass balance principle, the proper amount of additive to be fed
into the reactor depends on its efciency in removing the impurity
(Kim et al., 2007).
Concept 1: Consider a certain period of time, Additive Utilization Efciency (AUE) is the ratio of additive practically involved in
impurity removal when a certain amount of additive is added into
the reactor

z real
=
z all

(6)

where zall is the amount of additive added, zreal is the amount of

additive involved in impurity removal.
Consider a block of solution with volume V, for reactor i
(i = 1, 2, , N ), denote i as its AUE, ci 1 and ci as the impurity ion
concentration before and after its retention in the reactor, assume
that i is constant during the retention, then the required additive
dosage is

z i = i 1

MA
V (ci 1 ci )
MB

(7)

where MA and MB are the atomic weight of additive and impurity,

respectively.
3.1.2. Impurity removal ratio
Concept 2: Consider a block of leaching solution with volume V
and impurity ion concentration c0, its retention in solution purication process is essentially a gradually decline process of the
impurity ion concentration along the reactors. For reactor i
( i = 1, 2, , N ), the Impurity Removal Ratio (IRR) is the ratio of
impurity ion removed in it

ci 1 ci
i =
c0

(8)

IRR of each reactor needs to be arranged in order to achieve the

MA
Vc0
MB
MA
Vc0
MB

i 1 i = z w 0 i 1 i
i=1

i=1

(9)

is the ideal additive dosage needed to remove

all the impurity ion.

Eq. (10) indicates that AUE can not only be used to estimate
additive dosage of a single reactor, but also optimize additive
consumption of purication process composed of multiple reactors. If more IRR is assigned to the reactor with larger AUE, fewer
IRR is assigned to the reactor with smaller AUE, the overall additive consumption could be optimized. Moreover, IRR of a reactor
reects its internal reaction state. A stable IRR indicates a stable
reaction status. So limiting IRR of each reactor in suitable ranges is
benecial to the stability of purication process.
Thus for a block of leaching solution with volume V and impurity ion concentration c0, assume that AUE of each reactor are
constant, then the problem of optimizing the additive consumption can be formulated as an allocation problem of IRR according
to AUE while considering the stability and purication requirement:
N

min

z w (1, 2, , N ) = z w0

i 1 i
i=1

st.

0 < c N = c0 1

mini i maxi ,

i cdp
i=1
N

i = 1, 2, , N

(10)

where cdp is the desired efuent impurity ion concentration, mini

and max i are predened lower and upper bounds of IRR of reactor
i.
3.3. Two layer control scheme
Solving Eq. (10), which would obtain the optimized IRR of each
reactor, is equal to nd the optimized setting values of the efuent
impurity concentration of each reactor, or in other words, nd a
best decline gradient of impurity ion concentration along the reactors (Fig. 7). However, due to the control difculties discussed in
Section 2, solving Eq. (10) itself is not sufcient to achieve the
required purication performance. A robust adaptive controller,
which is capable of driving the efuent impurity ion concentration
of each reactor to follow the optimized value, is required to be
designed. Thus, a complete control scheme of solution purication
process should include two layers (Fig. 8). The upper layer, which
works on a slow time scale, solves the estimated economical optimization problem. The lower layer, which works on a fast time
scale, handles the model uncertainties, disturbance and saturation.
3.4. Upper layer: estimated economical optimization
The main task of the upper layer is to calculate the best decline
gradient of impurity ion concentration according to the impurity
ion concentration in the leaching solution c0 and AUE of the reactors i ( i = 1, 2, , N ). However, for each reactor, AUE is related
to the reaction status and changes with time. The impurity ion

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

93

Gradient optimization

Impurity ion concentration

c0

Gradient 1
Gradient 2
Gradient 3
Gradient 4
Required technical index
out of predefined range of IRR
failed to meet the required
purification performance

Fig. 9. Structure of a RBF neural network.

within
predefined
range of IRR
0
0

t1

output value of the ith neuron in the hidden layer. The most
commonly used radial basis function is the Gaussian function:

t2
time

t3

t4

Fig. 7. Gradient optimization along reactors (ti means the time when solution
outows from reactor i (i = 1, 2, , N )).

2 / 2
i

i ( ) = e ci

(12)

where c i is the Euclidean distance between and c i , c i and

si are center and spread of the ith ( i = 1, 2, , m) node in the
hidden layer.
The performance of a RBF neural network is determined by the
number of hidden neurons, center and spread of each neuron and
the connecting weight = [1, 2, , m ]. According to the way the
centers are selected, there exist four types of training method,
including random center selection, self-organized center selection,
supervised center selection and orthogonal least squares learning
algorithm. In this paper, the network is trained using orthogonal
least squares (OLS) method (Chen, Cowan, & Grant, 1991), while
the aim is to minimize the following criterion:
P

min J =

Fig. 8. Two layer control scheme.

i i ( )
i=1

di |2

i=1

concentration of the leaching solution is also not always constant.

