You are on page 1of 10

March/April 2006 1

INTRODUCTION
Grade uncertainty and in situ variability continue
to be major obstacles for optimizing open pit mine
design, as shown in recent studies. Dimitrakopoulos
et al. (2002) and Dimitrakopoulos (2005) show that
there are substantial conceptual and economic differ-
ences between risk-based frameworks and traditional
approaches. Ramazan and Dimitrakopoulos (2004a)
show that the scheduling patterns generated by con-
ventional mixed integer programming (MIP) models
that use multiple simulated orebody input models
may be signicantly different from each other, as well
as from the scheduling pattern generated by using an
estimated input model. These signicant variations in
the scheduling patterns suggest that a new mathe-
matical scheduling method is neededone that can
account for orebody variability in optimization.
Past efforts to deal with uncertainty attempt to
sequentially link stochastic orebody models with con-
ventional optimization formulations (Ravenscroft,
1992; Dowd, 1997). This sequential process is ineffi-
cient and, although it allows the assessment of risk in
a schedule, it does not produce optimal scheduling
solutions in the presence of uncertainty as shown in
Dimitrakopoulos et al. (2002). In addition, these
efforts do not consider multi-element deposits that
have complex ore quality constraints, such as nickel
laterites, iron ore, or magnesium deposits. Further-
more, dealing with orebody uncertainty and in situ
grade variability accentuates the need to consider
issues of equipment access and mobility on the
related stochastic optimization formulations. The
Probabilistic practical optimization of production
scheduling for multi-element deposits
S. Ramazan, Rio Tinto, Perth, Western Australia, Australia, and
R. Dimitrakopoulos, McGill University, Montreal, Quebec
ABSTRACT Optimizing production scheduling in open pit mines is important for managing cash
ows of mining operations. This paper describes a new, probabilistic, mixed integer programming
(MIP) formulation which has been developed to minimize the risk of not achieving planned pro-
duction targets, in terms of metal production and ore tonnage feed for the mill during production,
while considering maximization of total discounted economic value. The new MIP model also rec-
ognizes that to minimize movement of excavators over long distances and accommodate equip-
ment access, the blocks scheduled for mining within a period cannot be spread too widely. The
model is applied to a laterite nickel-cobalt deposit, as an example of multi-element deposits that
have complex ore quality constraints. The results show that the proposed model generates feasible
schedules. When compared with traditional optimization techniques, the new model is superior in
terms of minimizing the risk of not meeting production targets, especially at the early stages of pro-
duction, and in terms of generating practical scheduling patterns.
KEYWORDS Probabilistic schedule, Uncertainty, Mathematical programming, Mine planning
number of binary variables required in mathematical
modelling for open pit mine planning (Ramazan,
2001; Ramazan and Dimitrakopoulos, 2004b) is also
one of the major obstacles for application of MIP-type
models for large open pit mines. Godoy and Dimi-
trakopoulos (2004) developed a long-term produc-
tion scheduling model that is based on simulated
annealing, is robust to risk, and is focused on meet-
ing the optimized ore production targets to maximize
the total achievable net present value (NPV) of the
project, using multiple simulated orebody models to
deal with grade uncertainty. Dimitrakopoulos and
Ramazan (2004) developed a long-term production
scheduling method that is based on a linear pro-
gramming (LP) model. The LP formulations eliminate
partial block mining problem in using linear variables
instead of binaries to improve the efficiency in solu-
tion time, and the model also considers orebody
uncertainty integrating the issue of equipment access
and reduced movement of large equipment. They
proposed the use of grade probabilities integrated
with a parameter that discounts orebody risk for
meeting production targets. Two windows generating
two sets of constraints for each block in the model
are implemented for handling equipment access and
mobility, but these multiple windows and constraints
increase the size of the models. In addition, the
reader is referred to recent developments and prac-
tices in the industry, presented in Dimitrakopoulos
(2005).
This paper presents an application of a new,
probabilistic, production scheduling formulation for
complex, multi-element deposits. The formulation is
Paper 6 Metal Mining
based on expected block grades and block economic
values that are weighted by probabilities of the blocks
containing certain amounts of metal and ore, and of
ore quality and grades being above required cutoffs
or within dened intervals that are derived from
jointly simulated ore deposit models (Dimitrakopou-
los, 2004). The block economic values weighted by
the probabilities referred to as probabilistic values in
this paper. Expected economic values, block grades,
and probabilities are integrated with equipment con-
straints and the practical feasibility of mining
sequencing in an MIP model. A key effect of such a
probabilistic approach is that the relatively more cer-
tain (rich) areas of the deposit are mined in earlier
production periods, leaving uncertain (poor) areas
for later periods, when additional information usually
becomes available. Application of only one set of con-
straints to integrate the equipment access and mobil-
ity improves the efficiency of the model. Furthermore,
integrating probabilities with economic values in the
objective function eliminates the probability target
constraints and aims to maximize the combined value
of discounted economic value and the probabilities.
In the following sections, the new, probabilistic
production scheduling method based on an MIP for-
mulation is rst presented, and then combined with
equipment and other constraints, to generate practi-
cal scheduling patterns. Subsequently, the formula-
tion is applied to a nickel-cobalt laterite deposit,
elucidating the practical aspects of the formulation.
