Minerals Engineering 93 (2016) 65–79

Contents lists available at ScienceDirect

Minerals Engineering
journal homepage: www.elsevier.com/locate/mineng

Dynamic hybrid modeling and simulation of grinding–flotation circuits

for the development of control strategies
P. Karelovic ⇑, E. Putz, A. Cipriano
College of Engineering, Pontificia Universidad Católica de Chile, Av. Vicuña Mackenna 4860, Santiago, Chile

a r t i c l e i n f o a b s t r a c t

Article history: Process simulation is a very important tool for the design, development, analysis and optimization of
Received 5 February 2015 technical processes in the mineral industry. The ability to simulate process behavior without the cost
Revised 23 January 2016 of test runs can prevent the loss of man-hours and production, in addition to providing a platform for
Accepted 30 January 2016
the development of control tools and strategies. The usefulness of a simulator ultimately relies on how
Available online 3 May 2016
accurately the underlying mathematical model represents real behavior. In mineral processing, due to
varied complexities such as strong non-linearities, variable coupling, time varying parameters, etc.; the
Keywords:
development of accurate process models becomes an increasingly difficult task.
Hybrid systems
Process modeling
This paper describes the modeling and simulation of the main components of a concentrator plant, the
Process simulation grinding and flotation circuits. A hybrid dynamic model was favored to better represent the different
Mineral processing modes of operation and non-linearities exhibited by the plant. Industrial plant data was used to calibrate
the models used in the simulator. Simulation tests show that the performance of the simulator is qual-
itatively similar to that of a real plant, and it can be effectively used as a tool for the design and simulation
of control solutions.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction operation such as Semi-Autogenous Grinding (SAG) mills, ball

Copper minerals are too diluted in ore (0.5–2% of Cu) to be
smelted directly. Given the great amount of worthless rock, direct
heating and smelting would demand too much energy and too
much furnace capacity. To address this issue, ores which are des-
tined to pyrometallurgical processing are first concentrated, result-
ing in a product which contains about 30% of Cu (Biswas and
Davenport, 1976). The process of concentration begins with crush-
ing and grinding the mineral to achieve the liberation of the copper
contained in the ore. Undergrinding results in a coarse product that
may have a low degree of liberation, which will result in dimin-
ished recoveries. On the other hand, overgrinding wastes energy
and can increase the difficulty of processing valuable minerals.
Afterwards, copper minerals are physically separated from non-
Cu minerals by froth flotation, resulting in a Cu rich concentrate.
Extensive research has been conducted on the development of
mathematical models to represent the processes involved in a con-
centrator plant, in order to satisfy different objectives such as:
plant design and optimization, design of control systems and oper-
ator training. Grinding circuits, which are composed of several unit
⇑ Corresponding author.
E-mail addresses: pakarelo@uc.cl (P. Karelovic), eiputz@uc.cl (E. Putz),
aciprian@ing.puc.cl (A. Cipriano).
mills, vibratory screens, pebble crushers, sumps, centrifugal pumps
and hydrocyclones; have been extensively studied and numerous
models have been developed, ranging from static to dynamic mod-
els, and phenomenological to empirical models (Austin et al., 1987;
Améstica et al., 1993; Salazar et al., 2009; Orellana, 2010; Salazar
et al., 2010).
Similarly, several models have been developed to represent
flotation cells. Dynamic models based on hydraulic principles used
to evaluate control strategies for pulp levels stand out (Stenlund
and Medvedev, 2002; Yianatos et al., 2008), as well as models
based on mass balances of a set of phases (Pérez-Correa et al.,
1998; Casali et al., 2002). In dos Santos et al. (2014), a combination
of the compartment model and phenomenological models is pre-
sented. The proposed structure is successfully calibrated using
experimental data from a pilot plant. Similarly, Newcombe
(2014) presents an evaluation of current modeling methods for
flotation tested against data gathered from a flash flotation cell.
Additionally, a novel method is proposed to interpret and describe
the behavior observed, which uses the residence time from within
the settling zone of the cell to determine recovery by size at
varying heights relative to the mixing zone.
An integration of the grinding and flotation processes, and a
development of constrained predictive control strategies for this