Owing to the time varying property of c0 and i (i = 1, 2, , N ), Eq.
(10) needs to be solved periodically. Selection of the optimization
period is related to the time-variation characteristic of the target
process. If c0 or i ( i = 1, 2, , N ) changes drastically and frequently, then a short optimization period is required. If c0 and i
( i = 1, 2, , N ) are stable with small variations, then a long optimization period is preferred.
Besides the selection of optimization period, the estimation of
AUE is another important issue. There exist various kinds of
methods to estimate the value of AUE (Kim et al., 2007), such as
regression method and BoxJenkins method. In this paper, Radial
Basis Function Neural Network (RBFNN) (Moody & Darken, 1989),
which has been applied successfully in many engineering problems (Han, Qiao, & Chen, 2012; Iliyas, Elshafei, Habib, & Adeniran,
2013; Seshagiri & Khalil, 2000), is selected to estimate AUE.
RBF neural network is a feedforward neural network composed
of three layers including one input layer, one hidden layer and one
output layer. Fig. 9 shows a RBF neural network with n input, m
hidden node and a single output.
The most signicant characteristic of RBF neural network is the
use of nonlinear activation function, i.e., radial basis function, in
the hidden layer. The output of a RBF neural network is a weighted
linear combination of the output of hidden neurons (Fig. 9)

y=

|yi

(11)

where = [1 , 2 , , n ]T and y are the input and the output of the

network, i ( i = 1, 2, , m) is the connecting weight between the
ith neuron in the hidden layer and the output layer, i is the

(13)

where P is the number of training data samples, d and y are the

practical value and the output of the RBF neural network,
respectively.
A complete development procedure of the RBF neural network
for AUE estimation includes following steps:
Step 1: Select the input variables . According to the reaction
mechanism, for the AUE estimation of each reactor, the input
variables should include ow rate and impurity ion concentration
of the inlet solution, ow rate of recycled underow, dosage of
additive and catalyst and ORP.
Step 2: Select the training data. The training data is selected
from routinely collected production data of real plants. The
training data should cover as many production mode as possible.
However, the training data must be selected carefully, the data
from the cases with an excessive additive dosage should be
avoided.
Step 3: Select the centers { c i } using OLS, and choose the spread
of each neuron as the closest Euclidean distance between its
center and centers of other neurons.
Step 4 : Adapt the connecting weights using the adaptive
gradient descent procedure (Iliyas et al., 2013).
3.5. Lower layer: robust adaptive controller of impurity ion
concentration
The function of the lower layer is to eliminate the effect caused
by the model uncertainties, disturbance and saturation. A controller which forces the impurity ion concentration to track the
desired reference is designed.
3.5.1. Nominal state space model
According to Eqs. (3)(5), the nominal state space model of the

94

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

Besides model uncertainties d g i ( , t ) and approximation error

process can be represented as

x = s 0 + Ax F (x) g (v)

(14)

where x = [x1, x2, , x N ]T = [c1, c2, , cN ]T is the outlet impurity

ion concentration of the reactors which is considered as the system state, v = [v1, v2, , vN ]T = [eorp1, eorp2, , eorpN ]T is the ORP
F

setting value of the reactors, s 0 = [ V c0, 0, , 0]T . A is the system

matrix

a 0
0

a a 0
0
a a
A=

0 0
0 0

0
0
0

a
0

0
0

a 0
a a

x 0 = f0 (, p, t )
x1 = a 0 x 0 ax1 x1ga1 (v1) + d1 ( , v1, t )
x j = ax j 1 ax j x j ga j (v j ) + d j ( , v j , t )

(22)

in which j = 2, 3, , N , x0 is the impurity ion concentration of the

input leaching solution, is the physical and chemical characteristics of raw ore, f0 ( , p, t ) is an unknown bounded function,
F
a0 = , di ( , vi, t ) = x i (d g ( , t ) + da (vi )) + d i (t ) (i = 1, 2, 3, , N ) is
i

(15)

F +F

In which, a = inV u . F (x) and g (v) are functions of state and input
variables, respectively

x1

0
F (x) =

dai (vi ), external disturbance d i (t ) also affects the process dynamics,

thus a more comprehensive description of the process is

the synthetical effect of model uncertainties, approximation error

and external disturbance.
Before the development of the robust adaptive tracking controller, following assumptions are made:
Assumption 1. Solution purication process is input-to-state
stable (ISS).

0 0

x2 0

0 xN

(16)

Assumption 2. Impurity ion concentration can be determined

online.
Assumption 3. Variables, such as impurity ion concentration and
the identied parameter sets i ( i = 1, 2, , N ), are bounded and
have their own physical sense.

g1 (v1)

g (v2 )
g (v) = 2

g
N (vN )

(17)

In which,

gi (vi ) = A si A fi e(Eei+ 2 i F [(pi v i + qi ) eeq ]) / RTc , i = 1, 2, , N

i = {Ai0 , Eei, i, pi , qi } is the
( i = 1, 2, , N ) to be identied.

parameter

set

(18)
of

reactor

3.5.2. Model uncertainties, disturbance and saturation

The discussion in Section 2.3 indicates that the dynamics of
purication process is not always consistent. The identied parameter sets i ( i = 1, 2, , N ) exhibit uncertainties
i
(i = 1, 2, , N ) which are time varying and depend on the internal
and external environments. So function gi (vi ) can be written as

gi (vi ) = gi (vi ) + d g i ( , t )

(19)

]T ,

In which, = [ p, c
g i (vi ) and d g i ( , t ) are the deterministic part
and the uncertain part of gi (vi ), respectively.
In Eq. (19), g i (vi ) is saturated when input vi exceeded certain
value as shown in Fig. 6. There is a sharp corner when the saturation happens, thus backstepping technique cannot be directly
applied. In order to remove this barrier, the saturation is approximated by a smooth function (Wen, Zhou, Liu, & Su, 2011)

ai vi + bi vi2 + ci if vi vMi
gai (vi ) =

if vi < vMi
gi (vi )

3.5.3. Robust adaptive tracking controller

In this section, reactor 1 is used as an example to illustrate the
design of robust adaptive controller with the additive dosage as
the control input.
Dynamics of reactor 1 is approximated and augmented as follows:

x 0 = f0 (, p, t )
x1 = a 0 x 0 ax1 h1 + d1 (t )
h1 =
where h1 = x1ga1 (v1),

(23)

is an auxiliary control to be determined.