Next, the practical differences between this approach
and the traditional scheduling approach are dis-
cussed. Lastly, the conclusions of this study follow.
A PROBABILISTIC MIP MODEL FOR PRACTICALLY
FEASIBLE PRODUCTION SCHEDULING
In the developed MIP-based scheduling model
that has been developed, a probability is assigned to
each block to represent the desirability of that
block being mined in a given period (Dimitrakopoulos
and Ramazan, 2004). This probability represents the
chances that a block will contain the desired quantity
of metal, quality and quantity of ore, ore grades
being above given cutoffs, and deleterious elements
being within required ranges.
The probability is calculated
from jointly simulated orebody
models representing the min-
eral deposit (eg. Dimitrakopou-
los and Fonseca, 2003; Boucher
and Dimitrakopoulos, 2005) as
further discussed in a subse-
quent section. The probability is
multiplied by the economic
value, which is referred to as
probabilistic value in the objec-
tive function of the model. The
orebody model used in the MIP
formulation is obtained by aver-
aging the values in the jointly
simulated orebody models that
are equi-probable representations of the actual ore-
body deposit. The model contains an objective func-
tion and a set of constraints that are discussed next.
OBJECTIVE FUNCTION
The objective function formulation is:
T N M
[ (k
n
*p
n
)
t
b
n
t
wd
t
m
] (1)
t=1 n=1 m=1
where T is the total time period for scheduling; N is
the total number of blocks in the model; k
n
is the eco-
nomic value to be generated by mining block n in
period t; p
n
is the probability of block n to have hav-
ing the desired metal quantity, and quality and quan-
tity of ore grades, including the grades above given
cutoffs or within ranges; (k
n
*p
n
)
t
is the discounted
probabilistic value to be generated by mining block n
in period t; b
n
t
is a binary decision variable, which is
set to 1 if the block n is mined in period t, and to 0 if
it is not mined; w is the cost of a unit deviation from
generating a smooth scheduling pattern as discussed
in the next paragraph; d
m
t
is the deviation from a
smooth pattern when mining block m, as calculated
in Equation 2; and M is the total number of blocks
centred on block m within the large window shown
in Figure 1.
A smooth scheduling pattern can be dened as the
pattern generated by assigning the same periods to
several blocks that are adjacent to each other to mini-
mize the equipment movement of equipment during
mining operations. The model in Equation 1 requires a
suitable cost coefficient (w) for deviations from a
smooth schedule, and it this coefficient is derived
through a trial and error approach, as follows. Initially,
a low cost penalizes the deviation from the smoothness
in Equation 1, to result in a widely spread mining pat-
tern for various periods, while the probabilistic value
will be expected to be relatively high in the rst period.
Incremental cost increases in the model will lead to a
required mining width and suitable equipment access in
the schedule. This process of repeating the scheduling
process at least a few times for the purpose of under-
standing what scheduling
parameters should be used is the
common approach taken with
any scheduling model, so as to
identify the most suitable param-
eters.
The schedule generated by
the MIP model is considered as
optimum for maximizing the dis-
counted probabilistic value
(k
n
*p
n
), given the degree of the
smoothness or feasible scheduling
patterns obtained. Discounting
the probabilistic value, instead of
discounting economic values rst
and then multiplying by probabili-
ties, makes identies the blocks
Metal Mining
Fig. 1. The blocks surrounding block m. Solid lines inside
show block m and adjacent blocks.
2 CIM Bulletin I Vol. 99, N 1093
March/April 2006 3
with higher probabilities as those better mined in the ear-
lier early periods than in late periods. The orebody risk
discounting concept introduced in Dimitrakopoulos and
Ramazan (2004) is implemented here in the same way as
the economic discount rate. Because the economic val-
ues are incorporated in the probabilistic value parameters
k
n
and p
n
, the economic discount rate is used to consider
the effect of time on the probabilistic value, instead of
using the orebody risk discounting rate for the probabil-
ity values. Because the economic value and the probabil-
ities are combined, the MIP model may tolerate some risk
on the schedule if it would signicantly increase the dis-
counted economic value. Due to the orebody risk dis-
counting effect, the low-grade and high-risk areas of the
deposit will be scheduled for later periods of production.
The objective function in Equation 1 maximizes
the net present value (NPV), conditional to on the
generated higher probabilities of meeting the pro-
duction targets at the early periods, and on the feasi-
ble scheduling patterns that are decided by the
decision-maker changing the cost parameter (w) as
discussed above.
CONSTRAINTS FOR EQUIPMENT ACCESS AND
MOBILITY
Constraints for mining the block m at period t
are:
B

e
j
(b
t
m
b
t
j
) d
t
m
0 (2)
j
where j is the counter of the blocks surrounding
block m (shown in Fig. 1); B is the total number of
surrounding blocks; e
j
is a weighting factor used to
discriminate the adjacent and non-adjacent blocks
within the window, with e
j
being set to 2 if block j is
adjacent to block m and to 1 if it is not adjacent to
m; and d
m
t
is the deviation from smoothness for
block m.