http://dx.doi.org/10.1016/j.mineng.2016.01.021
0892-6875/Ó 2016 Elsevier Ltd. All rights reserved.
66 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79
integrated plant is presented in Muñoz and Cipriano (1999). The
proposed control strategy is a two-level scheme, with a regulatory
level based on linear Model Predictive Control (MPC), and a
optimizing level using non-linear dynamic models to maximize
economic profit. In Sosa-Blanco et al. (2000), an integrated
grinding-flotation simulator for a lead-silver-gold plant, calibrated
with real plant data, is presented. The simulator is used to perform
economic optimization of the flotation circuit through grinding
circuit tuning, increasing the economic efficiency of the plant by
10–20%. Finally, a simulator for a closed-circuit grinding process
with flash flotation is presented in Mackinnon et al. (2003). The
simulator is developed to estimate the interaction between flash
flotation and the grinding circuit and to predict the effect of vari-
ous changes to the operating conditions on circuit performance.
A comparison of the simulator results with operation data from a
real plant shows that the simulator closely represents the plant
circuit while operating under normal conditions.
In this paper, we propose a methodology to represent, by means
of hybrid models, a concentrator plant composed of a grinding
circuit followed by a rougher flotation line. Hybrid models incorpo-
rate both continuous and discrete variables (meaning that it can
only take values from an enumerable set) into the mathematical
relations that describe the dynamics of the systems considered.
For the rougher flotation circuit, a hybrid model is developed to
closely represent the different operating modes exhibited by the
cells, using discrete variables to indicate the mode that is currently
active. For the grinding circuit, on the other hand, each unit oper-
ation is represented by non-hybrid mathematical models, which
are interrelated through their input and output streams. These
interrelations are represented by discrete variables, which allows
the model to increase or decrease the number of processing lines,
simulate the blockage of pulp lines, switch between open circuit
and closed circuit operation, and so on. An integrated grinding-
flotation simulator is essential for an accurate prediction of the
copper recovery in response to changes in plant operation, as show
in Hatton et al. (2010). Additionally, the hybrid models developed
in this paper can be used to design and implement hybrid con-
trollers that take into account the different modes of operation
found in concentrator plants.
The organization of this article is as follows. Section 2 offers an
introduction to hybrid systems. The model structure, along with
the main mathematical equations, are presented in Section 3.
Model calibration is described in Section 4. Details of the imple-
mentation of the dynamic simulator are presented in Section 5.
Validation of the previous models is performed through simulation
in Section 6. Conclusions and future work are presented in
Section 7. Finally, the complete models and equations used to rep-
resent the various processes in the simulator are detailed in
Appendix A.
2. Hybrid systems
The model of a system is often associated with differential
equations (for continuous-time systems) or difference equations
(for discrete-time systems), typically derived from the physical
laws governing the dynamics of the systems under study or from
empirical observations. Additionally, in many applications the sys-
tems under consideration include parts described by logical rules,
which allow these systems to transition between different
operating modes. The interaction between continuous laws and
equations, and logic rules motivates the development of a frame-
work for modeling and controlling these type of systems
(Bemporad and Morari, 1999).
A common approach for representing such systems is by using
piecewise functions:
xðk þ 1Þ ¼ f i ðxðkÞ; uðkÞÞ
; for dðkÞ ¼ i; ð1Þ
yðkÞ ¼ gi ðxðkÞÞ
where xðkÞ is the state of the system, uðkÞ is the input vector and
yðkÞ is the output vector. dðkÞ is a discrete variable that represents
the mode of operation that is currently active at time-step k. Each
subsystems i 2 f1; 2; . . . ; sg is defined by the (possibly nonlinear)
functions f i and gi , and describes the dynamic evolution of the sys-
tem in the corresponding mode of operation i.
The methodology used in this paper combines non-linear
dynamic equations and binary variables to represent the different
operating modes of the grinding and rougher flotation circuit. This
methodology considers different dynamics for each condition and
is suitable to represent non-linear systems, which are very com-
mon in mining processes.
3. Model description
As previously stated, the plant studied in this article is consti-
tuted by a grinding circuit and a rougher flotation circuit. Both cir-
cuits are interconnected through the output stream of the grinding
circuit, which serves as the input of the rougher flotation circuit.
The objective of the simulator developed from the mathemati-
cal models presented is to be used as a tool for designing and test-
ing different control strategies. As such, the mathematical models
used focus on the transient behavior as well as the steady-state
behavior. Multiple models are used in both the grinding circuit
and the rougher flotation circuit, with the objective of providing
a qualitatively close representation of a real concentrator plant.
For simplification, the delays introduced by conveyors belts and
pipes are not included in the current version of the simulator,
although they will be considered in future work for the develop-
ment of control strategies.
3.1. Grinding circuit
Fig. 1 depicts the grinding plant modeled, which is divided into
two stages. In the primary stage, fresh ore coming from a previous
crushing stage is fed into a SAG mill, along with the necessary
water feed. Then, the mineral pulp from the SAG mill is classified
on a vibratory screen. The oversized ore is sent to a pebble crusher,
which feeds into the ball mills. The undersized ore, on the other
hand, continues to the water sump in the secondary stage. From
the sump, the mineral pulp is pumped to two hydrocyclone batter-
ies. The underflow of each hydrocyclone battery is fed into a ball
mill, which feeds back into the sump. The final product of this
grinding plant is the hydrocyclone overflow, which is fed to the
flotation cells.
Various mathematical models are used to represent the differ-
ent unit operations included in the grinding plant. The following
equations, used for the SAG mill, are derived from the models
developed in Austin et al. (1987) and Orellana (2010). This
dynamic model, based on mass balance, considers two processes
operating simultaneously inside the mill: breakage of the mineral
particles inside the grinding chamber and classification through
the discharge grate. The following equation represents the mass
balance used to model the SAG mill:
d
wsag ¼ f sag  psag  Kwsag ð2Þ
dt
where wsag is the hold-up of mineral inside the SAG mill, f sag is the
mineral feed to the mill and psag is the product of the mill. The
lower-triangular matrix K represents the breakage matrix.
Power consumption in the SAG mill is modeled with the follow-
ing equation:
 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79 67

Fig. 1. Simulated grinding plant.


2:5 0:1
~ / 1  910 /
S ¼ c1 ds ls Jð1  c2 JÞ q ð3Þ pch ¼ ðI  Cch ÞðI  Bch Cch Þ1 f ch ð8Þ
2
where f ch is the feed flow, pch is the product flow, Bch is a lower tri-
where c1 and c2 are empirical parameters, ds is the inner diameter angular matrix that represents the breakage process and Cch is a
of the mill, ls is the length of the mill and / is the fraction of the crit- diagonal matrix that represents the classification process.
ical speed of the mill. The parameter J is the fraction of occupied The sump model is based on a volumetric balance, considering
volume and q ~ is the equivalent density of the material inside the perfect mixture of its contents. Two dynamic equations are used to
mill. represent the variation of the total volume and the variation of the
To determine the final product of the mill, the following equa- mass of pulp inside the sump over time:
tion is used:
dV s d F
¼ ðAs hs Þ ¼ s
þ q þ Gs  p ð9aÞ
d   dt dt qm s
psag ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn ffi I  Csag wsag ð4Þ
i¼1 wsag;i
dms d
s qs
¼ ðAs hs qs Þ ¼ F þ qs þ Gs  p ð9bÞ
where Csag is the diagonal matrix of the classification efficiencies, dt dt
wsag;i is the i-th element of the hold-up distribution wsag and d is where V s is the volume of the pulp, ms is the mass of the pulp, As is
an empirical parameter. the area of the sump, hs is the pulp level in the sump, F is the total
For the water process in the SAG mill, a scalar dynamic equation mineral feed, qm is the mineral density, qs is the water feed in the
is used to represent the evolution of the water hold-up: s is the product volumetric flow
pulp, Gs is the external water feed, p
d and qs is the pulp density in the sump.
W W ¼ qin  qout ð5Þ The centrifugal pump static model, as presented in Orellana
dt
(2010), considers volumetric flows and pressures. The pulp flow
where qin is the feed water flow and qout is the product water flow. through the pump is proportional to the pump speed and inversely
The product flow is assumed to be proportional to the water hold- proportional to the difference in input and output pressure:
up:
! f p ¼ jp;1 v  jp;2 Pp;2  Pp;1  ð10Þ
j2
qout ¼ j1 þ Pn 4 WW ð6Þ where f p is the volumetric feed flow, v is the velocity of the pump,
i¼1 wsag;i
Pp;2 and P p;1 are the output and input pressures, respectively. jp;1
where j1 and j2 are empirical parameters. and jp;2 are empirical constants.
The vibratory screen is modeled empirically as a static effi- For the hydrocyclone battery, the equations and constant values
ciency curve fitted to experimental data, as derived in Orellana used are developed in Lynch and Rao (1975) and Plitt (1976). In
(2010), that determines the oversized and undersized ore: this model, pulp rejection is influenced by the amount of water
present in the feed pulp and by the actual operation of the hydro-
pover ¼ g f sc ð7aÞ
cyclone. Therefore, the hydrocyclone battery is represented
through a static model composed of two rejection processes in ser-
punder ¼ ðI  gÞ f sc ð7bÞ
ies. The first one only depends on the water density of the pulp:
where f sc is the mineral feed of the screen and g is a diagonal matrix ^r ¼ RH f hc ð11Þ
that contains the efficiency for each particle size.
The pebble crusher model used in this paper is presented in where ^r is the rejected mineral flow, f hc is the mineral feed flow and
Orellana (2010) and is a simplification of the crusher model pre- RH is defined as the ratio between the rejected water flow and the
sented in Whiten (1972). This model considers a single breakage feed water flow.
zone and a probability of entering or re-entering this zone. This The second rejection process is modeled as a rejection curve,
assumption leads to a static model similar to the SAG mill operat- similar to the vibratory screen:
ing in steady state:
68 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79