The bound of d1 (t ) is denoted as D1 which is not assumed to be

known. v1 is determined for a given state x1 and h1.
The following change of coordinates is made:

z1 = x1 x r1
z2 = h1 1

(24)

where xr1 is the optimized setting value of x1, z1 is the tracking

error of impurity ion concentration, z2 is due to the new state
variable h1, 1 is the virtual control law to be designed.
The backstepping procedure is developed below:
Step 1: Consider the Lyapunov function

Vz1 =

1 2
z1
2

(25)

The derivative of Vz1 is

Vz 1 = z1z1
= z1 [a 0 x 0 ax1 z2 1 + d1 (t ) xr1]

(26)

in which the parameter ai, bi, ci and vM i are chosen according to

g i (vi ) (Fig. 6).
Then g i (vi ) can be expressed as

If d1 (t ) and x0 are bounded by D1 and D0. Due to the fact that

z1d1 (t ) |z1|D1, z1x0 |z1|D0

gi (vi ) = gai (vi ) + dai (vi )

Vz 1 z1 [ ax1 z2 1 xr1] + a 0 |z1|D0 + |z1|D1

(20)

Design

where dai (vi ) is the bounded approximation error

|dai (vi )| = |gi (vi ) gai (vi )| Dai

(21)

(27)

1 as

^
1 = c1z1 ax1 xr1 + a 0 sgn (z1) x^0 + sgn (z1) D1

(28)

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

with the adaptive law

x^0 = 0 a 0 |z1|

(29)

where x^0 is the estimation of D0.

Thus for the following Lyapunov function:

V1 = Vz1 +

(30)

where x 0 = D0 x^0 is the estimation error of D0.

The following equation is obtained:

1
x 0 x 0
0

(31)

Step 2: Consider the Lyapunov function

(32)

with its derivative

= z2 { [(c1 a) x1 + a 0 sgn (z1) x^0 c1xr1

^
x r(12) + sgn (z1) D1]}

z 3 cv if z 3 cv

sgv (z 3 ) = 0
if |z 3 | < cv

z 3 + cv if z 3 cv

z2 { [(c1 a)(a 0 x 0 ax1 h1

(33)

^
Thus, for the following Lyapunov function, where D 3 = D3 D3
is the estimation error of D3, Theorem 2 can be obtained

(34)

(35)

For the following Lyapunov function:

1 2
D1
21

V2 = V1 + Vz2 +

(36)

where D 1 = D1 D^1 is the estimation error of D1.

The following equation is obtained:

V2 = z1z1 + z2 z2 +

1
1
x 0 x 0 + D 1D 1
0
1

c2 z22

(37)

Theorem 1. For solution purication process described by Eq. (1), it

is global asymptotically stable by applying the control law:

h1 =

1
2
2
2 (z 3 cv ) + D3 /(23 ) if z 3 cv
2
V3 = D 3 /(23 )
if |z 3 | < cv

2
1
(z + c )2 + D /(2 ) if z c
3
v
v
3
3
2 3

Theorem 2. For the ORP-Additive dosage system described by Eq.

(39), the approximation error of v1 by using approximation Eq. (40)
converges to [ cv, cv ], with that cv can be dened arbitrarily small
by the user.
Proof. The derivative of V3 is

Based on Steps 1 and 2, Theorem 1 can be derived.

v1 = g a1 1(h1/x1)

where cv is a positive design parameter.

with the adaptive law

D1 = 1 (|z1| + |z2 (c1 a)|)

sgn (z 3 ) if |z 3 | cv
sat (z 3 ) =
if |z 3 | < cv
z 3/cv

Design

= c2 z2 + z1 + (c1 a)(a 0 x 0 ax1 h1)

^

sgn (z2 )|c1 a|D1 + a 0 sgn (z1) x^0 c1xr1

^
x r(12) + sgn (z1) D1

(40)

^
+ a 0 sgn (z1) x^0 c1xr1 x r(12) + sgn (z1) D1]}

sgn (z2 ) sgn (c1 a) D1) + a 0 sgn (z1) x^0

^
c1xr1 x r(12) + sgn (z1) D1]}

^
D3 = 3 |sgv (z 3 )|

which is the bound of 1 (x1, , v1, t ), function sgv (z3 ) and sat (z3 ) are
dened as

= z2 { [(c1 a)(a 0 x 0 ax1 h1 + d (t ))

with

in which 1 and the dynamics of 1 (x1, , v1, t ) are unknown.