For example, if all the surrounding blocks other
than one adjacent block are mined when block m is
mined in period t, the deviation, d
m
t
, will be equal to
e
j
, which is 2. On the other hand, if a non-adjacent
block is not mined, d
m
t
will be 1, which incurs half the
cost of the case of not mining an adjacent block for
the objective function.
In this new model, the windows centred on m is
not moved over all the blocks but only over some of the
blocks, to thereby eliminating most of the constraints
from the model proposed by Dimitrakopoulos and
Ramazan (2004). If the window is moved over all the
blocks, the MIP model becomes too large and a feasi-
ble solution may not be obtained. Initially, the central
window is put over the block on the second row and
second column. The window is then moved on the
same row to the fth column, which means two blocks
are skipped. The window is moved until it reaches the
end of the row by skipping two blocks at a time and
setting the center central block on the third one. If the
central block comes over a block on the edge, that is
part of the window is outside the orebody model, the
window is ignored and constraints are not written.
After one row is nished, two rows are skipped and the
central block is put on the block at the third row from
the previous row. If the initial window is set on the rst
or third row, the model would still function in the same
way to generate feasible mining patterns.
The more traditional constraints of reserve, grade
blending mill requirements, metal production, pro-
cessing input capacity, and mining capacity are pre-
sented next.
RESERVE CONSTRAINTS
The set of reserve constraints specifies the
requirement that all the available blocks be mined
during one of the periods. The following formulations
are written for each block.
T

b
t
n
=1 (3)
t=1
where n =1, 2, , N.
GRADE BLENDING CONSTRAINTS
Upper bound constraints require the average
grade of the material sent to the mill for each of the
periods, t, to be less than or equal to a certain value,
Gr
max
.
N
(

(Gr
n
-Gr
max
)*O
n
*b
t
n
)0 (4)
n=1
where Gr
n
is the average grade of block n, O
n
is the
total ore tonnage in block n, and t=1, 2, , T.
Lower bound constraints require the average
grade of the material sent to the mill for each of the
periods, t, to be greater than or equal to a certain
value, Gr
min
.
N
(

(Gr
n
-Gr
min
)*O
n
*b
t
n
)0 (5)
n=1
where t=1, 2, , T.
METAL PRODUCTION CONSTRAINTS
Metal production constraints may also be consid-
ered as the processing constraints which are the phys-
ical limitations on the output. The total amount of
metal to be produced in a period cannot be more
than the processors maximum output capacity, Q
max
.
N

q
n
*b
t
n
Q
max
(6)
n=1
Metal Mining
where q
n
is the amount of recoverable metal con-
tained in block n.
Lower bound constraints ensure a minimum
amount of metal production, Q
min
, for each period, t.
N

q
n
*b
t
n
Q
min
(7)
n=1
PROCESSING INPUT CAPACITY CONSTRAINTS
Processing is constrained by the maximum pro-
duction capacity of the plant (PCap
max
) and the mini-
mum production requirement (PCap
min
). These upper
and lower bounds are necessary to ensure a smooth
feed of ore to mill.
Upper bound constraints for each period:
N

O
n
*b
t
n
PCap
max
(8)
n=1
Lower bound constraints for each period:
N

O
n
*b
t
n
PCap
min
(9)
n=1
MINING CAPACITY CONSTRAINTS
Mining capacity constraints represent the actual
available equipment capacity (MCap
max
) during each
production period.
N

r
n
*b
t
n
PCap
max
(10)
n=1
where r
n
is the total tonnage (ore and waste) con-
tained in block n.
CASE STUDY IN A NI-CO LATERITE DEPOSIT
The steps followed in this the case study using a
multi-element orebody can be summarized as fol-
lows:
1. Provide jointly simulated models of the
deposit attributes of interestNi, Co, Mg, and Al
grades, volume of percent rock, and the thick-
nesses of two layers (LS and RS)identified as
possibly containing nickel, all at the required
scheduling block size (40 by 40 m
2
).
2. Assign a probability to each block for recover-
able Ni metal content between 153 and 225 tons, Mg
less than 4.7%, and Al less than 0.6% from the
jointly simulated models of the deposit prepared in
Step 1.
3. Generate the e-type orebody model by aver-
aging the undiscounted economic values and grades
in each scheduling block.
4. Schedule the orebody mining using the for-
mulations in equations 1 through 10.
5. Quantify the risk in the optimal production
schedule using the jointly simulated deposit models
of pertinent attributes in Step 1.
OREBODY MODEL
The geology of the orebody shows a layer of
waste material on top of limonite and saprolite layers,
here combined to a zone (LS) with rocky saprolite (RS)
below. Both LS and RS may contain high-grade nickel.
For classication of ore and waste, the cutoff grade is
set at 0.5% Ni. The deposit is characterized by seven
attributes: Ni, Co, Mg, and Al grades, volume of per-
cent rock (Vol%R), thickness of LS, and thickness of
RS. The main mineral considered for prot in this proj-
ect is nickel. Cobalt is a by-product with limited con-
tribution to overall mine cash ows. Magnesium and
aluminum are relevant to the acid consumption at the
processing plant and have a major inuence on pro-
cessing costs.
