Minerals Engineering 53 (2013) 113–123

Contents lists available at ScienceDirect

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

Benefits of optimisation and model predictive control on a fully

autogenous mill with variable speed
C.W. Steyn, C. Sandrock ⇑
Department of Chemical Engineering, University of Pretoria, c/o Roper and Lynwood Road, Pretoria, South Africa

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

Article history: Autogenous (AG) milling is utilised around the world for particle size reduction. The system exhibits
Received 19 August 2012 highly non-linear behaviour in addition to being subject to unmeasured variability associated with most
Accepted 22 July 2013 ore bodies. Anglo American Platinum aimed at improving online optimisation of the circuit by imple-
menting industrial model predictive control (MPC) to reduce system variability and continuously drive
towards the optimal operating point within system constraints.
Keywords: The industrial dynamic matrix controller commissioned on the AG mill with a variable speed drive
Autogenous
resulted in a 66% reduction in power and a 40% reduction in load standard deviation. These are the main
Milling
Optimisation
controlled variables of the mill. The controller also improved the objective function, effective power util-
Response surface analysis isation, by 11%. This reduction in operated variable variability enabled a test campaign where the mill
Model based control was controlled at various operating regions in order to establish the conditions conducive to the finest
Benefit analysis product size at a given mill feed rate.
Moving the mill operating region from the benchmarked plant to the optimal grind environment and
stabilising the mill at this point with the model predictive controller provided an estimated potential
recovery increase of 0.32% (absolute) due to better liberation.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction The next step in the milling optimisation initiative was to

implement a similar solution on its AG mills with the following
A large portion of the total operating cost in the mineral bene- objectives:
ficiation process is associated with comminution circuits. The high-
est costs in these operations are normally grinding media and 1. Obtain similar variability reduction in the system’s main con-
energy. Since the global economic downturn at the end of 2008, trolled variables, power and load.
the control of these cost elements, together with optimised perfor- 2. Find the optimal operating using grindcurves (Powell and Main-
mance especially on the milling component of these circuits, has za, 2006; van der Westhuizen and Powell, 2006) and exploit
become increasingly important. this in the MPC algorithm to operate the mill in a way that max-
Anglo American Platinum employs numerous fully autogenous imises financial benefit.
(AG) primary mills. All are under Neuro-Fuzzy control, applied in 3. Quantify the real benefits of reducing the variability in the mill
a manner suitably robust to perform in the non-linear, highly and stabilising the circuit at the optimum operating conditions.
unmeasured environment of primary milling (Steyn et al., 2010).
These mills have recently been fitted with model predictive con- The primary AG mill at Rustenburg Platinum Mines Amandelb-
trollers that enable real-time optimisation and minimise a power ult UG2 No2 (RPMA2) plant was selected for this application due to
utilisation objective function (kW h/t75 lm). The algorithm was the additional potential of its variable speed facility. To achieve the
originally piloted on one of Anglo American Platinum’s ROM mills, objectives listed above, the following goals were established:
Mototolo JV. The success of the controller was immediately evi-
dent, demonstrating a reduction in variability of between 15% Benchmark the system based on the key performance
and 20% on the primary milling parameters (see Fig. 1) and also indicators that have been established for
improving energy efficiency. A reduction in controller input vari- comminution circuits.
ance with a clear bias towards the local optimum in the top left Model the relationship between the identified
corner of the feasible region is indicated in Fig. 1. variables in a dynamic empirical model
suitable for use in an online controller.
⇑ Corresponding author. Tel.: +27 124202197.
E-mail address: [email protected] (C. Sandrock).

0892-6875/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.mineng.2013.07.012
114 C.W. Steyn, C. Sandrock / Minerals Engineering 53 (2013) 113–123

