Professional Documents
Culture Documents
A Stochastic Mixed Integer Programming Framework F
A Stochastic Mixed Integer Programming Framework F
fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2970480, IEEE Access
Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2020. Doi Number
ABSTRACT Conventional mine planning approaches use an estimated orebody model as input to generate
optimal production schedules. The smoothing effect of some geostatistical estimation methods cause most
of the mine plans and production forecasts to be unrealistic and incomplete. With the development of
simulation methods, the risks from grade uncertainty in ore reserves can be measured and managed through
a set of equally probable orebody realizations. In order to incorporate grade uncertainty into the strategic
mine plan, a stochastic mixed integer programming (SMIP) formulation is presented to optimize an
underground cut-and-fill mining production schedule. The objective function of the SMIP model is to
maximize the net present value (NPV) of the mining project and minimize the risk of deviation from the
production targets. To demonstrate the applicability of the SMIP model, a case study on a cut-and-fill
underground gold mining operation is implemented.
INDEX TERMS production schedule optimization, mining industry, stochastic mixed integer programming,
grade uncertainty, underground mining
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2970480, IEEE Access
deterministic formulations described above consider a single NPV when compared to the deterministic design. Leite and
estimated orebody model and are incapable of dealing with Dimitrakopoulos [17] tested a SIP formulation at a copper
in-situ variability of orebodies [11]. The deterministic mine to generate a 29% higher NPV in comparison with the
representation of orebodies smoothens the metal grades and conventional mine planning method. Benndorf and
do not adequately represent the deposit. Thus, resulting in Dimitrakopoulos [18] developed a SIP formulation to address
unrealistic mine plans and unlikely revenue forecasts. joint multi-element grade uncertainty and successfully
Drillhole data is generally made up of sparse and discrete applied it to an open-pit iron ore mine. The optimal schedule
point information, which is far from enough to determine demonstrated that the SIP model can generate a higher
accurate three-dimensional orebody information. The mineral probability in achieving production targets, which diminishes
grades of blocks in the block model are estimated in the the overall project risk and increases the project profit. Asad
unsampled regions based on these sample data. As the main and Dimitrakopoulos [19] also incorporated joint financial
input of mine optimization models, the uncertainty of grade and geological uncertainty in the design of open-pit mines.
estimation can affect the metal content of blocks as well as All the studies above illustrate the merits of using a
economic block values. Consequently, this will impact the stochastic approach to incorporate geological uncertainty in
production plan and NPV of the mining operation. These open-pit production scheduling optimization. However, little
limitations have motivated researchers to seek to develop work has been done in terms of incorporating geological
risk-based optimization frameworks for mine planning that uncertainty in underground mining production schedules,
can incorporate and manage geological uncertainty [12]. because of the complexities associated with the development
With the development of simulation methods, the risks from and implementation of underground mining methods. Based
geological uncertainty in ore reserves can be measured and on Ramazan and Dimitrakopoulos’s approach [16],
managed through a set of equally probable orebody Carpentier et al. [20] presented a SIP model with recourse
realizations. Dimitrakopoulos et al. [13] quantified the optimization formulation to generate the optimal long-term
geological uncertainty from grade, tonnage, and rock-type for schedule for an underground mining project extension while
a conventional mine production schedule using 50 equally considering geological uncertainty, economically minable
probable simulated realizations of the deposit. The lenses, and cut-off grade. However, this work was developed
probabilistic approach for risk analysis showed that the for a specific purpose and needed more generalization for
expected NPV only had a 5% chance of being realized and stope extraction sequencing. Petro [21] and Montiel et al. [22]
the median NPV from the stochastic simulations was 25% implemented mining complexes optimization comprising
lower than the scheduling result from the estimated orebody. multiple open-pits, underground operations, and processing
Although this method can be used for sensitivity analysis of destinations under geological uncertainty. Although some
the geological risk, it cannot be capable for integrating and efforts have been made in incorporating geological
evaluating the uncertainty and risk accurately. The earliest uncertainty in underground mining production scheduling,
attempt to develop a long-term mining schedule while the application of stochastic optimization is relatively recent
incorporating geological uncertainty was established and and the methods available in the literature remains to be
implemented in an open-pit mine [14]. The authors used a verified when applied to different types of deposits and
multistage optimization framework to obtain the most underground mining methods. The research question here is;
profitable mining sequence while compatible with orebody “can a practical mine plan be generated taking into
uncertainty. This approach made significant improvements, consideration grade uncertainty for underground cut-and-fill
yet it is a step-wise framework and does not integrate extraction with backfilling”?
