RNMPC Applied To A Nonlinear Model of A

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

RNMPC Applied to a Nonlinear Model of a

ROM Ore Milling Circuit


L.C. Coetzee ∗ I.K. Craig ∗∗ E.C. Kerrigan ∗∗∗

MINTEK, Private Bag X 3015, Randburg, 2125, South Africa.
(Email: loutjiec@mintek.co.za, Tel: +27 11 709 4174)
∗∗
Department of Electrical, Electronic and Computer Engineering,
University of Pretoria, Pretoria, 0002, South Africa.
(Email: icraig@postino.up.ac.za)
∗∗∗
Department of Aeronautics and Department of Electrical and
Electronic Engineering, Imperial College London, Exhibition Road,
London SW7 2AZ, United Kingdom.
(Email: e.kerrigan@imperial.ac.uk)

Abstract: This paper presents simulation results of applying robust nonlinear model predictive
control (RNMPC) to a nonlinear model of a run-of-mine (ROM) ore milling circuit. The model
consists of nonlinear modules for the individual process units of the milling circuit, which allow
arbitrary milling circuit configurations to be modelled. This study aims to cast a complex
problem of a run-of-mine ore milling circuit into a robust nonlinear model predictive control
framework without losing the flexibility of the modularised nonlinear model and implement the
robust nonlinear model predictive controller using open-source software modules.

Keywords: Robust nonlinear model predictive control, run-of-mine ore milling circuit,
RNMPC, ROM

1. INTRODUCTION are present. This is accomplished by incorporating an


explicit description of the uncertainties and disturbances
into the prediction model of the controller. The controller
The ROM ore milling circuit plays an important role in then usually calculates the optimal control sequence or
processing mined ore. This forms part of the comminution feedback laws that optimise a given objective function for
process of liberating the desired metals or minerals from the worst-case realisation of the system as defined by the
the ore. At the heart of the milling circuit is usually a ball, worst-case objective function.
rod, semi-autogenous or autogenous mill.
There are a number of ways to approach RNMPC.
The milling process is complicated by significant input Some authors, such as Mhaskar and Kennedy (2008) and
and plant uncertainties, because the feed ore forms part Mhaskar et al. (2006, 2005) use Lyapunov stability theory
of the grinding medium and the variation in feed ore to formulate the RNMPC that also defines the set of initial
contributes to the input uncertainty. The feed ore distri- conditions that will satisfy the stability conditions. This
bution in terms of hardness and size affects the breakage ensures that the controller is feasible for the initial state
kinetics and contributes to the plant uncertainties. Proper and as a result remains feasible for all future states under
design of the milling circuit alone cannot eliminate the control.
disturbances. Feedback control systems play an important
part in reducing the effects of disturbances and increasing Bravo et al. (2006) and Limon et al. (2005) use reachable
efficiency (Craig and MacLeod, 1995). sets to describe the predicted state evolution subject
to uncertainty and disturbances inside their controller
In milling circuits it is difficult to control important vari- formulations.
ables such as the product particle size, because indepen-
dent control of the amount, size and hardness of the Lazar et al. (2008); Limon et al. (2006) and Magni et al.
grinding medium in the mill is not possible. This causes (2003) developed a closed-loop min-max RNMPC. Being
significant uncontrollable disturbances and uncertain plant closed-loop, the effect of feedback is taken into considera-
dynamics (Craig and MacLeod, 1995). tion when predicting the future state trajectories subject
to uncertainty and disturbances, which reduces the spread
Applying robust nonlinear control, such as RNMPC, fur- in predicted future state trajectories compared to open-
ther simplifies the process by eliminating the need for loop formulations. This reduced spread in trajectories,
different simulation and control models. The nonlinear increases the feasible region of the controller and reduces
nature and the uncertainties are explicitly incorporated in the conservatism of the controller.
the controller to provide the best performance and stability
trade-off. Robust control aims to guarantee stability and The RNMPC controller presented in this paper is based
improved performance when uncertainty and disturbances on an open-loop min-max scheme, which might cause the