~r ¼ YH ðf hc  ^rÞ ð12Þ line are Cu-rich concentrate, which is sent to a regrind stage; and
Cu-poor tail, which is sent to tailing dams.
where ~r is the mineral flow rejected in the classification and YH is a
The mathematical equations used to represent the rougher
diagonal matrix that represent the classification efficiency.
flotation circuit are derived in Putz and Cipriano (2015). Each flota-
Finally, the ball mill model is based on mass balance, similar to
tion cell can operate in one of three distinct operation modes:
the one used for the SAG mill, as described in Orellana (2010):
d 1. No overflow of concentrate.
wball ¼ f ball  pball þ Bball wball ð13Þ 2. Normal operation with overflow of concentrate.
dt
3. Overflowing slurry (no froth).
where wball is the hold-up of mineral inside the ball mill, f ball is the
mineral feed to the mill and pball is the product of the mill. The
Two auxiliary logical variables (d1 and d2 ) are used to represent
lower-triangular matrix Bball represents the breakage matrix.
the current operating mode of each flotation cell. Changes between
As the ball mill doesn’t have a discharge grate, the discharge
the different modes are determined by the pulp level of each cell.
rate of the mill is independent of particle size:
Logical rules are used to determine the value of these variables:
pball ¼ aM1 wball ð14Þ i i
di1 ¼ 1 () hp 6 hc ð17aÞ
where aM1 is an empirical parameter. i i i
di2 ¼ 1 () hp þ hf > hc ð17bÞ
The water processes in the ball mill follows the same model as
the SAG mill: i i i
where hp ; hf and hc are the pulp level, froth level and height of cell i,
d respectively. Fig. 3 shows a state diagram that describes the behav-
W W;ball ¼ qin;ball  qout;ball ð15aÞ
dt ior of the flotation cell.
Mass flows are represented in the rougher flotation circuit as
qout;ball ¼ aM2 W W;ball ð15bÞ particle size distributions composed of 4 intervals, from
0.025 mm to 0.106 mm of diameter. The behavior of the rougher
where aM2 is an empirical parameter.
flotation cell is represented through a mineralogical model
The stockpile feeding the SAG mill is modeled as a hybrid
(Pérez-Correa et al., 1998) of the interaction between the pulp
system where the mineral feed is a combination of three different
phase and the froth phase. By performing a mass balance in each
particle size distributions:
phase, considering the collection flow and drainage flow between
f sag ¼ a1 f c þ a2 f m þ a3 f f ; ð16Þ them, hybrid dynamic equations are derived for the pulp mass
and froth mass. These equations take into account the operation
where f sag is the feed particle size distribution to the SAG mill, f c is a modes described previously to determine the presence or absence
coarse particle size distribution, f m is a medium particle size distri- of output flows and flows between phases:
bution and f f is a fine particle size distribution. Additionally, a1 ; a2 " #
and a3 are binary variables that represent the presence or absence Q it
_ ijk
m ijk ij ijk
p ¼ M f þ af mf  ap þ
ij
mijk
p ð18aÞ
of each particle size in the feed. By using (16) to model the mineral V ip
feed to the SAG mill, the effects of disturbances in the feed size dis-
tribution on the operation of the grinding plant can be modeled. " #
Q ic
Additionally, a continuous variable is used to represent the _ ijk
m ij ijk ij
f ¼ ap mp  af þ mijk
f ð18bÞ
hardness of the feed ore. This variable is can take random values V if
between 0 and 1 and modifies the breakage rates of the mineral where j is the mineral class number and k is the mineral size. Addi-
fed to the mill. A value close to 1 represents a harder mineral, while i i
tionally, Mijk
f is the feed mass flow; Q t and Q c are the tail and con-
a value close to 0 represents a softer mineral.
A second binary variable is introduced to represent the operat- centrate volumetric flow, respectively; V ip and V if are the pulp and
ing state of each one of the secondary grinding lines. By clearing froth phase volume, respectively; mijk ijk
p and mf are the mass of min-
this variable, a complete secondary grinding line is taken out of eral in the pulp phase and the froth phase, respectively; and aijp and
operation. This permits the simulation of pulp line blockages,
aijf are the collection rate and the drainage rate for the pulp phase
scheduled maintenance and changes in the layout of the plant.
and the froth phase, respectively.
From the perspective of control system design, manipulated
The tail grade and concentrate grade for cell i are calculated as
variables for the grinding plant are the mineral feed tonnage, water
follows:
feed, SAG mill speed, water addition to the sump and centrifugal
pump speed. Mineral feed particle size and hardness are consid- Pm Pr ijk ijk
i j¼1 k¼1 M t lcp
ered disturbances, while mill power draw, mill hold-up, sump lt ¼ lcu Pm Pr ijk
ð19aÞ
level, hydrocyclone pressure and product particle size are consid- j¼1 k¼1 M t
ered controlled variables. Additional discrete variables can be
manipulated when using a hybrid model predictive control frame- Pm Pr ijk ijk
i j¼1 k¼1 M c lcp
work, such as the number of active hydrocyclones per battery in lc ¼ lcu Pm Pr ijk
ð19bÞ
the current model and the number of operating lines of secondary j¼1 k¼1 M c
grinding.
where lcu is the copper grade of a chalcopyrite molecule (approxi-
ijk
mately 34.6%) and lcp is the chalcopyrite grade of cell i, mineral class
3.2. Rougher flotation circuit
j and mineral size k.
Fig. 2 shows the typical configuration of a rougher flotation line. The hybrid dynamic model is based on the following
The line is composed of a set of cells connected in series grouped assumptions:
into banks. Pulp is fed to the line from a distributor tank through
an on/off valve, and the banks are connected with each other 1. Constant air flow into the cell.
through controlled valves. The products of the rougher flotation 2. Each phase of the cell is perfectly mixed.
 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79
Fig. 2. Simulated rougher flotation circuit.
Fig. 3. Rougher flotation cell operating modes.
3. There is transfer between both phases, depending on the collec-
tion rate of pulp phase and the drainage rate of froth phase.
4. The mineral recovered is chalcopyrite (CuFeS2).
5. Each flotation cell has a horizontal cross section of constant
area.
Moreover, the pulp level inside the cell is derived from hydrau-
lic models (Kämpjärvi and Jämsä-Jounela, 2003); the tail flow is
derived from physical laws based on Torricelli’s principle
(Stenlund and Medvedev, 2002) and the concentrate flow is
defined as a linear function of the pulp level.
From the perspective of control system design, manipulated
variables are the output controlled valves of each cell, while feed
flow and feed grade are considered disturbances. Controlled vari-
ables are the pulp levels of each cell and the final tail grade of
the rougher flotation line.
4. Model calibration
For the grinding plant, each subsystem was calibrated indepen-
dently using data from various sources. Parameters for the SAG
mill, ball mill and hydrocyclone battery were calibrated using
actual plant data, as shown in Orellana (2010). Parameters for
the vibratory screen and centrifugal pump were calibrated using
data obtained from product catalogs. Finally, the parameters for
the pebble crusher were calibrated using values recommended in
Orellana (2010). The sump is represented through a volumetric
balance, considering perfect mixture of its contents; therefore
there are not identifiable parameters. For the rougher flotation cir-
cuit, industrial steady-state data was used to adjust the hybrid
model, as shown in Putz and Cipriano (2015). The data consist of
copper grades and mass flows (mineral and water) for the feed,
concentrate and tail.
To perform the calibration, the plant is assumed to be in steady-