geological uncertainty into the production scheduling There are currently limited production scheduling
optimization formulation directly. optimization models available for large scale underground
The next generation of stochastic optimization based on cut-and-fill mining. This research seeks to develop a risk-
stochastic integer programming (SIP), provides a more based optimization framework using stochastic mixed integer
robust solution in dealing with uncertainty. The SIP model programming (SMIP) that effectively integrates grade
was first introduced to integrate geological uncertainty and uncertainty into the optimization of long-term production
manage risk directly in the production scheduling scheduling in underground cut-and-fill mining. The objective
optimization by Ramazan and Dimitrakopoulos [11, 15]. function of the SMIP model is to maximize the net present
This approach maximizes the NPV and minimizes geological value and balance the risk associated with grade uncertainty
risk from production targets by penalizing the simulated simultaneously, while meeting all technical constraints and
realizations deviations in the objective function operational requirements.
simultaneously. Similar studies and applications have been
implemented along this line of thought. Ramazan and II. DEFINITIONS AND ASSUMPTIONS
Dimitrakopoulos [16] developed an efficient model based on Cut-and-fill mining provides better selectivity and is
two-stage SIP for optimizing the production schedule for a preferred for steeply dipping orebodies with an irregular
small gold deposit and showed a 10% increase in expected shape, discontinuities and dissemination. Together with
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2970480, IEEE Access
room-and-pillar mining, it is the most flexible of respectively, and one continuous variable to control the
underground mining methods. In cut-and-fill mining, there portion of a stope to be mined in one-time period. The
are horizontal development levels connected to the main advancement and mining activities are repeated until mining
shaft at regular vertical intervals along the orebody. The ore has progressed along the entire strike of the orebody on a
is extracted and removed from a stope. When a stope is level.
mined, the void is backfilled and allowed to cure before
Outside stope Inside stope
adjacent stopes could be mined with the backfill material
serving as a working platform or artificial pillar. A typical
layout of planned stopes to be developed and extracted on
selected levels is illustrated in Figure 1. Drilling Extraction Backfilling
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2970480, IEEE Access
advancement v; where I is the total number of advancement precisely, the first part (Part 1) of the objective function
activities in set Dv(I). refers to the total discounted cashflow from mining and
Ll(M) for each level l, there is a set Ll(M) defining all the processing ore material for all realizations. The second part
stopes on this level; where M is the total number of stopes in (Part 2) minimizes the costs associated with deviating from
set Ll(M). the ore grade operating targets for all realizations. The third
4) PARAMETERS part (Part 3) minimizes the total costs associated with
Rnt , s revenue, generated by selling the final product from deviating from the ore tonnage operating targets. Part 2 and 3
stope n in period t for realization s. aim to control grade uncertainty by minimizing the
g n, s average grade of ore portion of stope n for variability from the ore grade and ore tonnage targets.
realization s. The per-unit penalty costs (denoted by ctg and cto in
on the ore tonnage of stope n. Equation (1)) for the deviations are discounted in terms of
rt processing recovery; the proportion of mineral geological risk discounting (GRD) [23]. It can be used to
recovered in period t. control the grade and ore tonnage risk distribution over time.
pt selling price in terms of obtainable per unit of This further motivates the extraction of lower-risk and
product sold in period t. higher-grade stopes in the early periods and leaves the
CAvt total cost for advancement v in period t. higher-risk stopes for later periods. It should be noted that it
cavt cost per tonne for advancement v in period t. is crucial for the operation to mine less risky parts of the
adv the tonnage of advancement v. deposits in early periods with the limited geological
CEnt total cost in terms of mining for stope n in period t. information. As mining progresses, with more geological
cent cost per tonne in terms of mining for stope n in information available in the form of operational exploration,
period t. we can reproduce the schedule for the later periods. When
CPnt total cost in terms of processing ore from stope n in the GRD is set to zero, a balanced-risk profile is produced.