IFACMMM 2009. Viña del Mar, Chile, 14 -16 October 2009.


controller to be conservative and have a reduced feasible
region. Particle Size (PSE)

The RNMPC is implemented using open source modules Cyclone Feed (CFF)
that have good numerical stability. Open source software
has the added advantage of allowing tight integration of Sump Water
the software modules for optimisation purposes by mod- (SFW)
ifying the source code, which is not always possible with
proprietary software. Furthermore, open source software
is usually available free of charge. Implementing the con- Solids Feed (MFS)
troller with open source modules reduces the time needed Mill Water (MIW) Mill Load
to obtain a working controller compared to writing all the (LOAD) Sump
software modules in-house. Steel Balls (MFB) Level
(SLEV)
This study aims to (1) cast a complex problem of a
ROM ore milling circuit into a RNMPC framework, (2)
implement the RNMPC efficiently using open-source soft-
ware components, and (3) describes practically motivated
Fig. 1. Run-of-mine ore milling circuit.
simulation results of the mill model being controlled by a
RNMPC controller. 1a and 2) as well as power consumption and throughput
(objectives 2 and 4).
2. ROM MILLING CIRCUIT DESCRIPTION
2.2 Milling Circuit Model
The circuit is fed gold-bearing ore at about 100 tons/hour
and grinds it down to a product with a particle size The variables of the milling circuit (Fig. 1) that are typ-
of 80% smaller than 75 µm (P80 = 75µm). The ROM ically controlled are the product particle-size (PSE), the
mill is operated in closed circuit with a hydrocyclone fraction of mill volume filled with material (LOAD) and
that separates the fine material (product) from the coarse the level of slurry in the sump (SLEV). The inputs to the
material (out-of-specification material), which is recycled milling circuit that are typically manipulated automati-
to the mill. The gold is then extracted through a leaching cally are the volumetric flowrate of water to the sump
process downstream. (SFW), the volumetric flowrate of slurry to the cyclone
A typical mill has dimensions of 5 m in diameter and (CFF), the mass feed rate of solids to the mill (MFS), the
a length of 9 m. The mill is supported by pressurised- volumetric flowrate of water to the mill inlet (MIW) and
oil circumferential bearings. The mill features lifter bars mass feed rate of steel balls to the mill (MFB). Steel balls,
and solid white-iron liners and it is operated at 90% of which are usually added to the mill feed by the operator
critical speed (Stanley, 1987). The mill discharges slurry in discrete quantities, will for this study be treated as a
through an end-discharge grate into a sump. The slurry continuous variable.
is diluted with water in the sump and pumped to the The feed ore consists of rocks (that do not discharge from
hydrocyclone for classification. The hydrocyclone has an the mill), coarse ore (out-of-specification that do discharge
internal diameter of 1 m. The underflow of the cyclone, from the mill) and fine ore (in-specification material). The
water and feed ore constitute the mill feed. composition of the feed is described by the fraction that
consists of fines (αf ) and fraction that consists of rock
2.1 Objectives in Mill Control (αr ).
A complete description of the model is given in Coetzee
The control of the milling circuit has multiple objectives,
et al. (2009). The parameters that describe the ore and
firstly to stabilise the system and secondly to optimise
steel ball hardness (φf , φr and φb ) inside the mill, the cy-
the economics of the process (Hulbert, 1989). The eco-
clone classification (εc and αsu ) and the feed composition
nomic objective is divided into sub-objectives that each
(αf and αr ) are shown in Table 1.
contributes to the overall economic objective of the milling
process. A set of possible sub-objectives for the milling
circuit are to (Craig and MacLeod, 1995): 3. ROBUST NONLINEAR MODEL PREDICTIVE
CONTROL
(1) improve product quality
(a) by increasing grind fineness, and
NMPC utilises a nonlinear model to predict the behaviour
(b) decreasing the fluctuations in product size,
of the plant and calculate the optimal control moves or
(2) maximise throughput,
control laws with regard to a specified objective function.
(3) minimise the amount of steel that is consumed for
NMPC is derived from nonlinear optimal control over a
each ton of fines produced, and
constant or varying time interval into the future [tk , tk +T ].
(4) to minimise the power consumed for each ton of fines
Only the first control move or control law is implemented
produced, etc.
and a new state measurement is taken. The nonlinear
The objectives above are interrelated and require trade- optimal control problem is then recalculated for the new
offs to be made. The main trade-offs are between particle time interval [tk+1 , tk+1 + T ], which leads to receding
size of the product and the throughput of solids (objectives horizon control (Mayne et al., 2000).