state conditions. Using the data described previously and the
mathematical equations that define the dynamic system, parame-
ter identification is performed through mean square error
minimization.
The concentrator plant simulator is not intended to represent a
specific plant, given that calibration data was collected from differ-
ent sources. Additionally, this presents the difficulty of correctly
adjusting each calibrated model to obtain a proper performance
from the simulator when both the grinding circuit and the rougher
69
flotation line operate simultaneously. As such, the simulator is
designed to exhibit a behavior that is qualitatively close to reality.
Such an open structure allows the design and test of different
applications, among which stand out the development and evalu-
ation of predictive models, testing of decision support systems
and testing of control strategies.
5. Simulator implementation
The simulator was designed and implemented in MATLAB/
Simulink. Each process unit was developed as a single block, and
different units are connected through streams of mineral and
water. These streams are implemented as vectors, where the first
n components represent the particle size distributions, and the
component n þ 1 represents the water flow.
Fig. 4 shows a diagram of the implementation of the simulator
for the grinding plant. The primary grinding line is presented, as
well as a single secondary grinding line. In practice, any number
of secondary grinding lines can be connected to the output of the
primary line. In the case of the simulator used in this study, three
secondary lines were used.
Similarly, Fig. 5 shows a diagram of the implementation of the
simulator for a rougher flotation line. The line is composed of four
rougher flotation cells connected in series through the tail stream.
The configuration of the simulator permits the connection of any
number of flotation cells in a single line.
Each block, representing a single processing units, contains the
mathematical equations needed to calculate all the variables of
interest in the process. Each block has its own inputs, outputs, state
variables and disturbances; which are computed independently
from each other. Table 1 shows these variables for the grinding
plant, while Table 2 shows the corresponding variables for the
flotation plant.
6. Simulation based validation
In this section the behavior of the models developed previously
is tested and validated through simulations. The tests consist in
introducing disturbances in the inputs to the plant (feed size distri-
bution, ore hardness) as well as periodically changing the values of
the manipulated variables (mineral feed, water addition, SAG mill
speed, pump speed and the position of the input valves of the flota-
tion cells) and observing the behavior of the outputs (SAG mill
hold-up and power draw, sump level, product particle size of the
grinding circuit, pulp level in flotation cells and copper recovery).
The plant is expected to perform qualitatively similar to a real
concentrator plant.
Fig. 6 shows the evolution of the product particle size (defined
as the percentage of product retained by a 65 mesh sieve) and
specific energy consumption of the grinding plant when the feed
70 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79
Fig. 4. Grinding plant simulator implementation.
Fig. 5. Rougher flotation line simulator implementation.
size distribution varies. It can be seen that as the ore becomes
coarser, both the product particle size and the specific energy con-
sumption increase.
Fig. 7 shows the evolution of SAG mill hold-up, power and pro-
duct particle size of the concentrator plant when the hardness of
the feed varies between soft ore, medium ore and hard ore. As
ore becomes harder, power consumption and ore hold-up increase,
which is to be expected as harder ore requires more energy to be
grinded, resulting in a coarser product. Additionally, Fig. 8 shows
that ore hardness has a comparatively lesser effect on the concen-
trate grade, tail grade and copper recovery in the rougher flotation
circuit, as opposed to the effect observed in the grinding plant.
Fig. 9 shows the effect of changing the mineral feed tonnage on
the power consumption and hold-up of the SAG mill, as well as the
product particle size of the grinding circuit. Additionally, Fig. 10
shows the effect of this disturbance on the pulp levels and mode
of operations of the first three cells of the rougher flotation line.
As expected, an increase in mineral feed produces an increases in
the power consumption and hold-up of the SAG mill, as well as
increasing the product particle size and the pulp level of the flota-
tion cells, causing the first cell to enter the third mode of operation
(presence of pulp in concentrate overflow). The opposite effect is
seen when the mineral feed is decreased.
In Fig. 11, the speed of the SAG mill (represented as a fraction of
the critical speed) is changed and the effect of the grinding plant
values is shown. It can be seen that a change in the SAG mill speed
produces a similar change on the power consumption in the same
direction, and in the opposite direction on the SAG mill hold-up
and product particle size. While the effect of the SAG mill speed
is noticeable in the grinding plant, this effect is far less noticeable
in the rougher flotation line.
Fig. 12 shows several power consumption versus hold-up
curves of the SAG mill, for different conditions of operation,
defined by different values of water feed to the mill. As can be seen,
the curve resembles and upside-down parabola, which means that
the model is able to represent closely the normal operation and
overload condition of the SAG mill.
To better visualize the effect that disturbances in the grinding
circuit have over the flotation process, Fig. 13 shows the evolution
of the pulp level of the first three cells of the rougher flotation line,
as well as the operating mode of each cell, when the water added
to the sump in secondary grinding is varied. It can be seen that,
when the water addition increases, the pulp level of the flotation
cells also increases, causing the first cell of the line to enter the
third mode of operation (presence of pulp in the overflow). Simi-
larly, when the water addition decreases, pulp levels also decrease,
causing all of the cells to enter the first mode of operation (absence
of overflow). Finally, when the water addition returns to its nomi-
nal value, the flotation cells return to their normal operating mode.
Fig. 14 shows all the relevant variables of the first cell of the
rougher flotation line. A disturbance is added to the feed tonnage
of the cell to display the evolution of the tail and concentrate flows,
pulp and froth level, and the mass of pulp and froth inside the cell
for each particle size interval considered. Additionally, it can be
seen that the hybrid model switches between its three operating
modes, as described in Fig. 3.
7. Conclusions
In this paper the development of a dynamic hybrid model for a
concentrator plant was presented. The plant, which is composed of
a grinding circuit and a rougher flotation circuit, was modeled
through a combination of static, dynamic, phenomenological,
empirical and hybrid models. Subsequent simulations of the plant
show a high degree of representation of the studied processes.
The development of an integrated simulator for a grinding-
flotation plant allows the design and test of different applications,
among which stand out the development and evaluation of predic-
tive models, testing of decision support systems and testing of con-
trol strategies.
Future work will focus on the improvement of the simulation
models, such as modeling the effect of the air flow in the flotation
cell. Additionally, future work will be made towards the design,
 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79 71