Table 1 the feed of the mill while the undersize particles are gravity fed
List of variables. to an agitated 200 m3 flotation feed surge vessel (rougher feed
Symbol Description tank).
CV Controlled variable 2.2. Control infrastructure
MV Manipulated variable
J Objective function
f Open loop response vector of the system Anglo American Platinum follows a two-tiered approach when
SSCosti Steady state cost for MV i controlling its primary mills. The top tier consists of an in-house
G The multi-variable free response matrix of the system developed supervisory advanced control system that is based on
y System outputs (CVs)
fuzzy-logic rules and cascades to the bottom tier base-layer control
u System inputs (MVs)
gpower Effective power utilisation of the milling circuit in kW h/t (passing schema (Steyn et al., 2010). The base-layer control of the circuit
75 lm) produced runs on the PLC which accounts for the PID control algorithms,
S Mill speed in % critical speed interlocks and sequences. The mill is controlled by providing three
Wrat Mill inlet water ratio manipulated variables (MVs) with setpoints from either the hu-
F Mill ore feed in tph
man–machine interface (HMI) or the advanced control layer. These
G Product grind in % passing 75 lm
PR Potential Recovery MVs are (use Fig. 2 as reference):
P(i) Profit function in terms of variable i
C(i) Cost function in terms of variable i Total ore feed-rate: Controlled by a PID controller that cascades
PNP Potential net profit
down to a coarse ore ratio controller. The
PF Performance function
total ore process variable PV is obtained
from a weightometer situated on the com-
Optimise the operation of the mill by utilising the bined ore (fine + coarse) feed belt.
improved circuit stability in a trial to deter- Coarse ore ratio: This is the ratio of coarse ore mass flow to
mine the optimal operating region. total ore mass flow (excluding the recycled
Analyse the benefits that the advanced controller introduces to ore). The ratio controller cascades to two
the system. ore feed rate PID controllers, which both
receive their PVs from weightometers situ-
Table 1 shows the symbols used in this work. ated on the coarse and fine ore feed con-
veyor belts and actuate the respective ore
feeders.
2. Circuit benchmark Inlet water ratio: The PV of this controller is the calculated
ratio of inlet water flow to total ore fed to
2.1. Process description the mill. The SP is assigned to a ratio con-
troller that cascades down to a flow PID
Rustenburg Platinum Mines Amandelbult UG2 No2 (RPMA2) is controller which actuates the inlet water
one of Anglo American Platinum’s mineral ore processing plants control valve.
situated in the North West province of South Africa. Ore from the
mine undergoes one stage of top size reduction in a primary jaw- The advanced milling controller at RPMA2 utilises a 3  3 Qua-
crusher before it is classified by a 100 mm grizzly screen and con- si-Fuzzy Rules-Based controller that uses the same SP as the base
veyed to the fine and coarse ore silos. The fine and coarse ore is layer to control the power, load and screen current of the circuit.
treated by a 6.1 m /  8.5 m fully autogenous primary mill (see An MPC with real-time optimisation was added to the advanced
Fig. 2). Autogenous mills are defined by the ore being the only layer. The MPC is used during normal operation, but control is
source of grinding media and the absence of a secondary media switched to the fuzzy system during abnormal situations. The
such as steel. The circuit is operated in a closed loop with a selection between the two algorithms is governed by the supervi-
horizontal vibrating screen fitted with 630 lm polyurethane sory controller that detects normal or abnormal conditions based
classification panels. The oversize from the screen is recycled to on process state rules.
Power [MW]

Power [MW]

Load [ton] Load [ton]

Fig. 1. A 2D density plot of the power and load for Mototolo (JV) ROM primary mill on data (a) two weeks prior and (b) two weeks post MPC implementation.
 C.W. Steyn, C. Sandrock / Minerals Engineering 53 (2013) 113–123 115

HMI SP

HMI SP
Classification screen
PID

HMI SP
PID
Mill Inlet H2O
FFC
WFC

WT

Coarse ore feeder PID RT

Dilution H2O

FT

PID PID

FT
WT
PID
RT
WT
Fine ore feeder
HMI SP

LT PID

Primary AG mill
PID FT

Discharge sump

Fig. 2. Anglo American Platinum’s primary milling base-later control infrastructure at RPMA2.

2.3. Controller design and implementation 2. Obtaining a Feasible Solution: The next step in the DMCplus™
control law is to determine whether a feasible solution is possi-
In this instance, Anglo American Platinum has used AspenTech ble for any combination of inputs. A feasible solution is possible
DMCplus™ as the platform to develop its model predictive control when the allowed MV movement within the MV constraints can
infrastructure. bring all the CVs within limits (Fig. 3). If no solution is available,
the algorithm will relax its CV limits in accordance to a rank of
2.3.1. Control law for DMCplus™ importance.
The model predictive control solution is computationally inten- As soon as a solution is available, the algorithm proceeds to the
sive and in some cases it is difficult to ensure feasibility when solv- next step with an updated set of CV limits required for a feasible
ing the quadratic objective function given in: solution.
3. Steady-State Optimisation: The aim of optimisation is to mini-
X
N2
JðN1 ; N2 ; N u Þ ¼ ^ðt þ jjtÞ  xðt þ jÞ2
dðjÞ½y mise the objective function for a given system within the allow-
j¼N 1 able process constraints (Snyman, 2005). The mathematical
form of this constrained problem is given as follows:
X
Nu
2
þ kðjÞ½Duðt þ j  1Þ ð1Þ
j¼1 min f ðxÞ; x ¼ ½x1 ; x2 ; . . . ; xn T 2 R ð3Þ
In an attempt to industrialise the algorithm, AspenTech developed a Subject to constraints usually of the form:
staged procedure where feasibility is ensured upfront (possibly by
relaxing constraints). This enables an analytic solution of the MV g j ðxÞ 6 0; j ¼ 1; 2; . . . ; m
trajectory. This four staged procedure is explained as follows with
hj ðxÞ ¼ 0; j ¼ 1; 2; . . . ; r
each cycle abiding to the routine (AspenTech, 2000):
The AspenTech framework (DMCplus™) allows for the usage of
1. Open Loop Prediction: The open-loop prediction or free response either linear or quadratic programming (AspenTech DMCplus™
of the system can be extended as the sum of four effects; fu the course notes):
response of the system due to past control movement, Dd + fd
(a) Linear Programming: A special case of general optimisation
the response to known disturbances and fn unmeasured system
arises when both the objective function and the constraints
disturbances or model error.
are linear functions (Snyman, 2005). Such an instance is
f ¼ f u þ Dd þ f d þ f n ð2Þ called linear programming, stated in the following form:
116 C.W. Steyn, C. Sandrock / Minerals Engineering 53 (2013) 113–123