For the purpose of scheduling, the deposit is rep-
resented by 2,030 blocks, 40 by 40 m
2
along east-
west and north-south directions. An orebody model is
generated using the technique of joint conditional
simulation, as detailed below. The model comprises
total tonnage, economic value, % tons of the LS
layer, % tons of the RS layer at 2 mm, Ni, Co, Mg,
Al, volume % rock, and % total ore tons (% tons of
LS + % tons of RS at 2 mm) within each block.
The seven deposit attributes mentioned above
are simulated, jointly, using a 5 by 5 m
2
grid for 35
realizations. The 5 by 5 m
2
grids are then reblocked to
the 40 by 40 m
2
block size. Each set of joint simula-
tion models generated is equally likely to be the real
deposit, given the available information. The joint
conditional simulation of these attributes is based on
the so-called joint simulation with minimum/maxi-
mum autocorrelation factors. This is an approach that
spatially decorrelates the variables involved to non-
correlated factors. The independent factors are indi-
vidually simulated and back-transformed to the
conditional simulations of the correlated deposit
attributes that reproduce the cross-correlations and
individual correlations of the original variables (Des-
barats and Dimitrakopoulos, 2000). Simulated repre-
sentations of the orebody are used to generate
average block grades and probabilities of values of
different attributes to be within given ranges of inter-
est and as needed for the scheduling optimization
formulation in Equation 1.
LONG-TERM PRODUCTION SCHEDULING
This study considers that ore material sent to the
processing plant during each production period
should produce at least 110,000 tons Ni metal at
approximately 90% recovery, and that the maximum
capacity of the processing plant is 122,000 tons Ni
metal. The ore material processed during each period
is constrained to have a magnesium content of
around 4.50.2% and aluminum content within the
range of 0.55% to 0.60%. Total tonnage of ore to be
4 CIM Bulletin I Vol. 99, N 1093
Metal Mining
processed in a period is also constrained within the
range of 9.5 to 10 million tons.
The probability to be assigned to each block is
calculated based on Ni grade and ore quality param-
eters of Mg and Al content of each block, using the
simulated realizations. The probability is determined
for each block containing recoverable Ni between
153 tons and 225 tons, Mg less than 4.7%, and Al
less than 0.6%. The upper bound of the metal is used
to minimize the chance of sending more recoverable
metal than to the processor than it can produce in a
period. In this study, it is assumed that the upper
bounds on the Al and Mg are more critical than the
lower bounds in terms of producing the correct
blended ore tonnage. If the lower bounds were also
applied, the probabilities would result in very low val-
ues eliminating the impact of economic values on the
schedule. When the individual blocks are not likely to
have the full desired characteristics, it is necessary to
implement the critical bounds. The probability in this
model is viewed as a parameter that helps to achieve
the main objective. The high probabilities of the
blocks scheduled in the early periods of the mine
reects the higher chances of the periodical schedule
meeting the production criteria.
PROBABILISTIC PRODUCTION SCHEDULING
CPLEX software (CPLEX, 1998) is used to solve all
the MIP formulations in this study. The term gap is
often used as an indication of how good an MIP solu-
tion is in mathematical programming. The
gap in the CPLEX software is reported as
a percentage to indicate how close the
current feasible solution is to the optimal
solution. The gap is the difference
between the best possible solution
obtained without constraining the binary
variables as binary (that is 0 or 1 values
only) and the current best solution that
the solver has found by applying the
binary constraints. If binary variables are
not constrained as binary, they can be set
to any value, such as 0.8, and this solu-
tion is considered as the best linear solu-
tion. Although the best linear solution is
not feasible for the MIP problem, it gives
an indication that there may be one or
more solutions having a better objective
value than the solution found by the
solver currently. As the solver continues
solving the problem, another solution
with a better objective value may be if the
gap is greater than 0%, or may not be
found even if the gap is very high and
over 50%. Currently, the only certain
knowledge is that the solution for an MIP
problem is the optimal solution for that
specic problem when the gap is 0.
The NPV maximization objective used
in traditional mathematical mine opti-
mization models is integrated with the
equipment access and mobility-related formulations.
The schedule obtained using the MIP model whose
objective function is constructed as NPV maximization
and feasible mining patterns is referred to as the tra-
ditional schedule. Figure 2 shows direct output of the
MIP model (a) and the smoothed image of the MIP
scheduling pattern (b). Figure 2a shows that some of
the blocks are not feasible to minefeasibly mined
since because they do not have the equipment
access. This infeasibility of the blocks makes it neces-
sary to smooth the image of the scheduling pattern
as shown in Figure 2b. During the image smoothing
process, the blocks that are scheduled to be mined in
a period during which they do not have equipment
access are rescheduled according to the scheduling
periods of the neighbouring blocks. It may be argued
that by increasing the cost coefficient w for the devi-
ations from smooth mining, the image smoothing
processing may be avoided. Although the argument
is theoretically valid, increasing the coefficient w has
increased the solution time in this case. The schedul-
ing pattern given in Figure 2a is obtained from a solu-
tion that has around 2.1% gap. For a high coefficient,
such as 5,000, the program ran for almost 20 hours
of cpu time to reach approximately an 8% gap on a
PC (P4 1.4 GHz), and the improvement on the gap
was extremely slow (Ramazan and Dimitrakopoulos,
2004b). The scheduling pattern was spread over the
deposit that is caused by the gap in the MIP solution.