IFACMMM 2009. Viña del Mar, Chile, 14 -16 October 2009.


Table 1. Parameter values for milling circuit. method (Diehl et al., 2005). This is accomplished by
dividing the prediction horizon [0, T ] into N discrete time
ParamNominal Min Max %∆ Description
Fraction of fines in the ore.
intervals called nodes t1  0 < t2 < t3 < · · · < tk < · · · <
αf 0.1 0.05 0.15 50
[dimensionless] tN −1 < tN  T where the sampling time is defined as
Fraction of rock in the ore. τs  tk+1 − tk .
αr 0.1 0.05 0.15 50
[dimensionless]
Power needed for a ton of The functions of time x(·) and u(·) are replaced by their
φf 28 14 42 50
fines produced. [kW·hr/ton] values at the nodes xk ∈ Rnx and uk ∈ Rnu for k =
Rock abrasion factor. 1, . . . , N and some form of interpolation between nodes.
φr 69 55 83 20
[kW·hr/ton] The inputs (uk ) are kept piece-wise constant between
φb 94 89 99 5
Steel abrasion factor. nodes. The resulting nonlinear controlled discrete-time
[kW·hr/ton] system is xk+1  fk (xk , uk , p̃), k = 1, 2, . . . , N − 1. The
Coarse split.
εc 184 175 193 5 nonlinear optimal control problem can now be cast into
[dimensionless]
Fraction solids in
the following nonlinear parameter optimisation problem
αsu 0.16 0.15 0.17 5
underflow. [dimensionless]
min φ(s, q) (4)
s,q
The nonlinear optimal control problem is to find a control
profile u(·) such that it minimises some scalar performance s.t. g(s, q, p̃, xk ) = 0 (5)
index i = 1, . . . , nc ,
θi,j (sj , qj ) ≤ 0, (6)
j = 1, . . . , N − 1,
min φc (x, u, T ) (1)
x,u where φ : RN ·nx ×R(N −1)·nu → R is the performance func-
s.t. ẋ(t) = fc (x(t), u(t), p̃), x1  x(0) ∀t ∈ [0, T ] (2) tion to be optimised, g : RN ·nx × R(N −1)·nu × R(N −1)·np ×
Rnx → RN ·nx is the equality constraint function that
θc (x, u) ≤ 0 (3) describes the discrete time system dynamics, θi,j : Rnx ×
where x : [0, T ] → R is the state trajectory, u :
nx Rnu → R, i = 1, . . . , nc , j = 1, . . . , N − 1 are the inequal-
[0, T ] → Rnu is the control trajectory, x(t) ∈ Rnx is the ity constraint functions, si ∈ Rnx , i = 1, . . . , N are the
state vector, ẋ(t) ∈ Rnx is the state sensitivities to time, estimated state parameters, qi ∈ Rnu , i = 1, . . . , N − 1 are
u(t) ∈ Rnu is the control vector, (x, u) → φc (x, u, T ) is the the control parameters, p̃  (p̃, . . . , p̃) ∈ R(N −1)np is the
scalar performance function, (x, u) → θc (x, u, T ) is the sequence of nominal model parameters, s  (s1 , . . . , sN ) is
inequality constraints function, p̃ ∈ Rnp is the nominal the state sequence and q  (q1 , . . . , qN −1 ) is the control
parameter vector and fc : Rnx × Rnu × Rnp → Rnx is the sequence to be optimised in the nonlinear optimisation
ordinary differential equation describing the dynamics of problem (Diehl et al., 2005).
the plant. The plant dynamics are time-invariant and each
optimal control problem can therefore be solved from time The scalar performance function is defined as
t1 = 0 without affecting the result. The initial state value N
 −1
x1 ∈ Rnx is the currently measured state of the system. φ(s, q)  Li (si , qi ) + E(sN , qN −1 ) (7)
 T i=1
For the sequel the ordered pair (a, b)  aT bT is
where the scalar interval performance indexes are defined
defined as a column vector.
as Li (si , qi )  h(si , qi )T Qh(si , qi ) + ∆qiT R∆qi where Q
Plant models, such as the mill circuit model described and R represent the weighting matrices on the outputs
above, always differ from the real system owing to incom- and controls respectively, h : Rnx × Rnu → Rny is the
plete modelling, parameter uncertainty and unmodelled function that maps the current state and control vector
disturbances. The NMPC developed earlier in (Coetzee to the output vector using nominal model parameters and
et al., 2008) will be extended to explicitly deal with model ∆qi  qi −qi−1 . The output vector consists of PSE, LOAD,
parameter uncertainty . Consider an uncertain parameter SLEV, Rheology factor (ϕ), THROUGHPUT and mill
vector p ∈ Rnp and nominal parameter vector p̃ ∈ Rnp , motor power draw (Pmill ). The terminal performance index
which are assumed to be restricted to a generalised ball is defined as E(sN , qN −1 )  h(sN , qN −1 )T P h(sN , qN −1 )
P = {p ∈ Rnp | p − p̃ ≤ 1} defined by using a suitable where P is the terminal cost weighting matrix. The equal-
norm · in Rnp (Diehl et al., 2006). ity constraint function describing the discrete time system
A suitable uncertainty description for the parameters is dynamics is defined 
as
box uncertainty where the upper pu and lower pl bounds on  xk − s1 ,