CV2,HI
CV2,HI
CV1,LO

MV2 Optimum Point

MV2,HI

Decrease J
CV2,LO

Feasible Region

MV2,LO

MV1

MV1,LO MV1,HI

Fig. 3. Graphical representation of the feasible region defined by MV and CV constraints.

minf ðxÞ ¼ cTx Du ¼ ½GT G GT e

1
ð7Þ
x

such that ð4Þ (b) Minimise MV movement: Minimising CV error in the most
Ax 6 b aggressive manner might provide the best control solution.
Model error, circuit constraints, etc. can however, induce
Within DMCplus™ the vector c in Eq. (4) is an n-vector containing
cycling and other unwanted control effects if aggressive
the steady-state cost parameter SSCost allocated to each manipu-
MV movement is allowed. An MV movement suppression
lated variable. The SSCost vector represents the direction of mini-
parameter K is added to Eq. (7) where:
mising the linear programming objective function. Duss represents
the steady state MV step conducive to the lowest J: K  Du ¼ 0 ð8Þ
X
min J ¼ nðSSCosti  Duss Þ ð5Þ
The components in Eq. (7) change as a result of MV minimisation
i¼1
(Eq. (8)) to:
(b) Quadratic Programming: The objective function is defined as a  
positive-definite quadratic function subject to these linear G
G0 ¼ ð9Þ
constraints (Snyman, 2005). The minimisation function in KI
DMCplus™ is given by:
(c) CV Importance: To distinguish between the importance of vari-
!2
X ous CVs, a weight Wi is assigned to the error of each CV.
min J ¼ nðSSCost i  Duss Þ  M ð6Þ
i¼1 1
Du ¼ ½G0T G0  G0T eW ð10Þ
Where the maximum profit, M, is defined as the point on a plane
only subject to MV constraints (represented as ‘‘Optimum Point’’
in Fig. 3). 2.3.2. Modelling
4. Future MV movement: The next step in the DMC control law is to DMCplus™ conducts its process modelling by means of a step-
calculate the future input movement in order to get the outputs test campaign where all the independent parameters are moved in
to the steady-state end values. order to produce responses from the dependent variables. Further-
(a) Minimise CV Error: CV error e in this instance is defined as more, DMCplus™ utilises finite-impulse-response as well as state-
the difference between the open loop response and the space modelling (Camacho and Bordons, 2007), as techniques to
steady state value (or SP). MV movement to achieve mini- determine the matrix of linear models describing the milling cir-
mum CV error is given as: cuit relationships.
 C.W. Steyn, C. Sandrock / Minerals Engineering 53 (2013) 113–123 117

The following characteristics are observed from the response The linear objective function obtained used for the DMCplusTM
matrix obtained during September 2010: controller’s steady-state optimisation can thus be presented
General. dynamically as:
DJ ¼ Dgpower ¼ 6:6 þ 0:27DS þ 0:165DW rat  0:022DF ð11Þ
1. The model horizon (or prediction horizon) of the controller that
produced the most reliable models was 60 min.
2. One of the most significant observations is that the mill load 3. Optimisation
behaviour is described as an integrator over the allocated model
horizon. 3.1. Optimisation strategy
3. The inlet water ratio to mill power model presented a positive
gain for the operating region of the mill. This indicates, accord- As explained, the mill feed maximisation is regarded as the pri-
ing to Morell and Kojovic (1996), that the flowrate through the mary objective and mill speed minimisation as secondary. In doing
mill is low enough for the interstices within the charge not to be this, the mill inlet water will be moved and kept at its high limit, to
completely occupied and that no slurry pool has formed. allow maximum speed reduction. This is due to the fact that the in-
4. Speed increases produced a positive, higher order, under- let water ratio shares a gain and optimisation directionality with
damped response from power which concurs with the findings speed (considering load). The three degrees of freedom available
of van der Westhuizen and Powell (2006). to the system are consumed by the feed high limit, the load high
limit due to reducing speed and the inlet water ratio high limit.
Model Gain: If the system is scaled according to input range and This optimisation strategy means that the mill will be fed at the
output error (Skogestad and Postlethwaite, 2005), the ore feed target feed rate while operated at an optimum viscosity (Klimpel,
yields the largest power gain with inlet water the second highest. 1983; Napier-Munn et al., 1999), optimum volumetric fill and low-
Mill speed yields the largest gain on mill load for both scaled and est speed (van der Westhuizen and Powell, 2006). The last three
unscaled models with inlet water ratio in second. It is noticeable factors ensure the highest possible grind for the given ore target.
that the coarse ore ratio produced the lowest scaled gain for both
power and load (see Fig. 4). 3.2. Optimisation framework