When high coefficients for the smoothness parame-
ters are used in the objective function, the scheduling
Metal Mining
Fig. 2. a) Scheduling pattern generated by traditional MIP optimization model (left); and b) smoothed image (right).
March/April 2006 5
pattern would be expected to be smoother if the
smaller gap could be reached. Among several trials,
the scheduling pattern shown in Figure 2a is chosen
as the most practical one.
Tables 1a and 1b provide the summary of the tra-
ditional schedule. The tables include ore and total
tonnages mined, undiscounted economic value (UEV)
and NPV, average grades, and the average probability
per scheduling period. Table 1a shows that there is
not a violation on of any constraints on the direct out-
put of the MIP, but there are some violations with
respect to the e-type orebody model on the ore ton-
nage constraints after the scheduling pattern is
smoothed, as given in Table 1b. The smoothed feasi-
ble schedule produces about 10.3 Mt of ore during
the rst year and over 9.2 Mt during the second year.
The schedule produces 121,880 of Ni metal in the
rst year and 107,000 in the second year.
The table shows that the generated probabilities
representing the desirability characteristics of the
blocks for being mined are varying almost randomly.
The probability is lowest in the rst period, 12.6%,
and highest in the second period, 14%. This is not
unexpected, in that the traditional schedule does not
consider the risk for not meeting production targets
during operation.
The production scheduling results obtained by
applying the optimization formulation in Equation 1,
with w set to 660, and constraints in equations 2 to
9 to the orebody of the Ni laterite deposit 1 are
shown in Figure 3 and summarized in Tables 2a and
2b. Figure 3a shows that the scheduling pattern
resulting from the proposed MIP model contains
some blocks that do not have the equipment access,
as is the case in the traditional schedule. The solution
of this formulation had about a 7.7% gap. The
smoothed image of the schedule in Figure 3b shows
that the schedule is feasibly mined in terms of practi-
cality of the resulting pattern.
The average probability (AP) of the blocks sched-
uled in a period can be considered as an indication of
how good the schedule is to meet the production tar-
gets for that period. It can be calculated as
Bm Bm
AP=P
n
*O
n
/ O
n
(11)
n=1 n=1
Metal Mining
Table 2b. Summary results of the schedule after smoothing
Periods Ore Tonnage Undiscounted Net Present Rec. Ni Ni Co Mg Al Probability
Economic Value Value
(years) (10
6
tonne) (10
6
tonne) ($ 10
6
) ($ 10
6
) (10
6
kg) (%) (%) (%) (%) (%)
1 9.77 15.92 13.55 12.55 113.94 1.30 0.092 4.34 0.55 15.9
2 9.62 15.93 13.50 11.57 112.77 1.30 0.091 4.54 0.59 12.0
3 9.53 15.60 12.67 10.05 111.03 1.29 0.089 4.61 0.60 11.7
Total/Average 28.91 47.45 39.72 34.17 337.74 1.30 0.090 4.50 0.58 13.2
Table 1a. Summary results of traditional MIP schedule
Periods Ore Tonnage Undiscounted Net Present Rec. Ni Ni Co Mg Al Probability
Economic Value Value
(years) (10
6
tonne) (10
6
tonne) ($ 10
6
) ($ 10
6
) (10
6
kg) (%) (%) (%) (%) (%)
1 9.86 16.65 14.23 13.18 116.72 1.32 0.088 4.38 0.59 12.7
2 9.53 15.84 13.43 11.51 111.02 1.29 0.089 4.69 0.55 13.5
3 9.52 14.96 12.06 9.57 110.00 1.28 0.094 4.43 0.60 13.4
Total/Average 28.91 47.45 39.72 34.26 337.74 1.30 0.090 4.50 0.58 13.2
Table 1b. Summary results of the schedule after smoothing
Periods Ore Tonnage Undiscounted Net Present Rec. Ni Ni Co Mg Al Probability
Economic Value Value
(years) (10
6
tonne) (10
6
tonne) ($ 10
6
) ($ 10
6
) (10
6
kg) (%) (%) (%) (%) (%)
1 10.26 17.31 14.66 13.58 121.88 1.32 0.089 4.37 0.59 12.6
2 9.21 15.36 13.03 11.17 107.41 1.29 0.089 4.68 0.54 14.0
3 9.43 14.78 12.02 9.54 108.45 1.28 0.093 4.45 0.60 13.0
Total/Average 28.91 47.45 39.72 34.29 337.74 1.30 0.090 4.50 0.58 13.2
Table 2a. Summary results of uncertainty-based schedule
Periods Ore Tonnage Undiscounted Net Present Rec. Ni Ni Co Mg Al Probability
Economic Value Value
(years) (10
6
tonne) (10
6
tonne) ($ 10
6
) ($ 10
6
) (10
6
kg) (%) (%) (%) (%) (%)
1 9.85 16.04 13.63 12.62 115.19 1.30 0.092 4.30 0.55 16.3
2 9.55 15.80 13.40 11.49 111.89 1.30 0.091 4.50 0.59 12.3
3 9.51 15.61 12.68 10.07 110.66 1.29 0.088 4.70 0.60 10.9
Total/Average 28.91 47.45 39.72 34.18 337.74 1.30 0.090 4.50 0.58 13.2
6 CIM Bulletin I Vol. 99, N 1093
March/April 2006 7
where Bm is the number of blocks to be mined dur-
ing a period; Pn is the probability of that block to
have desired characteristics; and O
n
is the tonnage of
ore within the block.