Before the development of the control law, an auxiliary approximation system is designed (Rincn, Erazo, & Angulo, 2012)

in which, fv (u1) = 1 (x0 1u1), c3 is a positive design parameter,

^
z3 = v1 v^1 is the approximation error, D3 is the estimate of D ,

Vz 2 = z2 z2 = z2 ( 1 )

(39)

^
v^1 = fv (u1) + c3 sgv (z 3 ) + D3 sat (z 3 )

1 2
z2
2

c1z12

Steps 1 and 2 deduced the reference setting value of ORP. In

order to obtain suitable additive dosage that force the practical
ORP to follow its setting value, Steps 3 and 4 are required.
Step 3: Consider the ORP-Additive dosage system

v1 = 1 (x 0 1u1) + 1 (x1, , v1, t )

c1z12 z1z2 + |z1|D 1

Vz2 =

= c2 z2 + z1 + (c1 a)(a 0 x 0 ax1 h1)

^

sgn (z2 )|c1 a|D1 + a 0 sgn (z1) x^0 c1xr1

^
x r(12) + sgn (z1) D1
and adaptive laws (29) and (35).

1 2
x 0
2 0

V1 = z1z1 +

95

(38)

V3 = sgv (z 3 ) z 3 +

1
D3 D3
3

(41)

As z3 = v1 v^1 = fv (u1) + 1 (x1, , v1, t ) v^1:

V3 = sgv (z 3 )(fv (u1) v^1) + sgv (z 3 ) 1 (x1, , v1, t ) + D 3 D 3

3
1

sgv (z 3 )(fv (u1) v^1) + |sgv (z 3 )|D3 + D 3 D 3

3
Using Eq. (40), the following equation is obtained:

(42)

96

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

^
V3 c3 sgv (z 3 )2 sgv (z 3 ) sgn (z 3 ) D3 + |sgv (z 3 )|D3
1
^
^
^
D 3 D3 + sgv (z 3 )[sgn (z 3 ) D3 sat (z 3 ) D3 ]
3
c3 sgv (z 3 )2 +

1
^
D3 (3 |sgv (z 3 )| D3 )
3

^
^
+ sgv (z 3 )[sgn (z 3 ) D3 sat (z 3 ) D3 ]

(43)

^
^
Notice that sgv (z3 )[sgn (z3 ) D3 sat (z3 ) D3 ] = 0, so

V3 c3 sgv (z 3 )2

(44)

According to the Lyapunov stability theorem (Khalil & Grizzle,

2002), the approximation error z3 asymptotically converges to ,
where = {z3: |z3 | cv }, the estimation error of D3 asymptotically
converges to zero.

Step 4 : Dene the tracking error as

z5 = vr1 v1
= vr1 v^1 + v^1 v1
= z4 z3

(45)

copper, cobalt, nickel, cadmium, indium, gallium, arsenic, antimony, and germanium. The presence of these metal ion impurities
would cause large drops of current efciency during electrowinning in which metallic zinc is recovered, resulting in energy waste
and downgrade of product quality. Thus, before electrowinning,
these metal ion impurities need to be removed to an acceptable
level. Among the impurity ions, cobalt ion is difcult to be removed and has the most detrimental effect to electrowinning, so
the result of cobalt removal is used as an index of solution purication performance.
Cobalt removal process studied in this section is composed of
four consecutive continuous stirred tank reactors and a thickener
(Fig. 2). Spent acid is supplied to provide an acid reaction environment. The solution is heated to around 80 C to guarantee
enough reaction impetus. Zinc dust and arsenic trioxide are added
into the reactors to conduct complex chemical and electrochemical
reactions with cobalt ions and residual copper ions from previous
copper removal process.
The main reactions taking place in cobalt removal process include the following:

Cu2 + + Zn = Zn2 + + Cu

(49)

As3 + + 3Cu2 + + 4.5Zn = Cu3 As + 4.5Zn2 +

(50)

As3 + + Co2 + + 2.5Zn = CoAs + 2.5Zn2 +

(51)

Consider z 4 = vr1 v^1, its derivative is

z 4 = vr 1 v^1
^
= vr 1 fv (u1) c3 sgv (z 3 ) D3 sat (z 3 )
^
= 11u1 1x 0 c3 sgv (z 3 ) D3 sat (z 3 ) + vr 1

(46)

If

u1 =

1
^
[1x 0 + c3 sgv (z 3 ) + D3 sat (z 3 ) vr 1 c4 z 4 ]
11

(47)

then

z 4 = c4 z 4

(48)

which means that z4 converges asymptotically to zero.

From Step 3, it is known that z3 asymptotically converges to
= {z3: |z3 | cv }. Hence the tracking error z5 converges to
= {z5: |z5 | cv } asymptotically.
Remark 1. Steps above provide a robust adaptive controller for
solution purication process. The closed loop system is globally
stable. The tracking and transient performance can be controlled
by adjusting the design parameters. Increasing c1, c2, c3 and c4, or
decreasing cv can improve the tracking performance while increasing 1, 2, 3 and 4 can decrease the effects of the initial
error estimates on the transient performance (Zhou, Wen, &
Zhang, 2004). However, increasing c1, c2, c3 and c4 indicates a
higher control input. Thus these design parameters are suggested
to be determined through try and error at the development stage.