The objective of this framework is to verify the optimum oper-

2.3.3. Objective function ating region of the mill, defined for a certain throughput, at the
The objective of this controller is to minimise the effective load and inlet water ratio high limits. As the benefit of milling cir-
power usage gpower of the milling circuit, in units kW h/t75 lm. cuits is usually expressed in terms of liberation or particle size dis-
The direction and magnitude of the SSCost vector in Eq. (5) are ob- tribution of the product (Hulbert, 2002; Sosa-Blanco et al., 2000;
tained by modelling the effective power usage of the mill to the Hodouin et al., 2001), the optimal operating region is defined as
speed and inlet water ratio. the limits conducive to the best grind. These limits were obtained
In the absence of an online measurement, the 8 h (shiftly) com- by conducting two optimisation trials.
posite grind samples are used to calculate gpower. This calculation is First, the grind to load relationship was determined by operat-
used to model the objective function using the manipulated vari- ing the mill at three different load high limits. The grindcurves
ables speed S, inlet-water ratio Wrat and mill feed rate F. A linear which resulted from the first trial suggest a second order polyno-
fit was performed where, as per Eq. (5), the SSCost vector is defined mial grind/load model:
by the model coefficients. 450 valid shiftly samples were obtained GðLÞjW rat ¼0:24 ¼ 0:011x2 þ 4:4x  421:4 ð12Þ
during the period 1 January 2010 to 1 October 2010. The linear
regression model achieved an R2 = 20% which is fairly low under The maximum grind is obtained at dy dx
¼ 0:022x þ 4:4 ¼ 0 which
normal circumstances. However, taking into account that compos- equates to a mill load of VGrind,max = 211.2 ton mill load (see
ite average samples over an 8 h period were used, this model was Fig. 5). The average power obtained at these operating regions also
assumed to be good enough to serve as an indication for the opti- formed a parabola with maximum power of 2900 kW achieved to
misation directions. A pure-quadratic regression was also carried the left of VGrind,max.
out in order to validate the directionality of the linear model. A Next the effect of inlet water on grind was determined. Note
root-mean-square-error (RMSE) equal to 1.7 kW h/t75 lm was that this trial was aimed at expanding the response surface on
achieved, which relative to gpower equates to an acceptable RMSErel, the inlet water axis. The response surface of the load and inlet-
equal to 7%. Both the linear and pure-quadratic model resulted in water resulted in a model that obtained a root-mean-square-error
positive coefficients for water and speed but negative for feed. value of 1.06% or 3% relative to the mean of the grind. The pure-

Fig. 4. The step-response model gain and integrating rate for mill power and load respectively.
118 C.W. Steyn, C. Sandrock / Minerals Engineering 53 (2013) 113–123