As shown in Table 2b, the schedule has the high-
est probability of achieving the desired properties of
the ore produced in the rst year (15.9%), a lower
probability in the second year (12%), and the lowest
in the last year (11.7%). This shows that the available
risk of not achieving the production targets is mini-
mized in the rst period of production. In the present
case study, an 8% discount rate is used to discount
the probabilistic value.
Table 2b shows that the MIP model is scheduling
ore tonnages varying between around 9.5 and 9.8 Mt
during production periods, which are within the
range of the model constraints of 9.5 to 10 Mt.
Although there are some differences between Table
2a and Table 2b, the output of the schedule in Table
2b is still within the model constraints with respect to
the e-type orebody model.
COMPARISON OF OPTIMAL SCHEDULES
BETWEEN PROBABILISTIC AND TRADITIONAL
APPROACHES
In this section, the production schedules gener-
ated in the previous section are compared. The tradi-
tional scheduling optimization does not include
probabilities, and the objective is constructed as max-
imization of the total NPV of the project while mini-
mizing practical infeasibilities to provide equipment
access for the blocks to be excavated and continuity
of the excavation process. In the probabilistic sched-
ule (PS), probabilities are integrated in the economic
values of the blocks to take account of the deposit
variability and to minimize the risk of not achieving
the planned objective during actual mine production.
In Table 1, the periodical average probabilities of
meeting the production targets are calculated for the
traditional schedule (TS) using the probabilities of the
blocks that are scheduled in the same period. It
should be noted that average probabilities are calcu-
lated by weighting them with the amount of ore ton-
nage contained in each block. Evidently, the effect of
not factoring risk in the scheduling optimization for-
mulation generates lowest probabilities, or highest
risk (12.6%) for meeting production targets in the
rst year. In the TS, the probability of achieving the
desired properties of the ore produced is highest in
the second year (14.0%), and lower again in the last
year (13.0%). The distribution of risk for not meeting
the planned production target is totally random in the
schedule, as expected since there is no parameter in
the scheduling model to affect the risk distribution.
This trend is usually not desirable, because it is
expected that more information will be available as a
result of experience gained with the deposit and
more samples obtained as the mining operation pro-
ceeds. Uncertainty in the riskier areas would therefore
be expected to decrease over time, enabling the deci-
sion-maker to improve decisions on short-term pro-
duction scheduling and blending processes in the
future. In addition, the objective is usually to secure
production characteristics at early stages of a project,
so as to secure cash ows and loan repayment, as
well as improve over-
all nancial aspects of
a project.
Figure 4 summa-
rizes the comparison
of the proposed prob-
abilistic schedule (PS)
and the TS, showing
the average deviations
per mining period
from expected opti-
mal ore production
targets, and the prob-
ability of deviations in
ore production per
mining period occur-
ring. The term devia-
tion is used herein to
refer to the amount of
production (ore or
grades) above or
below planned tar-
gets. Figure 4 shows
the deviations and the
probabilities that are
specic to ore produc-
tion. The values plot-
ted in the gure are
generated by calculat-
Metal Mining
Fig. 3. a) Scheduling pattern generated by probability-based MIP optimization model (left); and b) smoothed image
(right).
ing the deviations of each schedule with respect to
the 35 jointly simulated orebody models. The bars in
the gure shows the average of these deviations. The
lines in the gure are obtained by nding the ratio of
the number of models in which the ore tonnage con-
straints are violated to the total number of simulated
orebody models. In the rst year, the TS model has
91% chance of deviating by around 383,000 t on
average. However, the PS model has only 57%
chance of deviating by only about 85,000 t on aver-
age. Furthermore, In in the second period, the PS has
only 34% chance of deviating by around 91,000 t,
while the TS has 74% chance of deviating by almost
270,000 t. Note that the PS schedule does not con-
sider minimizing the risk of not meeting production
targets for the last period. The risk in the earlier peri-
ods is considered more costly for the objective func-
tion than the risks on in the later periods by
discounting the probabilistic value as shown in Equa-
tion 1.
Figure 5 shows the average deviations in produc-
tion as bars and also the probability of the average
deviation occurring as lines for the recoverable metal
production, which is generated in a similar way to Fig-
ure 4. In the rst period, the TS has about
66% probability of deviating by around
3,280 t on average, while the PS has 29%
chance of deviating by only approxi-
mately 191 t. In the second period, the TS
has 57% chance of deviating by a little
over 2,513 t, but the PS method has only
6% probability of deviating by 321 t in
nickel production.