4. Experimental study
In order to evaluate the feasibility and ability of the proposed
scheme, an experimental study of a cobalt removal process in a
zinc hydrometallurgy plant (Sun et al., 2014) is investigated in this
section.
4.1. Process description
Cobalt removal is an intermediary step of the solution purication process in zinc hydrometallurgy (Fig. 1). Due to the impurity of zinc concentrate, zinc sulfate solution after leaching
contains not only zinc ion, but also other metal ions, such as

By forming alloys or metal compounds, such as CoAs, cobalt ions is

gradually precipitated. After retention in four consecutive reactors,
zinc sulfate solution ows into the thickener in which liquidsolid
separation takes place. Overow of the thickener is delivered to
subsequent cadmium removal process, while the underow which
contains crystal nucleus benecial to cobalt removal is recycled to
the rst reactor.
According to Section 2.1, the nominal model of cobalt removal
process can be described by

dx1
Fin
Fin + Fu
=
x0
x1 k1A s1x1
dt
V
V
dxi
Fin + Fu
Fin + Fu
=
xi 1
x i ki A si x i
dt
V
V
i = 2, 3, 4
in which V is the volume of the reactor, Fin is the ow rate of the
leaching solution, Fu is the ow rate of the underow, x0 is the
impurity ion concentration of the leaching solution, xi is the efuent impurity ion concentration of the ith ( i = 1, 2, 3, 4 ) reactor,
ki and A s i are the reaction rate and reaction surface area in unit
volume of the ith ( i = 1, 2, 3, 4 ) reactor, respectively, and

Ee + 2F [(peorp + q) eeq ]

k = A f exp
RTc

(52)

A s = g s

(53)

where Af is the frequency factor, Ee is the standard activation energy, eeq is the equilibrium potential of cobalt reduction, Tc is the
reaction temperature, R is the idea gas constant, F is the Faraday
constant, gs is the precipitant content in unit volume of reactor,
eorp is the oxidationreduction potential, is the inuence factor of
electrode potential variation to cathode activation energy, is the
relation factor between surface area in unit volume of reactor and
gs, p and q are the approximation terms between oxidationreduction potential and mixed potential. The details of these

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

parameters are summarized in Table 1.

The control objective of cobalt removal process is to minimize
the zinc dust dosage and guarantee that the cobalt ion concentration after purication is equal to or smaller than a predened threshold cindex = 0.5 mg/L
4

min

z w (u1, u2 , u3, u4 ) =

ui
i=1

st.

cout cindex

4.2. Experiment setting

In this simulation, the practical cobalt removal process is simulated using a kinetic model of normal production mode. The
parameters in the kinetic model are identied from the production
data. The performance of the kinetic model is shown from Figs. 11
to 14.
To test the proposed scheme and mimic the real process, following situations are also considered, including:

(54)

in which, zw is the zinc dust consumption of four reactors, ui is the

zinc dust dosage of reactor i, cout is the cobalt ion concentration
after purication.
The manipulated variables of cobalt removal include dosage of
zinc dust and arsenic trioxide, setting value of ORP, reaction
temperature, ow rate of leaching solution and spent acid. Among
these variables, ow rate of leaching solution, which is related
with preceding leaching process and subsequent electrowinning
process, is decided from the prospective of the running of the
whole zinc hydrometallurgy process. Flow rate of spent acid and
reaction temperature are in most cases kept constant to provide a
suitable reaction environment. Dosage of arsenic trioxide is determined according to the stoichiometric relationship indicated by
Eqs. (49) (51). In the plant, the cobalt removal process is controlled by adjusting dosage of zinc dust or the setting value of ORP,
which can then be controlled by a ORP-zinc dust PI controller
automatically (Fig. 3).
However, as discussed in Section 2, cobalt removal is a complex
multiphase reaction inuenced by numerous factors, such as pH,
grain size of zinc dust, type of elements and copper ion concentration in the leaching solution. These inuences would affect
the process dynamics by causing variations in the parameters of
the model. For example, if is increased, then the process is more
sensitive to the variation of ORP; if is increased, then more reaction surface is available and the reaction would be accelerated.
Besides the inuence of these factors, cobalt removal process also
encounters saturation. An excessive zinc dust dosage may cause
the generation of basic zinc sulfate which would then block the
surface of zinc dust and thus hinder cobalt removal (Fig. 10 shows
a sample of the solution when excessive zinc dust is added).

1. Model uncertainties: The nominal model adopted in the robust

adaptive controller is different with the kinetic model of normal
production mode.
2. Variation of impurity ion concentration and ow rate of leaching
solution: The cobalt ion concentration in the leaching solution
varies from 22 mg/L to 36 mg/L as shown in Fig. 15. The
variation of the ow rate is shown in Fig. 16.
3. External disturbance: The disturbance shown in Fig. 17 is acted
on the rst reactor.
4.3. Simulation steps
The simulation interval is 24 h, and every optimization step is
10 min. At each optimization step, following steps are taken:
Step 1: AUE estimation. AUE of each reactor is estimated by a
RBF neural network with 15 centers and following inputs:
1.
2.
3.
4.
5.
6.
7.
8.

Flow rate of inlet solution.

Inlet cobalt ion concentration of the reactor.
Outlet cobalt ion concentration of the reactor.
Copper ion concentration in the leaching solution.
Zinc dust dosage.
Arsenic trioxide dosage.
Oxidation reduction potential.
Precipitant content in unit volume of the reactor.The structure
of the RBF neural network discussed in Section 3.4 is adopted
here. The estimation result is shown in Fig. 18.
Step 2: IRR calculation. Calculate the IRR of each reactor according to estimated AUE and Eq. (10) (Fig. 19). Then calculate the
setting values of the impurity ion concentration according to IRR
and cobalt ion concentration in the leaching solution (Fig. 20). The
constraints on the IRR of each reactor are shown in Table 2.
Step 3: ORP setting. Calculate the ORP setting values of each
reactor using the controller Eq. (38) (Fig. 21).