period t. The per-unit value of penalty costs is defined based on the
cpnt cost per tonne in terms of processing ore from stope impact of one unit of deviation from the related production
n in period t. target on the overall mining strategy. They are defined based
cgt , − per-unit penalty cost in terms of the lower grade on trial and error with consideration of the magnitude of
target deviation in period t. different production targets. Therefore, the per-unit value of
cgt , + per-unit penalty cost in terms of the upper grade the penalty can be altered to prioritize the most critical
target deviation in period t. operational targets according to the optimization
gDevst , − continuous variable; the shortfall from the grade requirements in the mining complex. Two binary decision
upper bound in period t for realization s. variables are used to model the advancement and extraction
gDevst , + continuous variable; the excess over the grade lower in each period, and a continuous decision variable is used to
bound in period t for realization s. model the portion of a stope which is mined in each period.
cot , − per-unit penalty cost for the lower ore tonnage The total revenue is defined in Equation (2) and the
target deviation in period t. advancement, mining, and processing costs are defined in
cot , + per-unit penalty cost for the upper ore tonnage Equations (3)-(5). By this uncertainty-integrated objective
target deviation in period t. function, the SMIP model can reduce the risk of not meeting
oDev−t continuous variable; the shortfall from the ore the planned production targets and provide a feasible
tonnage upper bound in period t. schedule.
oDev+t
1 S T V N ( Rn , s − CEn − CPn ) xn − CAv av
continuous variable; the excess over the ore tonnage t t t t t t
- Part 2
period t. (1+g )
t
S s =1 t =1
N let the maximum number of active levels available for
T c t oDev t + c t oDev t
mining in period t. - o , + + o,− −
Part 3
( )
t
t =1 1+ g
N ed the maximum length of mining duration for stopes.
i the financial discount rate. (1)
g the geological discount rate; used to control the risk
Rnt , s = g n , s on r t pt (2)
distribution over time.
CAvt = cavt ad v (3)
B. OBJECTIVE FUNCTION
The objective function of the SMIP model in Equation (1) is CEnt = cent on (4)
to maximize the expected NPV and at the same time
minimize the deviations from the production targets. More CPnt = cpnt on (5)
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2970480, IEEE Access
( ad ant ) au t t 1,..., T
v
(o xnt ) eu t
N
n t 1,..., T (7)
n =1
Current Predecessor
2) PROCESSING QUANTITY CONSTRAINTS advancement advancement
Equation (8) is a flexible constraint for the total tonnage
processed in each period and in each realization. The Figure 3. Schematic representation of advancements.
continuous variables oDevt+ and oDevt- are used as buffers to
allow deviations but these deviations are penalized in the t
objective function. Penalizing such deviations ensure that the avt − aik 0 i Dv ( I ) , v 1,...,V , t 1,...,T (10)
k =1
proportion of ore processed at the plant is within the required
targets as much as is feasible. put and plt are the upper and t
lower bounds for ore tonnage mined by the limiting ent − aik 0 i Dn ( I ) , n 1,..., N , t 1,...,T (11)
k =1
processing capacity in period t, respectively.
2) NON-ADJACENT STOPE PRODUCTION CONSTRAINT
N
( on xn ) − oDev+ pu t 1,..., T
t t t Non-adjacent stope production constraint is a geotechnical
n =1 constraint which focusses on limiting the size of unsupported
N (8)
( o xt ) + oDevt pl t
n n − t 1,..., T void areas. Figure 4 shows horizontal and vertical view of
n =1 nine adjacent stopes in two kinds of spatial constraints
required to ensure the geotechnical stability of the mine. If
3) GRADE BLENDING CONSTRAINTS
the central stope H5 is in the mining state, all mining
Equation (9) seeks to adjust the average grade of ore
activities are forbidden in the adjacent stopes (H2, H4, H6, H8)
production within the desired range in each period and in
(Figure 4 (1)). Similarly, if the central stope V5 is in the
each realization. Similar to the processing quantity
mining state, all mining activities are forbidden in the
constraints, the excess and shortage of the desired grade
adjacent stopes (V2, V4, V6, V8) (Figure 4 (2)). Equation (12)
range in each period across all geological realizations are
ensures that all forms of simultaneous adjacent stope
penalized in the objective function. gut and glt are the
activities are avoided to prevent excessively large
respective upper and lower bounds for the desired grade
unsupported voids.
range in period t.