Table 1 Appendix A
Inputs, outputs, state variables and disturbances for the grinding plant.

Unit Inputs Outputs State Disturbances The mathematical models used to represent to concentrator
variables plant presented in this study are detailed in the following section.
SAG mill Mineral feed Mineral Mineral Ore The models related to the grinding plant are adapted from Orellana
product hold-up hardness (2010), while the model for the flotation cell is taken from Putz and
Water feed Water Water Size
Cipriano (2015). Additionally, the values used for the parameters
product hold-up distribution
Speed ratio Power draw are presented after its corresponding model.
The particle sizes used for the grinding and flotation circuits are
Vibratory Mineral feed Mineral Size
screen oversize distribution the following:
Water feed Mineral
undersize
xgrinding ¼ ½203:2 152:4 101:6 53:3 19:1 12:7 3:35
Water 1:7 0:85 0:425 0:212 0:106 0:053 0:025 0 ½mm
oversize
Water
ð20aÞ
undersize
Pebble crusher Mineral feed Mineral Size xflotation ¼ ½0:106 0:053 0:025 0 ½mm ð20bÞ
product distribution
Water feed Water
product A.1. SAG mill model
Water sump Mineral feed Mineral Pulp
product height The SAG mill is represented by a mass balance model which
Water feed Water Pulp takes into account the effects of grinding and classification. Min-
product density
eral flows are represented through vectors of n elements associated
External Mineral
water feed hold-up to different particle size ranges.
Centrifugal Input Output d
pump pressure pressure wsag ¼ f sag  psag  Kwsag ð21Þ
dt
Pump speed Volumetric
pulp flow where wsag is the hold-up of mineral inside the SAG mill, f sag is the
Hydrocyclone Mineral feed Mineral Size mineral feed to the mill and psag is the product of the mill.
battery product distribution The lower-triangular matrix K represents the breakage matrix
Water feed Water
product
and is defined as:
Rejected 8
> 0 if i < j
mineral <
A
Rejected ki;j ¼ ki if i ¼ j ð22Þ
water >
: A A
Feed pressure ki  ki1 if i > j
Ball mill Mineral feed Mineral Mineral Size A
where ki is the rate with which particles bigger than size i are bro-
product hold-up distribution
A
Water feed Water Water ken to sizes smaller or equal than i. The rates ki are dependent of
product hold-up the conditions of operation of the mill, including mineral hold-up,
power draw and intrinsic characteristics of the ore.
The product of the mineral hold-up a the grinding rates are pro-
Table 2 portional to the power draw of the mill:
Inputs, outputs, state variables and disturbances for the flotation plant.
A E
Unit Inputs Outputs State Disturbances Wki ¼ Ski ð23Þ
variables Pn
where W is the total mineral hold-up (i.e. W ¼ i¼1 wsag;i ), S is the
Flotation Mineral Concentrate mass Pulp mass Feed grade E
cell feed flow power draw of the mill and the rates ki only depend on the charac-
Water Concentrate Froth mass Size teristics of the mineral, and can be determined experimentally.
feed volumetric flow distribution To include the effect of the mineral hardness, the effective
Tail mass flow Pulp
grinding rates are calculated as follows:
height
Tail volumetric flow E soft hard
ki ¼ ð1  lÞki þ lk i ð24Þ

development and implementation of Hybrid Model Predictive Con- where l is a variable between 0 and 1 that represents the mineral
soft hard
trol (HMPC) strategies. These schemes take advantage of the hybrid hardness, and ki and ki are the grinding rates of previously
models presented in this paper, enabling the representation of sys- identified soft and hard minerals, respectively. In that way, when
tems incorporating logical variables, rules and continuous dynam- l varies from 0 to 1, the mineral fed to the mill will vary from a soft
ics. The controllers designed using this framework will be tested to a hard mineral.
and compared against conventional MPC strategies to evaluate The power draw of the SAG mill, as derived in Austin et al.
its performance. (1987), is calculated with the following equation:


0:1
Acknowledgments
2:5
~ / 1  910 /
S ¼ c1 ds ls Jð1  c2 JÞ q ð25Þ
2
This study was funded by the Fondecyt project No. 1120047, where c1 and c2 are empirical parameters, ds is the inner diameter
‘‘Distributed Hybrid Model Predictive Control for Mineral of the mill, ls is the length of the mill and / is the fraction of the crit-
Processing”. ical speed of the mill. The parameter J is the fraction of occupied
72 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79

Fig. 6. Simulation results for alternating feed size distribution.

Fig. 7. Grinding plant results for change in ore hardness.

volume and q ~ is the equivalent density of the material inside the

mill. Both parameters are determined as follows: Fig. 8. Rougher flotation circuit results for change in ore hardness.
Pn
1 i¼1 wsag;i
of the mill occupied by balls, W W is the water hold-up inside the
J¼ þ JB ð26Þ
1 qm V mill and W B is the total load of balls.
~ of the
To determine the final product of the mill, the output p
Pn grinding chamber (before classification) is assumed to have the
i¼1 wsag;i þ WW þ WB
q~ ¼ ð27Þ same normalized size distribution as the hold-up:
JV
~
p wsag
where  is the porosity of the mineral, qm is the density of the min- Pn ¼ Pn ð28Þ
~
i¼1 pi i¼1 wsag;i
eral, V is the inner volume of the mill, JB is the fraction of the volume
 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79 73

where qin is the feed water flow and qout is the product water flow.
The product flow is assumed to be proportional to the water hold-
up:

qout ¼ a W W
!
j2 ð32Þ
¼ j1 þ Pn 4 W W
i¼1 wsag;i

where j1 and j2 are empirical parameters.

Table 3 lists the values for the parameters used in the SAG mill
model.

A.2. Vibratory screen model

The vibratory screen model is based on a static efficiency equa-

tion, where the oversized and undersized mineral flows are repre-
sented as follows:

pover ¼ g f sc ð33aÞ
Fig. 9. Grinding plant results for change in mineral feed tonnage.
punder ¼ ðI  gÞ f sc ð33bÞ

where f sc is the mineral feed of the screen and g is a diagonal matrix

Additionally, the following relation is determined empirically:
that contains the efficiency for each particle size:
X
n qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xn h  a2 i
~i ¼ d
p w ð29Þ a1
xi
i¼1 sag;i gi ¼ 1  e d50
ð34Þ
i¼1

where d is an empirical parameter. The parameters a1 and a2 are identified from operation data, xi
Finally, the final product particle size after classification is is a representative size of interval i and d50 is the diameter for
determined as follows: which 50% of the feed particles are rejected. This last parameter
  is determined as follows:
psag ¼ I  Csag p ~
Pn !0:148
d   F ua
¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi I  Csag wsag ð30Þ i¼1 f sc;i
Pn d50 ¼ ð35Þ
1:6 a3
i¼1 wsag;i

where Csag is the diagonal matrix of the classification efficiencies. where F ua is the fraction of mineral smaller than the screen aperture
For the water process, a scalar dynamic equation is used to rep- and a3 is an empirical parameter.
resent the evolution of the water hold-up: For the water process, it is assumed that only 10% of the feed
water flow is rejected along with the oversized material.
d Table 4 lists the values for the parameters used in the vibratory
W W ¼ qin  qout ð31Þ
dt screen model.

Fig. 10. Pulp levels and modes of operation of flotation cells for change in mineral feed tonnage.
74 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79

Fig. 11. Grinding plant results for change in SAG mill speed.