Table 1
Parameters in kinetic model.
Parameter Physical meaning

Unit

Value

V
gs

m3

450

A0
R

Volume of reactor
Precipitant content in unit volume of the
reactor
Relation parameter between surface area
and gs
Frequency factor of cobalt removal reaction
Ideal gas constant

Tc
F

Reaction temperature
Faraday constant

Ee

Standard activation energy of cobalt removal reaction

Mixed potential of present solution
Oxidation reduction potential of present
solution
Equilibrium potential of cobalt ion
Inuence factor of electrode potential
change to cathode activation energy
linear term of linear approximation function
offset term of linear approximation function

emix
eorp
eeq

p
q

97

kg/m3
m2m3/kg
1/s
J/(molK)
K
C/mol

8.314472
96485

kJ/mol
V
V
V
1
1
1
Fig. 10. Hinder effect caused by excessive zinc dust dosage.

98

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

0.7
Outlet cobalt ion concentration (mg/L)

Outlet cobalt ion concentration (mg/L)

4.5
Kinetic model
Practical value

3.5

Kinetic model
Practical value
0.6

0.5

0.4

0.3

0.2

10

20

30

40

50

60

70

10

80

60

70

80

40
Cobalt ion concentration

Kinetic model
Practical value

2.5
2
1.5
1

10

20

30

40
50
test point

60

70

35

30

25

20

80

Fig. 12. Kinetic model performance of reactor 2.

10 12 14 16 18 20 22 24
time (h)

Fig. 15. Impurity ion concentration of input solution.

Flow rate of input solution

1.2
320

Kinetic model
Practical value

Flow rate of input solution

300

Flow rate (m3/h)

Outlet cobalt ion concentration (mg/L)

40
50
test point

Cobalt ion concentration of leaching solution

0.5

30

Fig. 14. Kinetic model performance of reactor 4.

Cobalt ion concentration (mg/L)

Outlet cobalt ion concentration (mg/L)

Fig. 11. Kinetic model performance of reactor 1.

20

0.8

0.6

0.4

0.2

280
260
240
220

10

20

30

40
50
test point

60

Fig. 13. Kinetic model performance of reactor 3.

70

80

200

10

12 14
time (h)

16

Fig. 16. Flow rate of input solution.

18

20

22

24

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

99

Impurity removal ratio

Disturbance

50%

Disturbance
40%
Impurity removal ratio

Disturbance

30%

Reactor 1
Reactor 2
Reactor 3
Reactor 4

20%

10%

10 12 14
time (h)

16

18

20

0%

22 24

10 12 14
time (h)

16

18

20

22 24

Fig. 17. Disturbance.

Fig. 19. Optimized impurity removal ratio of each reactor.

Step 4 : Zinc dust dosage determination. The dosage of additive

is obtained via Eq. (47) and shown in Fig. 22.

3. There is no large uctuation and excessive increment in the

additive dosage.

4.4. Performance evaluation

An important phenomenon is that the setting value of AUE in

the rst reactor is kept at 70% throughout the simulation. This
indicates that by using the idea illustrated in Sections 3.1 and 3.2,
most of the cobalt ion was removed in the rst reactor whose AUE
is higher than the rest reactors. The economical effect is that the
total zinc dust dosage is decreased, while the over excessive dosage of additive is avoided. A practical application of this idea is in
Sun et al. (2014), in which a decrease in the zinc dust consumption
was observed.

The control result is shown in Figs. 23 and 24. It is indicated

that by using the two-layer control scheme:
1. The nal impurity ion concentration always satises the purication requirement.
2. The impurity ion concentration and oxidation reduction potential of each reactor can track their reference trajectories well.
AUE of reactor 1

AUE of reactor 2

100%

100%
Practical value
RBFNN

80%

60%

AUE

AUE

80%

60%

40%

40%

20%

20%

0%

12
time (h)

16

20

Practical value
RBFNN

0%

24

AUE of reactor 3

20

24

100%
Practical value
RBFNN

80%
60%

60%

40%

40%

20%

20%

12
time (h)

16

20

Practical value
RBFNN

80%

AUE

AUE

16

AUE of reactor 4

100%

0%

12
time (h)

24

0%

Fig. 18. Additive utilization efciency of each reactor.

12
time (h)

16

20

24

100

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

Setting values of impurity ion concentration of reactors

Outlet impurity ion concentration (mg/L)

15
Reactor1
Reactor2
Reactor3
Reactor4
10

10 12 14
time (h)

16

18

20

22 24

Fig. 20. Setting value of the impurity ion concentration of each reactor.

Table 2
Constraints on IRR of each reactor.
Reactor

Reactor 1

Reactor 2

Reactor 3

Reactor 4

Constraint on IRR

[55% 70%]

[20% 40%]

[2.5% 5%]

[1% 3%]

ORP setting value of reactor 2

ORP setting value of reactor 1

620

560

610
ORP (mv)

ORP (mv)

550

540

600
590
580

530
570
520

12
time (h)

16

20

560

24

660

650

650

640

640

630

620

610

610
4

12
time (h)

16

16

20

24

20

24

630

620

600

12
time (h)

ORP setting value of reactor 4

660

ORP (mv)

ORP (mv)

ORP setting value of reactor 3

20

24

600

12
time (h)

16

Fig. 21. ORP setting value.