quadratic relationship of load and inlet water ratio on grind is gi- in a structured ON/OFF manner. This
ven as follows: could however not be performed due to
G0 ðL; W rat Þ ¼ 0:0076  L2 þ 3:2  L þ 266  W rat  592:48 the expense and time demands of a prop-
erly designed ON/OFF experiment on an
 W 2rat  336:6 ð13Þ entire circuit. In order to confirm the sig-
The point which maximised the grind based on Eq. (13) is obtained nificant reduction in variability observed
at @G ¼ 0 and @W@G
¼ 0 for load and inlet water ratio, respectively. in Fig. 6, the theoretical minimum vari-
@L rat
This amounts to a load of Lopt = 213 ton and an inlet water ratio of ance of the load was investigated to
Wrat = 0.224. It is important to note that this is only an empirical determine whether any change in ore
model, applicable to the primary mill and UG2 ore body of RPMA2 properties are observed.
at a nominal feed rate of 350 t/h. Due to the seemingly high model The power spectral density of the mill
error this optimum is not considered exact but rather an approxi- load indicates a decrease in amplitude
mate optimal operating region of the mill. during MPC operation at frequencies
x < 3  104 (Fig. 7). The mill load
3.3. Performance results appears to converge on the minimum var-
iance approximation at lower frequencies
As a preliminary indication of the MPC performance, the first (see Fig. 7). This concurs with the results
month of operation (19 October to 22 November 2010) was com- shown in Fig. 6. The minimum variance
pared to the baseline Fuzzy-logic Controller (FLC) performance of the load indicated very little difference
during the period 1 August 2010 to 9 September 2010. Only good between data for before and after the pro-
quality data at a sampling rate of 1 point per minute was used. This ject. This implies that best-case control
produced 16430 good data points during the MPC period and performance were similar during the
17180 during the FLC period. Good data was defined as plant run- two periods in question and that ore-
ning data, determined by whether the mill was receiving a feed properties can be considered similar.
rate higher than 100 tph. To further verify the reduction in vari-
The following results were observed: ance, the historical standard deviation of
load on a month to month basis was com-
Mill Load and Power: The mill load standard deviation was pared to that observed as a result of the
decreased from 6.8 ton to 2.3 ton, a 66% project. A load standard deviation consis-
decrease (Fig. 6). The standard deviation tent with the 6.8 ton observed in Fig. 6
of the mill power was reduced by was obtained for the 10 months before
145 kW, from 364 kW to 219 kW. The the project (Fig. 8). The reduction in load
2D density plot illustrates how both the standard deviation from 6.8 ton to 2.3
load and power plots are more dense ton are observed for three months after
and closer to the load high limit as a commissioning the MPC. This constitutes
direct result of the reduction in deviation. a 66% reduction in load variability, which
Hulbert (2002) and Craig and Koch (2003) is comparable to the 64% reduction
warn that comparing the performance of achieved by Rogers et al. (2010) during a
the improved, post-project circuit to the similar MPC application at Lumwana Cop-
benchmarked plant can lead to bias. The per. Further reductions in variability of
reason for this bias is due to unmeasured both MVs and CVs were reported in the
external factors such as ore properties literature when moving from fuzzy-logic
that might be different for the two cases. expert systems to MPC (Jonas, 2008; Rog-
The recommended way of removing such ers et al., 2010).
bias is to reducing the effect of these Objective Function: The effective power utilisation gpower was
external factors by conducting an experi- reduced by 11%, from 25.2 kW h/t75 lm
ment where both the cases are utilised to 22.4 kW h/t75 lm. The zero-hypothesis
of no difference between the means was
rejected with a 99% confidence using a
two-sample t-test, assuming unequal vari-
ance. A time based plot of the normalised
shiftly averages shows that this reduction
in gpower was probably as a result of a
change in operating philosophy with a vis-
ible change in operating areas for MVs
speed and inlet water (Fig. 9). Note that
the measured disturbance feed size
remained unchanged which indicates that
the change in objective function was prob-
ably not due to external factors.

3.4. Assessment of Neuro-Fuzzy and MPC performance

It can be argued that the success of the MPC over the Neuro-
Fuzzy controller was due to insufficient tuning or bad design on
Fig. 5. RPMA2 primary AG Mill grindcurves. the part of the latter. There are however fundamental differences
 C.W. Steyn, C. Sandrock / Minerals Engineering 53 (2013) 113–123 119

3800 3800

3600 3600

3400 3400

3200 3200

3000 3000
Power [kW]

Power [kW]
2800 2800

2600 2600

2400 2400

2200 2200

2000 2000

1800 1800

170 180 190 200 210 220 170 180 190 200 210 220
Load [ton] Load [ton]

Fig. 6. Primary mill power to load curve as a 2D histogram with controller limits (red) (a) base-case (1 August to 9 September 2010) and (b) first month of MPC operation (19
October to 22 November 2010). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

2
10 (2009). Note that this benefits analysis reports on potential recov-
MinVar Post-MPC ery and hence potential benefit of the project.
1 Actual Post-MPC
10 MinVar Pre-MPC
Actual Pre-MPC
4.1. Performance functions
Power Spectral Density