The auxiliary variables d1 and d2 are calculated with the follow-

ing equations:
d1 ¼ b0 þ b1 css þ b2 F ch þ b3 f 80 ð38aÞ
d2 ¼ b4 þ b5 css þ b6 F ch þ b7 f 80 ð38bÞ
where b0 through b7 are empirical parameters, css is the closed-side
setting of the crusher, F ch is the total mineral flow and f 80 is the size
at which 80% of the feed is retained.
The elements of the N  N breakage matrix Bch are defined as:
8
> 1 if i ¼ N ^ j P N  1
>
>
< wði  jÞ  wði  j þ 1Þ if j < i < N ^ j < N  1
bi;j ¼ ð39Þ
> 1  ðwð1Þ  wðN  jÞÞ
> if i ¼ N ^ j < N  1
>
:
0 otherwise
The function wðiÞ is defined as:
"
b8 #d3
xi
wðiÞ ¼ 1  1  ð40Þ
x1

Fig. 12. SAG mill power versus load curve.

d3 ¼ b9 þ b10 css þ b11 F ch þ b12 f 80 ð41Þ
where b8 through b12 are empirical parameters.
A.3. Crusher model
For the water process, it is assumed that the feed water flow is
equal to the product water flow.
For the crusher model, a static mass balance equation is used,
Table 5 lists the values for the parameters used in the crusher
which takes into account the processes of breakage and classifica-
model.
tion. The equations and constant values used in this model are
developed in Orellana (2010).
A.4. Water sump model
The static mass balance is characterized as:

pch ¼ ðI  Cch ÞðI  Bch Cch Þ1 f ch ð36Þ The water sump is represented trough a volumetric balance and
a mass balance:
where f ch is the feed flow, pch is the product flow, Bch is a lower tri-
angular matrix that represents the breakage process and Cch is a dV s d F
¼ ð As h s Þ ¼ s
þ q þ Gs  p ð42aÞ
diagonal matrix that represents the classification process. dt dt qm s
The classification rates ci of matrix Cch are determined as
follows: dms d
s qs
¼ ðAs hs qs Þ ¼ F þ qs þ Gs  p ð42bÞ
8 dt dt
>
<1 if xi > d2
d x where V s is the volume of the pulp, ms is the mass of the pulp, As is
ci ¼ 1  d22d1i if d1 6 xi 6 d2 ð37Þ
>
: the area of the sump, hs is the pulp level in the sump, F is the total
0 if xi < d1 mineral feed, qm is the mineral density, qs is the water feed in the
 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79 75

Fig. 13. Pulp levels and modes of operation of flotation cells.

Fig. 14. Hybrid behavior of first flotation cell.

s is the product volumetric flow

pulp, Gs is the external water feed, p qm 1  qs
cp ¼ ð44Þ
and qs is the pulp density in the sump. 1  qm qs
Eqs. (42a) and (42b) can be combined to determine the equa-
The mineral and water mass flows in the product are deter-
tion that models the evolution of the pulp density in the sump:
mined with the following equations, respectively:


 s qs c p
P¼p ð45Þ
dqs 1 q
¼ 1  s F þ ð1  qs Þðqs þ Gs Þ ð43Þ
dt As h s qm  
 s qs 1  c p
qout;s ¼ p ð46Þ
The percent solids of the pulp is determined with:
76 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79

Table 3 The input pressure is equal to the pressure exerted by the pulp
Parameters for the SAG mill model. in the sump:
h i
c1 ds 9.75 ½m
8.81 kW
m0:5 ton Pp;1 ¼ qs g hs ð50Þ
c2 1.1 ls 5.18 ½m
h i where g is the gravitational acceleration.
 0.4 qm 2.7 ton
m3
 The output pressure is the sum of the pressure exerted and the
V 359 m3 JB 0.08
h 0:5 i feed of the hydrocyclone battery, the pressure exerted by the col-
d 235.53 tonh WB 204 ½ton
h i h 4i
umn of pulp in the pipes that connect the pump and the hydrocy-
j1 3.77 h
1 j2 0.71 ton clone battery, the pressure drop due to friction and the pressure
h

ki
soft [0.3211, 0.3211, 0.3211, 0.3211, 0.4009, 0.4155, 0.3494, drop due to singularities. Friction drops are proportional to the
0.2818, 0.2155, 0.1602, 0.1171, 0.0854, 0.0618, 0.0432, length of the pipe and the square of the pulp velocity, and inversely
 ton 
0.0003] kWh proportional to the diameter of the pipe. On the other hand, singu-
ki
hard [0.1729, 0.1729, 0.1729, 0.1729, 0.2159, 0.2237, 0.1882,
larities drops are proportional to the square of the pulp velocity.
0.1518, 0.1161, 0.0862, 0.0631, 0.0460, 0.0332, 0.0232,
 ton 
0.0001] kWh lp v 2p v 2p
ci [1.0000, 1.0000, 1.0000, 1.0000, 0.9000, 0.4000, 0.3500, Pp;2 ¼ PH þ qs g hH þ jF þ jS ð51Þ
0.3000, 0.2500, 0.2000, 0.1500, 0.0100, 0.0023, 0.0009, 0]
dp 2 g 2g
where PH is the pressure exerted and the feed of the hydrocyclone
battery, hH is the height of the hydrocyclone battery relative to
Table 4 the pump, jF is the friction loss factor, jS is the singularities loss
Parameters for the vibratory screen model.
factor, lp is the length of the pipe, dp is the diameter of the pipe
a1 0.638 and v p is the velocity of the pulp flow.
a2 8.98 The velocity of the pulp flow can be determined with:
h i
a3 0.00013 ton
mm6:757
f p
vp ¼ ð52Þ
Ap
Table 5 where Ap is the area of the pipe.
Parameters for the crusher model.
Finally, the output pressure can be expressed as:
b0 0.1 ½mm b7 0 !2 !2
b1 1
 
b8 0.9 lp 1 f p 1 f p
b2 0 mm b9 0.03 Pp;2 ¼ PH þ qs g hH þ jF þ jS
ton
  dp 2 g Ap 2 g Ap
b3 0 b10 1.2 mm1
 
!
0.1 ½mm 1 lp jF 1 jS ð53Þ
¼ PH þ qs g hH þ f 2p
b4 b11 0.05 ton
  þ
b5 2 b12 0.01 mm1 A2p dp 2 g A2p 2 g
mm
b6 0
¼ PH þ qs g hH þ cpf 2p
ton

To simplify the model, the quadratic term of the volumetric feed

Finally, the mineral hold-up of the sump is modeled as: flow is disregarded by making the parameter cp equal to 0.
dws Table 7 lists the values for the parameters used in the centrifu-
¼ f s  ps gal pump model.
dt
s q cp ð47Þ
p
¼ f s  Pn s ws A.6. Hydrocyclone battery model
i¼1 ws;i

Table 6 lists the values for the parameters used in the water
sump model.
A.5. Centrifugal pump model
In the centrifugal pump, the volumetric pulp flow is propor-
tional to the velocity of the pump and to the pressure difference
between input and output:
f p ¼ jp;1 v  jp;2 Pp;2  Pp;1 
where f p is the volumetric feed flow, v is the velocity of the pump,
Pp;2 and P p;1 are the output and input pressures, respectively, jp;1 is
an empirical constant and jp;2 is defined as:
The hydrocyclone battery is represented with a mass balance
equations involving two distinct processes. The first process repre-
sents the pulp rejected from the battery due to the amount of
water in the pulp, while the second process represents the classifi-
cation performed by the hydrocyclone. The equations and constant
values used in this model are developed in Lynch and Rao (1975)
and Plitt (1976).
The first processed os characterized by:
ð48Þ ^r ¼ RH f hc ð54Þ
where ^r is the rejected mineral flow, f hc is the mineral feed flow and
RH is defined as the ratio between the rejected water flow and the
feed water flow:

476:0287
jp;2 ¼ ð49Þ
qs g

Table 7
Parameters for the centrifugal pump model.
Table 6 h i
jp;1 20.99 m3
Parameters for the water sump model. h RMP

  h i hH 7 ½m
As 20 m2 qm 2.7 ton
m3 cp 0
 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79 77

qr The lower-triangular matrix Bball represents the breakage

RH ¼
qin;hc matrix and is defined as:
qin;hc  qout;hc ð55Þ 8
¼ > 0 if i < j
>
qin;hc < h
> r
xi L
i
bi;j ¼ sj 1  xj 1 if i ¼ j ð64Þ
The output water flow is related to the feed water flow and the >
> h     i
>
: s 1  xi
r L r L
percent solids of the pulp by the following empirical equation: j xj
 1  xi1
xj
if i > j
 
qout;hc ¼ 1:1 qin;hc  3:96 cp þ jH1 nH ð56Þ where r L and sj are empirical parameters.
where jH is an empirical parameter and nH is the number of hydro- There is no classification process in the ball mill, so it is
cyclones in the battery. assumed that the particle size distribution of the product mineral
The process of classification is characterized by the following flow is the same as the mineral hold-up:
equation: pball ¼ aM1 wball ð65Þ
~r ¼ YH ~f where aM1 is an empirical parameter.
¼ YH ðf hc  ^rÞ ð57Þ The water processes in the ball mill follows the same model as
the SAG mill:
¼ YH ð1  RH Þf hc
d
where ~r is the mineral flow rejected in the classification and YH is a W W;ball ¼ qin;ball  qout;ball ð66aÞ
dt
diagonal matrix that represent the classification efficiency.
The elements yi of matrix YH are determined as follows: qout;ball ¼ aM2 W W;ball ð66bÞ
 aH
0:693
xi
where aM2 is an empirical parameter.
yi ¼ 1  e ð58Þ
x50
Table 9 lists the values for the parameters used in the ball mill
where x50 is the diameter for which 50% of the feed particles are model.
rejected and aH is the Plitt parameter.
The parameter x50 is determined with the following empirical A.8. Flotation cell model
equation:
The flotation cell model considers three operating modes deter-
dvf dsp PH qout;hc mined by the level of pulp and froth. The operating modes identi-
ln x50 ¼  þ  þ jH2 ð59Þ
6:6 8:9 10:7 47:2 nH fied are:
where dvf is the diameter of the vortex finder, dsp is the diameter of
1. No overflow of concentrate.
the spigot and jH2 is an empirical parameter.
2. Normal operation with overflow of concentrate.
The pressure of the pulp at the feed of the hydrocyclone battery
3. Presence of pulp in overflow of concentrate.
is determined with the following empirical equation:
!2
f p To represent the operating modes, two auxiliary binary vari-
PH ¼ 13:725  0:125 ð60Þ ables were defined:
1:48 dvf nH 1  cp jH3 h i h i
i i
di1 ¼ 1 () hp 6 hc ð67aÞ
where jH1 is and empirical parameter. h i h i
The final product mineral flow and rejected mineral flow are i i i
di2 ¼ 1 () hp þ hf > hc ð67bÞ
determined by:
i i i
where hp ; hf , and hc are the pulp level, froth level, and cell height
phc ¼ ~f  ~r ¼ ~f  YH ~f ¼ ðI  YH Þ~f ¼ ðI  YH Þð1  RH Þf hc ð61Þ
for cell i; respectively. Therefore, the first operating mode is
h i
rhc ¼ ^r þ ~r ¼ f hc  phc ð62Þ represented by di1 di2 ¼ ½1 0, the second operating mode is
h i
Table 8 lists the values for the parameters used in the hydrocy- represented by di1 di2 ¼ ½1 1, and the third operating mode is
clone battery model. h i
represented by di1 di2 ¼ ½0 1.
A.7. Ball mill model Mass balance equations were used to model the mass of the
pulp phase and the froth phase inside the cell, incorporating the
The ball mill is characterized by a mass balance model similar to feed flow, concentrate flow, and tail flow:
the one used in the SAG mill model: " #
Q it
_ ijk
m p ¼ M ijk
f þa ij ijk
f mf  a þ ij
p mijk
p ð68aÞ
d V ip
wball ¼ f ball  pball þ Bball wball ð63Þ
dt
where wball is the hold-up of mineral inside the SAG mill, f ball is the
mineral feed to the mill and pball is the product of the mill.
Table 9
Parameters for the ball mill model.
h i
Table 8 aM1 1
12.01 h
Parameters for the hydrocyclone battery model. h i
ton
aM2 12.01 h
1
jH1 122.55 h
aH 0.38
rL 7.70
jH2 2.56 dvf 26.42 ½cm
sj [970.40, 699.31, 513.22, 321.11, 152.29, 80.28, 36.91, 15.04, 8.16, 4.39,
jH3 2.09 dsp 15.24 ½cm h i
1
nH 6 2.36, 1.27, 0.69, 0.36, 0.14] h
78 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79

" #
Q ic Table 10
_ ijk
m ij ijk ij
f ¼ ap mp  af þ mijk
f ð68bÞ Parameters for the 4-cell rougher flotation line model.
V if  
Ai 25:48 m2 ; i ¼ 1 ip 0:08; 8i
 2
where j is the mineral class number and k is the mineral size. 51 m ; i ¼ 2; 3; 4 if 0:50; 8i
Additionally, Mijk i i 1 0.35 ½m 2 0.37 ½m
f is the feed mass flow; Q t and Q c are the tail and
hf hf
3
hf 0.39 ½m 4
hf 0.41 ½m
concentrate volumetric flow, respectively; V ip and V if are the pulp
Dh
i 1½m; i ¼ 1; 2; 3 i
hc 4:62 ½m; 8i
and froth phase volume, respectively; mijk ijk
p and mf are the mass of 0½m; i ¼ 4 lcu 34.61 [%]
h 2i h 2i
mineral in the pulp phase and the froth phase, respectively; and a1c 3024.02 mh a2c 6402.91 mh
aijp and aijf are the collection rate and the drainage rate for the pulp h 2i h 2i
a3c 230.08 mh a4c 3710.79 mh
phase and the froth phase, respectively. 1
h 2:5 i h 2:5 i
at 2667.26 mh a2t 3323.72 mh
The total volume for each cell is determined as follows: h 2:5 i h 2:5 i
 
3
at 1905.89 mh a4t 555.14 mh
1  ip h i h i
i
V ip ¼ Ai h p ð69aÞ a11
p 89.48 h
1 a12
p 0.29 h
1