 f1 (s1 , q1 , p1 ) − s2 ,
the parameters
 p are known: P  {p ∈ R np
|p ≤ p ≤ p } =
  l
box u
 � −1 � pl +pu  g(s, q, p, xk )  .. (8)
p ∈ Rnp | diag pu −p l
p −  ≤ 1 , where the 
 .
2 2
∞ 
centre of the box is defined as p̄  pl +p u
∈ Rnp . In general fN −1 (sN −1 , qN −1 , pN −1 ) − sN .
2
the centre of the box p̄ and the nominal parameter vector To add uncertainty into an optimisation problem, a min-
p̃ do not have to be the same point (p̃ = p̄). max optimisation can be done (Diehl et al., 2006; Ma and
The nonlinear optimal control problem, consisting of a Braatz, 2001). The worst-case values for the cost φ(s, q) is
system with continuous dynamics, needs to be discretised defined as ψ(q)  maxs,p φ(s, q) s.t. g(s, q, p, xk ) = 0
in order to be cast in terms of a nonlinear parameter and the worst-case values for the constraint functions
optimisation problem, using the direct multiple shooting θi,j (sj , qj ) are defined as ωi,j (qj )  maxs,p θi,j (sj , qj )

IFACMMM 2009. Viña del Mar, Chile, 14 -16 October 2009.