Another important phenomenon is that the uctuation in cobalt ion concentration is gradually attenuated from reactor 1 to
reactor 4. This is due to the constraints on the IRR of each reactor
and the use of robust adaptive controller. By using the controller,

the zinc dust dosage is neither kept constant nor changed drastically. As a result, the nal cobalt ion concentration is more stable
compared with the open loop control manner adopted in Sun et al.
(2014).

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

Additive dosage of reactor 1

Additive dosage of reactor 2

0.30

Dosage (kg/m3)

0.40

Dosage (kg/m3)

101

0.35

0.30

0.25

0.20

0.15

0.25

12
time (h)

16

20

0.10

24

0.10

0.08

0.08

0.06
0.04
0.02
0

12
time (h)

16

20

24

20

24

Additive dosage of reactor 4

0.10

Dosage (kg/m3)

Dosage (kg/m3)

Additive dosage of reactor 3

0.06
0.04
0.02

12
time (h)

16

20

24

12
time (h)

16

Fig. 22. Zinc dust dosage.

ORP of reactor 1

560

ORP of reactor 2

620
610
ORP (mv)

ORP (mv)

550

540

520

12
time (h)

16

20

590
580

Practical ORP
Reference ORP

530

600

Practical ORP
Reference ORP

570
560

24

660

650

650

640

640

630
620

600

12
time (h)

20

24

16

20

630
620

Practical ORP
Reference ORP

610

16

ORP of reactor 4

660

ORP (mv)

ORP (mv)

ORP of reactor 3

12
time (h)

Practical ORP
Reference ORP

610
24

600

Fig. 23. ORP tracking performance.

12
time (h)

16

20

24

102

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

Impiurity ion concentration of reactor 1

Impiurity ion concentration of reactor 2

10

10
Control result
Reference value

mg/L

mg/L

12
time (h)

16

20

24

Impiurity ion concentration of reactor 3

12
time (h)

16

20

24

10
Control result
Reference value

8
6

12
time (h)

16

20

Control result
Reference value

mg/L

mg/L

Impiurity ion concentration of reactor 4

10

Control result
Reference value

24

12
time (h)

16

20

24

Fig. 24. Cobalt ion concentration tracking performance.

5. Conclusion

References

This paper proposed an intuitive two layer control scheme for

solution purication process. The idea of gradient optimization
was derived based on two concepts (AUE and IRR) extracted from
the common characteristics of solution purication process. The
desired trajectories of efuent impurity ion concentration of each
reactor are tracked by adjusting the setting values of ORP and
controlling the additive dosage. The feasibility of the proposed
scheme is proved and illustrated through a case study. However,
there are still some drawbacks of the proposed scheme. The performance of this scheme is affected by the accuracy of the process
model and selection of the design parameters. It is suggested to
determine these design parameters at the development stage. If
the accuracy of the process model is not sufcient or the user is
lack of experience in tuning the design parameters, the performance of the scheme may deteriorate. Thus model free controller
design approach, and more precise and comprehensive process
modeling method still needs to be studied to increase the ability of
the scheme in the future.

Antonelli, R., & Astol, A. (2003). Continuous stirred tank reactors: easy to stabilise?.
Automatica, 39(10), 18171827.
Antropov, L., Beknazarov, A., 1972. Theoretical electrochemistry. Moscow: Mir.
Bckman, O., & stvold, T. (2000). Products formed during cobalt cementation on
zinc in zinc sulfate electrolytes. Hydrometallurgy, 54(2), 6578.
Brve, K., stvold, T., 1994. Norzink removal of cobalt from zinc sulphate electrolytes. In: Hydrometallurgy '94 (pp. 563577). Cambridge, Springer.
Boyanov, B., Konareva, V., & Kolev, N. (2004). Removal of cobalt and nickel from
zinc sulphate solutions using activated cementation. Journal of Mining and
Metallurgy, 40B(1), 4155.
Chen, S., Cowan, C. F., & Grant, P. M. (1991). Orthogonal least squares learning algorithm for radial basis function networks. IEEE Transactions on Neural Networks, 2(2), 302309.
Di Ciccio, M., Bottini, M., Pepe, P., & Foscolo, P. (2011). Observer-based nonlinear
control law for a continuous stirred tank reactor with recycle. Chemical Engineering Science, 66(20), 47804797.
Dib, A., & Makhlou, L. (2006). Mass transfer correlation of simultaneous removal
by cementation of nickel and cobalt from sulphate industrial solution containing copper: Part II: Onto zinc powder. Chemical Engineering Journal, 123(1),
5358.
Dib, A., & Makhlou, L. (2007). Mass transfer correlation of simultaneous removal
by cementation of nickel and cobalt from sulfate industrial solution containing
copper: Part I: Onto rotating zinc electrode disc. Chemical Engineering Journal, 130
(1), 3944.
Dreher, T. M., Nelson, A., Demopoulos, G. P., & Filippou, D. (2001). The kinetics of
cobalt removal by cementation from an industrial zinc electrolyte in the presence of Cu, Cd, Pb, Sb and Sn additives. Hydrometallurgy, 60(2), 105116.
Flett, D. S. (1992). Solution purication. Hydrometallurgy, 30(1), 327344.
Fugleberg, S., Jarvinen, A., E.Y., 1993. Recent development in solution purication at
Outokumpu Zinc Plant, Kokkola. World Zinc, 93, 241247.
Han, H., Qiao, J., & Chen, Q. (2012). Model predictive control of dissolved oxygen
concentration based on a self-organizing RBF neural network. Control Engineering Practice, 20(4), 465476.
Hoang, H., Couenne, F., Jallut, C., & Le-Gorrec, Y. (2012). Lyapunov-based control of
non isothermal continuous stirred tank reactors using irreversible thermodynamics. Journal of Process Control, 22(2), 412422.
Iliyas, S. A., Elshafei, M., Habib, M. A., & Adeniran, A. A. (2013). RBF neural network
inferential sensor for process emission monitoring. Control Engineering Practice,
21(7), 962970.
Khalil, H. K., & Grizzle, J. (2002). Nonlinear systems. Upper Saddle River: PrenticeHall.