0
10

-1
4.1.1. Profit function
10 The economic consequence in terms of profit is defined by the
effect of primary grind on PGM recovery in the flotation circuit
-2
10 (Wei and Craig, 2009; Sosa-Blanco et al., 2000). Due to numerous
factors influencing the final PGM recovery of a concentrator,
-3
10 extracting a representative model of the primary mill product size
to overall plant performance alone is not a trivial task. The poten-
-4
10 tial recovery to grind relationship for RPMA2 UG2 ore was obtained
from laboratory test work. Note that this grind represents the final
-5
10 grind and hence the final plant recovery by assuming perfect flota-
-5 -4 -3
10 10 10 tion operation and a linear recovery-grind relationship within the
Frequency (Hz) operating region. The next step is to obtain a model between pri-
mary and final grind. A linear model was used which resulted in
Fig. 7. The mill load frequency comparison between pre and post MPC as well as
an RMSE of 1.5% passing 75 lm. The RMSE of the model relates
the load’s theoretical minimum variance response of the pre and post MPC. The
theoretical variance remains similar, while the actual obtained variance is lower
to 20% of the range of the filtered final grind and is therefore con-
than the pre-MPC condition. sidered a poor model. Due to the model quality, the relationship
between primary and final grind remains an assumption rather
between the two algorithms which should lead to more favourable than a proven fact. The effect of primary grind G(L, IW) on potential
results during MPC operation. The feed-forward capability of the recovery PR can then be deduced by means of convolution:
MPC allowed for better disturbance rejection with the inclusion
of the ore feed size. The MPC uses future prediction to obtain a fea- PR ¼ 0:134  GðL; W rat Þ þ 79:61 ð14Þ
sible solution and to avoid possible constraints where the Neuro-
Fuzzy relies predominantly on feed-back (Jonas, 2008; Moyano with model boundaries:
et al., 2010). The minimisation of the effective power utilisation
objective is directly included in the MPC algorithm. It is certainly 185 < L < 230
possible to add the above mentioned functionality to the Neuro- 0:13 < W rat < 0:26
Fuzzy controller. However, due the poor scaling of Fuzzy control-
lers, this will probably result in rule explosion and substantial G(L, Wrat), obtained from the optimisation framework (Eq. (13)). The
engineering hours to design and maintain. profit PF in terms of the main process variables are defined as:

PðL; W rat Þ ¼ a  PR ð15Þ

4. Benefit analysis
with a = a1  F  HG.
The purpose of this analysis is to determine the financial bene- a1 is equal to the monetary value of a gram of PGMs in the final
fits achieved by stabilising the milling circuit at its optimum. The concentrate at a certain grade ci. F represents the amount of ore
financial benefit of the milling circuit will be determined using treated by the mill in ton per unit time and HG the head grade in
the performance function method devised by Wei and Craig g/t.
120 C.W. Steyn, C. Sandrock / Minerals Engineering 53 (2013) 113–123

Fig. 8. The mill load standard deviation comparison between pre and post MPC over time.

Fuzzy Control Commissioning MPC

FeedSize

Speed

ObjFunction

OreFeed

WaterRatio

Fig. 9. Normalised shiftly averages, time based plot of the objective function, manipulated variables and disturbance variable of the primary mill.

4.1.2. Cost function CðL; W rat Þ ¼ b  PðL; W rat Þ ð16Þ

As the purpose of this financial benefit model is to provide a
metric to compare various optimisation or control strategies, only Where b represents the cost of power in $/(kW h) and,
the operating cost will be included in the cost PF. Fixed costs are dP dP
considered to be constant. Energy and grinding media are the PðL; W rat Þ ¼ Lþ W rat ð17Þ
dL dW rat
two highest operating cost factors associated with comminution.
As the RPMA2 primary mill is a fully autogenous mill, the economic The effect of load on power P is described by the parabola obtained
effect of cost is defined only by the energy consumption of the mill. during trial 1 with dP
dL
¼ 2:39L þ 497:4 kW/ton (see Fig. 5). The inlet
Note that the cost of liners and the performance of the discharge water ratio to power relationship is obtained from the linear step-
sump are not included in the scope of this benefit analysis and response model where the steady-state gain represents
dP
are considered negligible. The cost function is defined as follows: dW rat
¼ 43 kW.
 C.W. Steyn, C. Sandrock / Minerals Engineering 53 (2013) 113–123 121

(a) (b)

3100

(c) (d) (e)

Power (kW)
2900

2700

2500
87 87.5 190 200 210 220 230
(f) (g) (h) 36

Grind (-75mic)
34

32

30

86.5 87 87.5190 200 210 220 230

Potential Recovery (%) Load

Fig. 10. Comparing the potential financial benefit of the benchmarked plant (red), benchmarked controller performance at the optimum load (green) and the performance of
the MPC at the load optimum (blue). (a) Normal distribution (ND) of potential recovery, (b) ND of load, (c) ND of power, (d) 2D projection of the potential net profit function,
(e) power to load curve, (f) ND of grind, (g) grind to potential recovery curve and (h) grind to load curve. The contours in d) represent 80% confidence intervals. (For
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

(a) (b)

2800
(c) (d) (e) Power (kW)
2600

2400

2200

(f) (g) (h) 36

Grind (-75mic)

34

32

30

86.5 86.9 87.3 0.15 0.19 0.23 0.27

Potential Recovery (%) Water Ratio

Fig. 11. Comparing the potential financial benefit of the benchmarked plant (red) and the inlet water ratio optimum (blue). (a) Normal distribution (ND) of potential recovery,
(b) ND of water ratio, (c) ND of power, (d) 2D projection of the potential net profit function, (e) power to inlet water ratio curve, (f) ND of grind, (g) grind to potential recovery
curve and (h) grind to water ratio curve. The contours in (d) represent 80% confidence intervals. (For interpretation of the references to colour in this figure legend, the reader
is referred to the web version of this article.)