s.t. g(s, q, p) = 0 where pi ∈ Pbox , i = 1, . . . N − 1 Table 2. Simulation Summary.
is defined as the unknown time varying model parame-
(N −1) Variable Value Variable Value
ters and p  (p1 , . . . , pN −1 ) ∈ Pbox is defined as the Prediction Horizon
sequence of time varying model parameters. The worst- (T )
60 s Nodes (N ) 6
case cost and constraint functions are calculated by max- Sampling Time (τs ) 10 s Simulation time 260 min
imising the cost function and constraint functions with
(N −1)
regard to the model parameter sequence p ∈ Pbox and Table 2. The milling circuit model contains large parame-
state values s ∈ R (N ·nx )
. The worst-case cost function ter uncertainties; this is especially true of the parameters
ψ(q) is then minimised by choosing the control moves related to the composition of the feed ore and the hardness
q  (q0 , q1 , . . . , qN −1 ) ∈ R((N −1)·nu ) subject to the worst- of the ore, which has an impact on the energy needed
case constraints ωi,j , i = 1, . . . , nc , j = 1, . . . , N − 1. to grind a ton of ore. The parameter variations used in
the simulation study are shown in Table 1. The parameter
This min-max optimisation problem is difficult to solve vector changes every 200 s, to allow the parameter distur-
for general nonlinear systems. The optimisation problem bances to sufficiently impact the simulation. The parame-
can however be simplified by approximating the worst- ters follow a uniform distribution to produce large changes
case calculations for the cost ψ̃ (s, q) and the constraints in the parameter values in order to properly demonstrate
ω̃i,j (s, q) , i = 1, . . . , nc , j = 1, . . . , N − 1. The ap- the disturbance rejection capabilities of the controller.
proximation of the worst-case cost ψ̃ (s, q) and constraints The milling circuit is simulated at the operating point as
ω̃i,j (s, q) can be done through liberation of the system described by the “OP” column in Table 3. The constraints
dynamics g (s, q, p, xk ) = 0, the cost φ(s, q) and constraint of the milling circuit are described by the “Min” and “Max”
θi,j (sj , qj ) functions. columns in Table 3. Metric ton is used throughout this
The approximation of the worst-case cost ψ (s, q) by paper. The weighting of the variables by Q and R in the
ψ̃ (s, q), is defined by a convex optimisation problem objective function (7) is described by the “W” column in
Table 3 and chosen based on the performance criteria of
∂φ(s, q) Section 2.1. Further, P = Q with no terminal constraints
max φ(s, q) + ∆s (9) (θN (sN ) ∈ Rnx ). The inputs are normalised according
∆s,∆p ∂s to their maximum range and outputs are normalised
∂g (s, q, p̄, xk ) ∂g (s, q, p̄, xk ) according to their setpoints in the objective function.
s.t. ∆s + ∆p = 0,(10)
∂s ∂p
Feed ore hardness and composition changes are major dis-
∆pl  ≤ 1, l = 1, . . . , N − 1, (11) turbances that milling circuit controllers have to contend
and the approximation of the worst-case constraints with, especially when the feed ore is switched between
ωi,j (s, q) by ω̃i,j (s, q) are defined as feeds that originate from different stockpiles. A feed ore
hardness increase is simulated by increasing the power
∂θi,j (sj , qj ) needed to produce a ton of fines (φf ) by 50% at time
max θi,j (sj , qj ) + ∆s (12) 10 min. A feed ore composition change is simulated by
∆s,∆p ∂s
increasing the fraction of the feed consisting of rock (αr )
∂g (s, q, p̄, xk ) ∂g (s, q, p̄, xk ) by 50% at time 100 min. These disturbances are very large
s.t. ∆s + ∆p = 0,(13)
∂s ∂p but not uncommon in practice.
∆pl  ≤ 1, l = 1, . . . , N − 1, (14) Coetzee et al. (2009) investigated the disturbance rejection
where ∆s ∈ R(N ·nx ) , p̄  (p̄, . . . , p̄) ∈ R(N −1)·np is a capability of the RNMPC and found it to be very good.
sequence of parameters at the centre of the box, ∆pl  pl − This simulation scenario investigates the ability of the
p̄ ∈ Rnp is the deviation of the model parameters from RNMPC to follow a large setpoint change in PSE in the
the centre of the box and ∆p  (∆p1 , . . . , ∆pN −1 ) ∈ presence of large disturbances described above.
R(N −1)·np is defined as the sequence of model parameter MFS (see Fig. 2) represents the throughput of new feed ore
deviations. The approximate min-max problem to the mill, while THROUGHPUT (see Fig. 2) represents
the product throughput that is sent downstream for fur-
min ψ̃(s, q) (15) ther processing. At steady state, these two values should be
s,q

i = 1, . . . , nc , Table 3. Constraints and operating point.