Acknowledgments
This research was supported by Innovation-driven Plan in
Central South University (Grant no. 2015cx007), the National
Natural Science Foundation of China (Grant nos. 61174133,
61273185) and Science Fund for Creative Research Groups of the
National Natural Science Foundation of China (Grant no.
61321003). Bei Sun and Mingfang He would like to thank China
Scholarship Council for the nancial support (Nos. 201206370097
and 201306370089). These are gratefully acknowledged.

B. Sun et al. / Control Engineering Practice 44 (2015) 89103

Kim, S., Kim, K., Park, E., Song, S., & Jung, S. (2007). Estimation methods for efciency of additive in removing impurity in hydrometallurgical purication
process. Hydrometallurgy, 89, 242252.
Knapp, T. D., Budman, H. M., & Broderick, G. (2001). Adaptive control of a CSTR with
a neural network model. Journal of Process Control, 11(1), 5368.
Lew, R. W., 1994. The removal of cobalt from zinc sulphate electrolytes using the
copperantimoney process (Ph.D. thesis). University of British Columbia.
Li, Y., Gui, W., Teo, K. L., Zhu, H., & Chai, Q. (2012). Optimal control for zinc solution
purication based on interacting CSTR models. Journal of Process Control, 22(10),
18781889.
Moody, J., & Darken, C. J. (1989). Fast learning in networks of locally-tuned processing units. Neural Computation, 1(2), 281294.
Nelson, A., Wang, W., Demopoulos, G., & Houlachi, G. (2000). The removal of cobalt
from zinc electrolyte by cementation: A critical review. Mineral Processing and
Extractive Metallurgy Review, 20, 325356.
Polcaro, A., Palmas, S., & Dernini, S. (1995). Kinetics of cobalt cementation on zinc
powder. Industrial & Engineering Chemistry Research, 34, 30903095.
Rincn, A., Erazo, C., & Angulo, F. (2012). A robust adaptive controller for an
anaerobic digester with saturated input: Guarantees for the boundedness and
convergence properties. Journal of Process Control, 22(9), 17851792.
Seshagiri, S., & Khalil, H. K. (2000). Output feedback control of nonlinear systems
using RBF neural networks. IEEE Transactions on Neural Networks, 11(1), 6979.
Sun, B., Gui, W., Wang, Y., & Yang, C. (2014). Intelligent optimal setting control of a
cobalt removal process. Journal of Process Control, 24(5), 586599.
Sun, B., Gui, W., Wu, T., Wang, Y., & Yang, C. (2013). An integrated prediction model

103

of cobalt ion concentration based on oxidation reduction potential. Hydrometallurgy, 140, 102110.
Tozawa, K., Nishimura, T., Akahori, M., & Malaga, M. A. (1992). Comparison between
purication processes for zinc leach solutions with arsenic and antimony trioxides. Hydrometallurgy, 30(1), 445461.
van der Pas, V., & Dreisinger, D. (1996). A fundamental study of cobalt cementation
by zinc dust in the presence of copper and antimony additives. Hydrometallurgy, 43(1), 187205.
Wang, L., Gui, W., Teo, K., Loxton, R., & Yang, C. (2012). Optimal control problems
arising in the zinc sulphate electrolyte purication process. Journal of Global
Optimization, 54(2), 307323.
Wen, C., Zhou, J., Liu, Z., & Su, H. (2011). Robust adaptive control of uncertain
nonlinear systems in the presence of input saturation and external disturbance.
IEEE Transactions on Automatic Control, 56(7), 16721678.
Wu, F. (2001). Lmi-based robust model predictive control and its application to an
industrial CSTR problem. Journal of Process Control, 11(6), 649659.
Yang, D., Xie, G., Zeng, G., Wang, J., & Li, R. (2006). Mechanism of cobalt removal
from zinc sulfate solutions in the presence of cadmium. Hydrometallurgy, 81(1),
6266.
Yu, D., Chang, T., & Yu, D. (2007). A stable self-learning PID control for multivariable
time varying systems. Control Engineering Practice, 15(12), 15771587.
Zhou, J., Wen, C., & Zhang, Y. (2004). Adaptive backstepping control of a class of
uncertain nonlinear systems with unknown backlash-like hysteresis. IEEE
Transactions on Automatic Control, 49(10), 17511759.