It should be noted that the average
deviations and probability of deviating are
decreasing consistently in ore and nickel
production as production continues in the
TS. This trend of having a higher risk of
not achieving production targets in the
early years of mine life can cause serious
nancial problems for the mining com-
pany. Minimizing risk in the rst period
and having an increasing risk prole is a
desirable risk distribution considering the
fact that as the mining operation contin-
ues, more information is accumulated
and experience is gained on the deposit
that can signicantly reduce the future
risk.
There are no signicant deviations in
Mg and Al contents, which means that
periodical constraints on these factors are
not as tight as processing input capacity
and metal production constraints. The
proposed probabilistic MIP schedule per-
forms substantially better than the tradi-
tional schedule when comparing the
overall deviations in ore feed and metal
production requirements during the rst
two periods.
CONCLUSIONS
This paper has presented a new,
probabilistic optimization formulation for
long-term production scheduling in open
pit mines. It is particularly suitable for
complex, stratied multi-element orebod-
ies such as Ni nickel laterites and iron ore
mines. The mathematical programming
formulation integrates orebody uncer-
tainty in respect of grade, ore quality and
quantity, and risk quantication, as well
Metal Mining
Fig. 4. Average deviations in ore tonnes and the probability of deviations occurring for the two schedules: traditional
(TS) and probabilistic (PS).
Fig. 5. Average deviations in nickel production and the probability of deviations occurring for the two schedules:
traditional (TS) and probabilistic (PS).
8 CIM Bulletin I Vol. 99, N 1093
March/April 2006 9
as equipment access and mobility, and other typical
operational requirements.
A key part of the formulation is that the proba-
bilities of grades of different elements being above
relevant cutoffs, or within a given range, are inte-
grated in the optimization process. This provides the
opportunity to generate schedules that aim to reduce
risk at early production stages, when secure cash
ows are most critical; and later production periods
will benet from additional information that becomes
available as mining operations proceed.
To generate production schedules with feasible
mining patterns, the formulation is coupled with
equipment accessibility and mobility constraints,
which aim to minimize inefficiencies in the utilization
of mining equipment. Generating the double-
window-type model instead of single-window allows
the constraints to be established for just 1/9th of the
total blocks considered, that which is extremely cru-
cial for the efficiency of the MIP model.
The practical aspects of the risk-based approach
were shown in an application at a Ni laterite deposit.
Relevant attributes and input to the scheduling for-
mulation (Ni, Co, Mg, and Al grades, volume of per-
cent rock, and thickness of layers) were jointly
simulated, conditional to all available drilling informa-
tion. Thirty-ve equally possible simulated representa-
tions of the deposit were used to generate
probabilities and averages for the optimization, as well
as to quantify the risk in meeting production targets.
The comparisons of the results with traditional
long-term production scheduling based on NPV opti-
mization veries the expectations for the new risk-
based formulation that risk in meeting production
targets is minimized and is lower in the rst period
than the second, and so forth. In addition, the sched-
uling patterns generated from the proposed
approach are feasible and superior to those from the
traditional optimization.
Although the proposed approach is not set up to
explicitly maximize NPV, it generates a realistic NPV,
which is the best under the scheduling considera-
tions. It is obvious that NPV can be increased by forc-
ing the probabilistic MIP model to mine high-grade
blocks in the early periods through the use of high
and tight grade constraints in those periods. How-
ever, increasing NPV will generally increase the risk of
not meeting production targets. The traditional
model shown in this study produced 2% higher total
NPV due to the high cash ows in the rst scheduling
period of the model. However, the risk of not meet-
ing production targets in the rst period was about
6% higher than for the proposed risk-based MIP
model. There are also the practical mining issues
mentioned above to take into consideration.
The proposed method can work well in three-
dimensional deposits that have precedence con-
straints. Due to the precedence constraints, the
schedule may not need to apply the smoothness
parameters in three-dimensional deposits, because to
extract a block at certain depth all the overlying
blocks must be mined. Therefore, the scheduling pat-
tern is more likely to result in a mineable smooth
shape than a laterite-type deposit where there is no
method of managing the smoothness of the pattern
without the proposed smoothness formulations in
this paper. It is clear that having precedence con-
straints in the model doesnt prevent the scheduler to
use the proposed risk-managed objective in the
scheduling process.
Future work could consider probabilistic design
and scheduling in underground mines (e.g. Grieco
and Dimitrakopoulos, 2005) and a more direct inte-
gration of orebody uncertainty in production schedul-
ing formulations. Since the work presented herein,
direct integration of uncertainty and new formula-
tions for stochastic production scheduling can be
found such as in Godoy (2003), Dimitrakopoulos
(2004), Ramazan and Dimitrakopoulos (2005), and
Menabde et al. (2005).
Paper reviewed and approved for publication by the
Geological Society of CIM and the Metal Mining
Society of CIM.
Salih Ramazan is a specialist mine planning engineer with Rio
Tinto, Australia. He holds a M.Sc. in mining engineering
(optimization and open pit mine planning) and a Ph.D. in mining
engineering (optimization and mine planning) from the Colorado
School of Mines, United States, and a M.E. (geostatistics) from
cole des Mines de Paris, France.