s.t. ω̃i,j (s, q) ≤ 0, (16)
j = 1, . . . , N − 1.
Variable Min Max OP Weight Units
g (s, q, p̄, xk ) = 0 (17) MIW 0 100 33 0.01 m3/hour

can be solved efficiently by casting it into a form that MFS 0 200 100 0.01 tons/hour

preserves the sparsity of the problem as well as smooth MFB 0 4 2 0.01 tons/hour
3
CFF 400 500 442 0.01 m /hour
objective and constraint functions (Diehl et al., 2006). m3/hour
SFW 0 400 267 0.01
PSE 60 90 80 100 % < 75µm
4. SIMULATION AND RESULTS LOAD 30 50 45 100 m3
SLEV 2 9.5 5.0 1 m3
In this simulation scenario the “actual” plant differs from ϕ 0 1 0.51 0 dimensionless
the nominal model. The prediction horizon, sampling time, THROUGHPUT 100 0 200 1 tons/hour

number of nodes and simulation time are summarised in Pmill 0 2000 2000 0 kW

IFACMMM 2009. Viña del Mar, Chile, 14 -16 October 2009.


90 500
(% < 75 µ m)

(m / hour)
80

CFF
70

3
PSE

60 0
0 50 100 150 200 250 0 50 100 150 200 250
200

(tons / hour)
(% Full)

40
LOAD

100

MFS
20 0
0 50 100 150 200 250 0 50 100 150 200 250
100 500

(m / hour)
(% Full)
SLEV

50

SFW
3
0 0
0 50 100 150 200 250 0 50 100 150 200 250
100
THROUGHPUT

200

(m / hour)
(tons / hour)

100 50

MIW
3
0 0
0 50 100 150 200 250 0 50 100 150 200 250
1 5

(m / hour)
Rheology
Factor

0.5

MFB
3
0 0
0 50 100 150 200 250 0 50 100 150 200 250

2000
Pmill (kW)

(kg/m )
1.4

3
CFD
1000
1.3
0 1.2
0 50 100 150 200 250 0 50 100 150 200 250
Time (minutes) Time (minutes)

Fig. 2. RNMPC: Controlled variables (PSE, LOAD, SLEV and THROUGHPUT ), manipulated variables (CFF, MFS,
SFW, MIW and MFB) and other important variables (Rheology Factor, Pmill and CFD). The dashed lines indicate
the constraints on the variable and the vertical dotted lines indicate the start of the disturbance events. The
dash-dot line indicates the setpoint.