21
h i h i
a 1 a22 1
i  p 89.48 h p 2.34 h
V if ¼ Ai h f 1  i
f ð69bÞ a31
p 89.48 h
h
1
i
a32
p
h
1.56 h
1
i

h i h i
ip if
41
where and are the air hold-up of the pulp phase and froth ap 89.49 h
1 a42
p 1.80 h
1

11
h i h i
phase, respectively, and Ai is the area of the cell. af 0.00 h
1 a12
f 157.03 h
1

h i h i
To determine the level of the pulp phase, a volumetric balance a21
f 0.00 h
1 a22
f 157.02 h
1

is performed (considering that the froth level is negligible com- h i h i

a31 1 a32 1
f 0.00 h f 157.05 h
pared to the pulp level). h i h i
a41
f 0.00 h
1 a42
f 157.05 h
1

1 
i
dhp i1k
lcp ½0:4663 0:3026 0:7114 0:3842; 8i
¼ i Q if  Q it  Q ic di1 ð70Þ
dt A i2k
lcp ½0:0059 0:0194 0:0094 0:0039; 8i

where Q if is the feed volumetric flow

The tail volumetric flow was derived from Torricelli’s law, while
the concentrate volumetric flow was represented with an empiri- Finally, the final recovery of the rougher stage is calculated as
cal function of pulp level, froth level, and cell height: follows:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

Q it ¼ ait v ic
i
hp  hp þ Dh
iþ1 i
ð71aÞ lt
R ¼ 100 1  ð74Þ
lf
    
i i i
Q ic ¼ aic hp þ hf  hc di1 di2 þ Q if  Q it 1  di1 ð71bÞ where lt and tail grade of the last cell of the rougher line and lf is the
feed grade.
where ait and aic are tuning constants, v ic 2 ½0; 1 is the position of Table 10 lists the values for the parameters used in the 4-cell
i rougher flotation line model.
the output valve of cell i; Dh is the level difference between con-
secutive cells and Q if is the feed volumetric flow.
Mass flows are determined by combining mineral masses and References
volumetric flows:
Améstica, R., González, G., Barría, J., Magne, L., Menacho, J., Castro, O., 1993. A SAG
Q it mill circuit dynamic simulator based on a simplified mechanistic model.
M ijk
t ¼ mijk
p ð72aÞ Proceedings of XVIII International Mineral Processing Congress, vol. 1, pp. 117–
V ip 129.
Austin, L., Menacho, J., Pearcy, F., 1987. A general model for semi-autogenous and
Q ic
M ijk
c ¼ mijk
f ð72bÞ autogenous milling. Proceedings of the 20th International Symposium on the
V if Application of Mathematics and Computers in the Mineral Industries, vol. 2, pp.
107–126.
Bemporad, A., Morari, M., 1999. Control of systems integrating logic, dynamics, and
where Mijkt and M ijk
c are the tail mass flow and concentrate mass constraints. Automatica 35 (3), 407–427.
flow, respectively. Biswas, A., Davenport, W., 1976. Extractive Metallurgy of Copper. Pergamon
The tail grade and concentrate grade for cell i are calculated as International Library, Pergamon Press.
Casali, A., Gonzalez, G., Agusto, H., Vallebuona, G., 2002. Dynamic simulator of a
follows: rougher flotation circuit for a copper sulphide ore. Miner. Eng. 15 (4), 253–262.
Pm Pr ijk ijk
dos Santos, N.A., Savassi, O., Peres, A.E.C., Martins, A.H., 2014. Modelling flotation
i j¼1 k¼1 M t lcp with a flexible approach – integrating different models to the compartment
lt ¼ lcu Pm Pr ijk
ð73aÞ model. Miner. Eng. 66–68, 68–76.
j¼1 k¼1 M t Hatton, D., Connolly, J., Valencia, L., 2010. An investigation into the effect of grind on
flotation performance by modelling a large copper concentrator. In:
Pm Pr ijk ijk
Proceedings of the 7th International Mineral Processing Seminar, GECAMIN,
i j¼1 k¼1 M c lcp pp. 301–308.
lc ¼ lcu Pm Pr ijk
ð73bÞ Kämpjärvi, P., Jämsä-Jounela, S., 2003. Level control strategies for flotation cells.
j¼1 k¼1 M c Miner. Eng. 16 (11), 1061–1068.
Lynch, A., Rao, T., 1975. Modelling and scale-up of hydrocyclone classifiers. In:
where lcu is the copper grade of a chalcopyrite molecule (approxi- Proceedings of the 11th International Mineral Processing Congress, pp. 245–
ijk 269.
mately 34.6%) and lcp is the chalcopyrite grade of cell i, mineral class Mackinnon, S., Yan, D., Dunne, R., 2003. The interaction of flash flotation with closed
j and mineral size k. The equation used to calculate the tail grade is circuit grinding. Miner. Eng. 16 (11), 1149–1160.
valid in steady state. However, according to Pérez-Correa et al. Muñoz, C., Cipriano, A., 1999. An integrated system for supervision and economic
optimal control of mineral processing plants. Miner. Eng. 12 (6), 627–643.
(1998), this equation can also be used to represent its behavior in Newcombe, B., 2014. A phenomenological model for an industrial flash flotation
dynamic state. cell. Miner. Eng. 64, 51–62.
 P. Karelovic et al. / Minerals Engineering 93 (2016) 65–79
Orellana, R., 2010. Modelo, control y simulador de planta de molienda
semiautógena y molienda secundaria (Modeling, control and simulation of
semi-autogenous and secondary grinding plant) Thesis for degree of Electrical
Engineer. Universidad de Chile.
Pérez-Correa, R., González, G., Casali, A., Cipriano, A., Barrera, R., Zavala, E., 1998.
Dynamic modelling and advanced multivariable control of conventional
flotation circuits. Miner. Eng. 11 (4), 333–346.
Plitt, L., 1976. A mathematical model of the hydrocyclone classifier. CIM Bull. 69
(776), 114–123.
Putz, E., Cipriano, A., 2015. Hybrid model predictive control for flotation plants.
Miner. Eng. 70, 26–35.
Salazar, J., Magne, L., Acuna, G., Cubillos, F., 2009. Dynamic modelling and
simulation of semi-autogenous mills. Miner. Eng. 22 (1), 70–77.
79
Salazar, J.L., Cubillos, F., Valdés-González, H., 2010. Advanced Simulation for Semi-
Autogenous Mill Systems: A Simplified Models Approach. InTech, pp. 145–156
(Chapter 8).
Sosa-Blanco, C., Hodouin, D., Bazin, C., Lara-Valenzuela, C., Salazar, J., 2000.
Economic optimisation of a flotation plant through grinding circuit tuning.
Miner. Eng. 13 (10–11), 999–1018.
Stenlund, B., Medvedev, A., 2002. Level control of cascade coupled flotation tanks.
Control Eng. Pract. 10 (4), 443–448.
Whiten, W., 1972. The simulation of crushing plants with models developed using
multiple spline regression. J. S. Afr. Inst. Min. Metall., 257–264
Yianatos, J.B., Larenas, J.M., Moys, M.H., Diaz, F.J., 2008. Short time mixing response
in a big flotation cell. Int. J. Miner. Process. 89 (1–4), 1–8.