4.1.3. Overall performance function PNP ¼ a  PRðL; W rat Þf ðL; W rat Þ þ b  PðL; W rat Þf ðL; W rat Þ ð18Þ
The potential nett profit (PNP) given as profit – cost are pre-
sented as:
122 C.W. Steyn, C. Sandrock / Minerals Engineering 53 (2013) 113–123

The probability density function (PDF) of the grind is defined by the 4.2.2. Financial benefit – Inlet water ratio
PDF of the load and inlet water ratio which relationship was found Since no significant difference in water ratio variability was ob-
to be quadratic (Eq. (13)). served (Wrat standard deviation difference of 0.001 is considered
negligible), the financial benefits for water will be estimated by
investigating two scenarios both at the optimum load of 213 ton:
4.2. Comparing financial performance
1. At the benchmarked inlet water ratio mean of 0.178.
One of the main objectives of this project is to determine the 2. The optimum inlet water ratio, Wrat = 0.224.
financial benefit of stabilising and optimising the primary milling
circuit. The PNP function indicates that only load and inlet water Eq. (19) is used to derive the distribution for power, grind and
ratio are necessary to determine financial impact. Comparing potential recovery from the distribution of inlet water ratio. A
financial performance in terms of minimising PGM losses in the grind increase of 1.3% passing 75 lm from 34.0% to 35.3% was ob-
concentrator, a1 is considered a function of basket price of refined served by increasing the inlet water to the optimum. This results in
PGMs, less the operating cost to treat the additional recovered a PR increase of 0.17% from 87.17% to 87.33% and an increase on
material in the subsequent smelting and refining processes. These the PNP curve as illustrated in Fig. 11.
costs are confidential. The financial benefit of recovering more The total grind increase observed by moving the load and water
PGMs far outweighs the cost of power, therefore a  105b. The ratio mean to 213 ton and 0.224 respectively and stabilising load at
highest financial benefit is achieved at the maximum grind and a standard deviation of 2.3 ton was calculated at 2.2% passing
hence potential recovery. 75 lm. This equates to a potential recovery increase of 0.32%.
The financial benefit is derived by considering the impact of Recovery increases of smaller than 1% are very difficult to prove
load and inlet water ratio separately: on an industrial site. The benefit calculated from this project re-
mains therefore a theoretical construct, which is subject to various
assumptions.
4.2.1. Financial benefit – Load
As mentioned, the standard deviation of the load was reduced 4.3. Benefit of variable speed milling (VSD)
from 6.8 ton to 2.3 ton with the implementation of the MPC. The
financial impact of this project in terms of load is estimated by The majority of the primary mills in Anglo American Platinum
comparing the following three scenarios at the optimum water ra- do not have a VSD and hence only have two degrees of freedom
tio Wrat = 0.224: for control, they are, ore and water feed rate. Since ore feed rate
is expected to be at the target throughput, the water is utilised
1. The benchmarked plant with load at a mean of 204 ton and more often to reject any system disturbances. The additional de-
standard deviation of 6.8 ton. gree of freedom made possible by the mill VSD allows the control-
2. The benchmarked stability at the optimum load, a mean of 213 ler more leverage to reject disturbances while maintaining
ton and at deviation of 6.8 ton. volumetric fill and viscosity at the respective optimum. The cost
3. MPC performance at the optimum load, a mean of 213 ton and a of deviating from the optimum during upset conditions can be se-
deviation of 2.3 ton. vere, as was explained in Section 4.2. It is therefore important to
appreciate that the potential benefit is as a result of the VSD.
To derive the distribution of the grind, potential recovery and
power from the distribution of load, the change of variables for dis-
tributions equation is used. The distribution of variable Y = g(X), X 5. Conclusions
with probability density function fX(X), is calculated as follows:
5.1. Control
 