Roussos Dimitrakopoulos holds the Canada Research Chair in
Sustainable Mineral Resource Development and Optimization
under Uncertainty and the BHP Billiton Chair in Mine Planning
Optimization at the Department of Mining, Metals, and Materials
Engineering, McGill University. He holds a Ph.D. in stochastic
modelling from cole Polytechnique, Montreal, and a M.Sc. from
the University of Alberta, Edmonton.
Metal Mining
REFERENCES
BOUCHER, A. and DIMITRAKOPOULOS, R., 2005. A new effi-
cient joint simulation framework and application in a multi-
variable deposit. Orebody Modelling and Strategic Mine
Planning, The Australasian Institute of Mining and Metallurgy,
Spectrum Series, 14, p. 303-312.
CPLEX, 1998. Using the CPLEX Callable Library, Including
Using the CPLEX Base System with CPLEX Barrier and Mixed
Integer Solver Options. ILOG, Inc., Incline Village, Nevada,
United States.
DESBARATS, A. and DIMITRAKOPOULOS, R., 2000. Geostatis-
tical simulation of regionalized pore size distributions using
min/max autocorrelation factors. Mathematical Geology, 30,
p. 919-942.
DIMITRAKOPOULOS, R., 2004. Orebody uncertainty, risk
assessment and protability in recoverable reserves, ore selec-
tion, and mine planning. Proceedings, Short Course, SME
Annual Meeting and Exhibit. The Society for Mining, Metal-
lurgy and Exploration, 355 p.
10 Copyright2006. All rights reserved. ISSN 1718-4169 CIM Bulletin I Vol. 99, N 1093
Metal Mining
REFERENCES
DIMITRAKOPOULOS, 2005. Orebody modeling and strategic
mine planning: Uncertainty and risk management models.
The Australasian Institute of Mining and Metallurgy, Spectrum
Series, 14, 376 p.
DIMITRAKOPOULOS, R. and FONSECA, M. B., 2003. Assess-
ing risk in grade-tonnage curves in a complex copper deposit,
northern Brazil, based on an efficient joint simulation of mul-
tiple correlated variables. Computer Applications in the Min-
eral Industries. Edited by F.A.C. Calzolari. South African
Institute of Mining and Metallurgy, p. 373-382.
DIMITRAKOPOULOS, R. and RAMAZAN, S., 2004. Uncertainty
based production scheduling in open pit mining. SME Trans-
actions, 316, p. 106-112.
DIMITRAKOPOULOS, R., FARRELLY, C., and GODOY, M.C.,
2002. Moving forward from traditional optimization: Grade
uncertainty and risk effects in open pit mine design. Transac-
tions of the Institute of Mining and Metallurgy, Section A:
Mining Industry, 111, p. A82-A89.
DOWD, P.A., 1997. Risk in minerals projects: Analysis, percep-
tion and management. Transactions of the Institute of Mining
and Metallurgy, Section A: Mining Industry, 106, p. A9-18.
GODOY, M.C., 2003. The Effective Management of Geologi-
cal Risk in Long-term Production Scheduling of Open Pit
Mines. Ph.D. thesis, Bryan Research Centre, The University of
Queensland, Brisbane, Queensland, 256 p.
GODOY, M.C. and DIMITRAKOPOULOS, R., 2004. Managing
risk and waste mining in long-term production scheduling.
SME Transactions, 316, p. 43-50.
GRIECO, N. and DIMITRAKOPOULOS, R., 2005. Grade uncer-
tainty in stope design: Improving the optimization process.
Orebody Modelling and Strategic Mine Planning. The Aus-
tralasian Institute of Mining and Metallurgy, Spectrum Series,
14, p. 249-256.
MENABDE, M., FROYLAND, G., STONE, P., and YEATES, G.,
2005. Mining schedule optimization for conditionally simu-
lated orebodies. Orebody Modelling and Strategic Mine Plan-
ning. The Australasian Institute of Mining and Metallurgy,
Spectrum Series, 14, p. 347-352.
RAMAZAN, S., 2001. Open Pit Mine Scheduling Based on
Fundamental Tree Algorithm. Ph.D. thesis, Mining Engineer-
ing Department, Colorado School of Mines, Golden, Col-
orado, 164 p.
RAMAZAN, S. and DIMITRAKOPOULOS, R., 2004a. Traditional
and new MIP models for production scheduling with in-situ
grade variability. International Journal of Surface Mining,
Reclamation and Environment, 18, p. 85-98.
RAMAZAN, S. and DIMITRAKOPOULOS, R., 2004b. Recent
applications of operations research in open pit mining. SME
Transactions, 316, p. 73-78.
RAMAZAN, S. and DIMITRAKOPOULOS, R., 2005. Stochastic
optimization of long-term production scheduling for open pit
mines with a new integer programming formulation. Orebody
Modelling and Strategic Mine Planning. The Australasian Insti-
tute of Mining and Metallurgy, Spectrum Series, 14, p. 353-
360.
RAVENSCROFT , P.J., 1992. Risk analysis for mine scheduling
by conditional simulation. Transactions of the Institute of Min-
ing and Metallurgy, Section A: Mining Industry, 101, p. A104-
A108.

You might also like