the same. The throughput settles at a lower value, because ore hardness disturbance by increasing MIW relative to
the feed ore hardness was increased, thus deviating from MFS in order to reduce the mill discharge density to
the nominal operating conditions. The controller reduces maintain PSE at setpoint. The increase of MIW relative
THROUGHPUT in order to maintain PSE at setpoint, to MFS leads to an increase in the rheology factor and
which is weighted as being more important. the RNMPC increases MFB (see Fig. 2) to compensate
for the higher rheology factor in order to maintain proper
The PSE setpoint is lowered from 80% to 70% at time 100
breakage conditions inside the mill.
min to increase THROUGHPUT, because THROUGH-
PUT and PSE is inversely proportional. At time 100 min, the controller further increases MIW
and as a result increases the rheology factor that leads
PSE (see Fig. 2) is controlled through changing the cut
to a decrease in MPOW. The drop in MPOW results in
of the cyclone. The cut of the cyclone is changed through
a coarser grind and, therefore, a decrease in PSE from
the CFF (see Fig. 2) and the cyclone feed density (CFD).
80% to 70%. This strategy is not ideal, because reducing
The CFD (see Fig. 2) can be changed by changing the mill
grinding efficiency decreases THROUGHPUT when the
discharge density through the ratio of MIW (see Fig. 2) to
PSE setpoint is lowered, but the strategy is consistent with
MFS and, more importantly, by changing the sump slurry
the weightings of the variables (see Table 3). The control
density through SFW (see Fig. 2). The controller therefore
strategy can be corrected by removing the weighting on
has three degrees of freedom with which to control the
SLEV and increasing the weighting on MIW relative to the
PSE. The SLEV (see Fig. 2) is given a setpoint, because
other MVs, which should result in the RNMPC changing
most operators do not trust the controller if it allows SLEV
PSE with SFW and CFF, rather than MIW and CFF, and,
to vary too much. The weighting on SLEV is, however,
therefore, maintain grinding efficiency of the mill and as a
very low (see Table 3) to allow the controller to vary SLEV
result increase THROUGHPUT when the PSE setpoint is
when needed in order to control more important variables,
lowered.
such as PSE.
PSE and LOAD are maintained at their desired setpoints
under RNMPC (see Fig. 2) regardless of the active distur- 5. CONCLUSIONS
bances.
The rheology factor (see Fig. 2) is a function of the The results of a practically motivated simulation show that
water and solids inside the mill. The RNMPC manipulates an RNMPC controller can successfully control important
MIW and MFS to maximise throughput by maintaining milling circuit variables in the face of large disturbances
optimum breakage conditions inside the mill, which are that are not uncommon in practice. The simulation results
obtained when the rheology factor is at the optimum show that the correct weightings of the variables in the
value of 0.51. The controller compensates for the feed RNMPC objective function is very important to produce
the correct behaviour.

IFACMMM 2009. Viña del Mar, Chile, 14 -16 October 2009.