d 
fY ðyÞ ¼  ðg 1 ðyÞÞ:fX ðg 1 ðyÞÞ ð19Þ A model predictive controller that utilises the step-response
dy
models was implemented and commissioned. The objective func-
tion of the controller was designed to optimise effective power
where g1 denotes the inverse function. utilisation. The controller improved the circuit stability by reduc-
A grind increase of 0.56% passing 75 lm or a potential recovery ing the standard deviation of the mill load and power by 66%
increase of 0.1% from 87.2% to 87.3% (Eq. (14)) was observed by and 40% respectively. The controller also reduced its objective
moving the load mean from 204 ton to the optimum 213 ton (com- function, effective power utilisation by 11%, from 25.2 kW h/t pass-
paring scenarios 1 and 2). A further 0.35% passing 75 lm grind or ing 75 lm to 22.4 kW h/t passing 75 lm.
0.05% PR increase was observed when reducing the load variance
around the optimum from 6.8 ton to 2.3 ton (comparing scenarios
2 and 3). Fig. 10 illustrates how the financial impact manifests by 5.2. Benefit analysis
reducing the operating region of the mill and then moving it higher
up on the PNP surface. The impact of this AG mill optimisation project was to improve
In order to interpret the benefits analysis plots in Figs. 10 and the grinding efficiency of the mill, which, resulted in a theoretical
11, it is important to note that due to the high amount of noise increase in overall the PGM recovery of 0.32%. This was achieved
in the data, the correlations presented in these plots could not be by operating the mill at the optimum volumetric fill and viscosity
considered exact. The distributions of the parameters power, po- for a given throughput, made possible by the presence of a VSD:
tential recovery and grind were thus assumed to be independent.
The contours of the various scenarios on all the power to potential  moving the load from an average of 204 ton to 213 ton (contrib-
recovery plots, represents the 80% probability of the relative distri- uted 26% of total benefit);
bution. A one standard deviation distribution of the exact model  stabilising the load at the optimum grind by reducing the stan-
value is presented as a dotted line in the colour of the respected dard deviation from 6.8 ton to 2.3 ton (contributed 16% of total
scenario on the power to recovery contour plot. benefit);
 C.W. Steyn, C. Sandrock / Minerals Engineering 53 (2013) 113–123 123

 moving the inlet water ratio average from 0.178 to 0.224 (con- Automining 2010: 2nd International Congress on Automation in the Mining
Industry.
tributed 58% of total benefit).
Napier-Munn, T., Morell, S., Morrison, R., Kojovic, T., 1999. Mineral Comminution
Circuits – Their Operation and Optimization, second edition. Julius Kruttschnitt
Mineral Research Centre, The University of Queensland.
References Powell, M., Mainza, A., 2006. Extended grinding curves are essential to the
comparison of milling performance. Minerals Engineering 19 (15), 1487–1494.
Rogers, M., Almond, D., Maru, T., Becerra, K., 2010. Model predictive control
AspenTech, 2000. AspenTech pty. ltd. DMCplus Course Notes. application at Lumwana copper concentrator. In: Babarovich, V. (Ed.),
Camacho, E.F., Bordons, C.A., 2007. Model Predictive Control, second ed. Springer. Automining 2010: 2nd International Congress on Automation in the Mining
Craig, I., Koch, I., 2003. Experimental design for the economic performance Industry.
evaluation of industrial controllers. Control Engineering Practice 11 (1), 57–66. Skogestad, S., Postlethwaite, I., 2005. Multivariable Feedback Control: Analysis and
Hodouin, D., Jms-Jounela, S.L., Carvalho, M.T., Bergh, L., 2001. State of the art and Design, second ed. Wiley-Interscience.
challenges in mineral processing control. Control Engineering Practice 9 (9), Snyman, J.A., 2005. Practical Mathematical Optimization: An Introduction to Basic
995–1005. Optimization Theory and Classical and New Gradient-Based Algorithms.
Hulbert, D., 2002. Evaluation of the economic benefit of milling circuit control. In: Springer.
Proceedings of the 15th IFAC Triennial World Congress. Elsevier Science Sosa-Blanco, C., Hodouin, D., Bazin, C., Lara-Valenzuela, C., Salazar, J., 2000.
Barcalona, Spain. Economic optimisation of a flotation plant through grinding circuit tuning.
Jonas, K., 2008. Advanced control for mineral processing: Better than expert Minerals Engineering 13 (10–11), 999–1018.
systems. In: Orford, I. (Ed.), 40th Annual Meeting of the Canadian Mineral Steyn, C., Brooks, K., de Villiers, P.G.R., Muller, D., Humphries, G., 2010. A holistic
Processors. pp. 421–443. approach to control and optimization of an industrial run-of-mine ball milling
Klimpel, R.R., 1983. Slurry rheology influence on the performance of mineral/coal circuit. In: Gorden, J. (Ed.), IFAC MMM Conference 2010. pp. 137–141.
grinding circuits. Part II. Minerals Engineering 35, 21–26. van der Westhuizen, A.P.P., Powell, M.S., 2006. Milling curves as a tool for
Morell, S., Kojovic, T., 1996. The influence of slurry transport on the power draw of characterising sag mill performance. In: SAG Conference 2006 – Part 2.
autogenous and semiautogenous mills. In: Proc SAG ’96. Vancouver, pp. 373– University of British Columbia, Dept of Mining and Mineral Process Engineering.
389. Wei, D., Craig, I.K., 2009. Grinding mill circuits – a survey of control and economic
Moyano, G., Viveros, P., Cortes, G., 2010. Sag and secondary grinding multivariable concerns. International Journal of Mineral Processing 90 (1–4), 56–66.
predictive control coordinated: division codelco norte. In: Babarovich, V. (Ed.),