The RNMPC was implemented using open-source optimi- Craig, I.K. and MacLeod, I.M. (1995). Specification
sation software IPOPT (Kawajir et al., 2006) and open- framework for robust control of a run-of-mine ore milling
source automatic differentiation software CppAD (Lougee- circuit. Control Engineering Practice, 3(5), 621–630.
Heimer, 2003). The simulation executed with an aver- Diehl, M., Bock, H.G., and Kostina, E. (2006). An approx-
age time of about 26 s and a maximum time of 123 s imation technique for robust nonlinear optimization.
per iteration on a Dell PowerEdge 1955 blade with Intel Mathematical Programming: Series A and B, 107(1),
Xeon 5140 (Dual-Core) 2.33GHz processor, 2GB RAM 213–230.
and 1333MHz FSB. The current RNMPC implementation Diehl, M., Bock, H.G., and Schloder, J.P. (2005). A
is not feasible for practical implementation, because the real-time iteration scheme for nonlinear optimization in
maximum and average calculation times are much longer optimal feedback control. SIAM Journal on Control and
than the recommended sampling time of 10 s (Craig and Optimization, 43(5), 1714–1736.
MacLeod, 1995). Hulbert, D.G. (1989). The state of the art in the control
of milling circuits. In 6th IFAC Symposium on Automa-
Tuning the controller will include the selection of the
tion in Mining, Mineral and Metal Processing (Buenos
prediction horizon for stability and performance, while
Aires).
maintaining a reasonable calculation time. The computa-
Kawajir, Y., Laird, C., and Wachter, A. (2006). Introduc-
tion time can also be reduced by using less nodes without
tion to Ipopt: A tutorial for downloading, installing, and
affecting the prediction horizon T , leading to an effective
using Ipopt, Revision: 799. Carnegie Mellon University,
shorter control horizon. The control algorithm is hampered
Pittsburgh, PA, USA. URL https://projects.coin-
mainly by a large number of slack variables for imple-
or.org/Ipopt.
menting robustness and the Hessian calculations. With
Lazar, M., Munoz de la Pena, D., Heemels, W.P.M.H.,
the continuous increase in computing power, this should
and Alamo, T. (2008). On input-to-state stability of
become less of an issue in the foreseeable future.
min-max nonlinear model predictive control. Systems &
The simulation further assumed full-state feedback which Control Letters, 57(1), 39–48.
is not available on real plants. Typically the controlled Limon, D., Alamo, T., Salas, F., and Camacho, E.F.
variables would be measured online (Wei and Craig, 2009) (2006). Input to state stability of min-max MPC con-
from which an observer would be needed to infer the model trollers for nonlinear systems with bounded uncertain-
states. ties. Automatica, 42(5), 797–803.
Limon, D., Bravo, J.M., Alamo, T., and Camacho, E.F.
The adoption of advanced control by the mineral process- (2005). Robust MPC of constrained nonlinear systems
ing industry will probably be determined by the trade- based on interval arithmetic. IEE Proceedings of Control
off between the added complexity of implementing and Theory and Applications, 152(3), 325–332.
maintaining an advanced controller such as RNMPC, and Lougee-Heimer, R. (2003). The Common Optimiza-
the benefits that can be derived from such an imple- tion INterface for Operations Research. IBM Jour-
mentation. Results given in this paper suggest that if a nal of Research and Development, 47(1), 57–66. URL
milling circuit regularly experiences large changes in feed http://www.coin- or.org/CppAD/.
ore hardness and composition, when for example the feed Ma, D.L. and Braatz, R.D. (2001). Worst-case analysis
ore is switched between feeds that originate from different of finite-time control policies. IEEE Transactions on
stockpiles, RNMPC might well warrant a closer look. Control Systems Technology, 9(5), 766–774.
Magni, L., De Nicolao, G., Scattolini, R., and Allgöwer,
ACKNOWLEDGEMENTS F. (2003). Robust model predictive control of nonlinear
discrete-time systems. International Journal of Robust
This paper is published with the permission of Mintek. and Nonlinear Control, 13, 229–246.
This work was sponsored by the Nation Research Foun- Mayne, D.Q., Rawlings, J.B., Rao, C.V., and Scokaert,
dation of South Africa, Mintek and the University of P.O.M. (2000). Constrained model predictive control:
Pretoria. Thanks to Dr. Dave Hulbert of Mintek for the Stability and optimality. Automatica, 36, 789–814.
nonlinear milling circuit model and for discussions on the Mhaskar, P., El-Farra, N.H., and Christofides, P.D. (2005).
modelling and control of milling circuits. Robust hybrid predictive control of nonlinear systems.
Automatica, 41(2), 209–217.
Mhaskar, P., El-Farra, N.H., and Christofides, P.D. (2006).
REFERENCES Stabilization of nonlinear systems with state and control
constraints using Lyapunov-based predictive control.
Bravo, J.M., Alamo, T., and Camacho, E.F. (2006). Ro- Systems & Control Letters, 55(8), 650–659.
bust MPC of constrained discrete-time nonlinear sys- Mhaskar, P. and Kennedy, A.B. (2008). Robust model pre-
tems based on approximated reachable sets. Automatica, dictive control of nonlinear process systems: Handling
42(10), 1745–1751. rate constraints. Chemical Engineering Science, 63(2),
Coetzee, L.C., Craig, I.K., and Kerrigan, E.C. (2008). 366–375.
Nonlinear model predictive control of a run-of-mine ore Stanley, G.G. (1987). The extractive metallurgy of gold
milling circuit. In proceedings of the 17th IFAC World in South Africa. Technical Report Vol 1, South African
Congress, July 6-11. Seoul, Korea. Institute of Mining and Metallurgy, Johannesburg.
Coetzee, L.C., Craig, I.K., and Kerrigan, E.C. (2009). Wei, D. and Craig, I.K. (2009). Grinding mill circuits – A
Robust nonlinear model predictive control of a run-of- survey of control and economic concerns. International
mine ore milling circuit. IEEE Transactions on Control Journal of Mineral Processing, 90(1-4), 56 – 66.
Systems Technology, accepted for publication.

IFACMMM 2009. Viña del Mar, Chile, 14 -16 October 2009.

You might also like