N
( g n , s -gu ) on xn − gDevs , + 0 t 1,...,T , s 1,..., S
t t t V1 V2 V3
H1 H2 H3
n =1 H4 H5 H6
N H7 H8 H9 V4 V5 V6 Unmined
( gl t -g ) o x t − gDev t 0
n =1
n, s n n s,− t 1,...,T , s 1,..., S Mining
Forbidden
V7 V8 V9
(9) (1) (2)
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2970480, IEEE Access
These constraints are typically used in a cut-and-fill mining A. IMPLEMENTATION OF THE SMIP MODEL
scheduling model. Active levels control constraints take into The proposed model is implemented on an underground cut-
account the equipment mobility and access required for the and-fill gold mining operation. The deposit orientation is
underground mine production scheduling optimization N45°E, dipping to SE at an angle between 30°~50°. The
framework. As we all know, a large number of active levels main metallic minerals in the deposit are pyrite, electrum and
might lead to serious operational problems. Therefore, a native gold. The geological database consists of 490 drill
defined maximum number of simultaneous active levels holes with 13967 composites covering an area of 1300 by
allow mine planners to adjust production strategies to enable 600 m2, approximately. Using the available geological
different application scenarios depending on available information, one mineralized domain is defined and modeled.
infrastructure, production rates, equipment capacities, and the The deposit is estimated to contain 32 million tonnes of ore,
stability of the surrounding rocks. In addition, the active with an average grade of about 1.8 g/t Au. Figure 5 shows
levels control can also be considered as indirect geotechnical the schematic orientation of the orebody. This case study area
constraints. Equations (13) and (14) control the maximum contains 200 stopes distributed on 8 levels. The size of a
number of simultaneous active levels in each period of the stope is 50 m in length, 40 m in height, and the width follows
schedule. In general, the active levels control constraints help that of the orebody ranging from 10 to 30 m.
the optimization model to generate a more operationally
feasible mining pattern, indirectly leading to a practical
maximum NPV.
t
1 M
M x
k =1 m =1
k
m − lelk 0
m Ll ( M ) , l 1,..., L , t 1,...,T
(13)
L
le
l =1
t
l N let t 1,..., T (14)
x t
n 1 n 1,..., N (17) outline the robustness and impact of the risk-based mine plan
t =1 generated by the SMIP model. The two case study scenarios
T are implemented as follows:
e
t =1
t
n N ed n 1,..., N (18) (1) The stochastic optimization method (SMIP): a
stochastic mine production schedule is generated based on
fifteen orebody realizations using the SMIP model with
IV. CASE STUDY
reasonable technical, economic and risk-control parameters.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2970480, IEEE Access
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2970480, IEEE Access
targets will have a strong likelihood of being realized in the C. ANALYSIS OF THE OPTIMIZATION RESULTS:
early periods depending on the geological information CONTROLLING GEOLOGICAL RISK
available now. To reduce this risk and obtain more accurate Besides providing a reasonable solution and risk analysis for
forecasts for the later periods, new drill hole data and the mining operation, a major contribution of the SMIP
exploration information need to be updated over time and the model is the ability to control risk. To demonstrate its
schedules generated again. effectiveness, another fifty realizations of the deposit were
generated. Using different realizations for both the
optimization and the risk assessment provides a better
representation of the risk resilience of the SMIP schedule.
As we all know, the actual value of the deposit will never
be known except at locations where data are collected in the
future. Conditional simulation provides a model of
uncertainty by a set of equally probable simulated
realizations. In other words, each realization has an equal
probability to be the actual orebody. A comparison between
grade and NPV is described by randomly selecting three
different simulation realizations; the 10th, 25th, and 40th
realizations. The orebody model is replaced with these three
realizations and the stochastic and conventional schedules are
run with the same mining and economic parameters. Figure
Figure 6. Ore tonnage mined in SMIP and CM schedules over 18 periods. 10, Figure 11 and Figure 12 show the average ore grade (left)
and cumulative NPV (right) of the stochastic schedule (red)
and the conventional schedule (gray) for three different
realizations, respectively. For these representative
realizations, the expected ore grade in the stochastic results
are more compatible within the desired range and show less
fluctuations when compared to the conventional results.
Evidently, these merits can be attributed to incorporating in-
situ variability and uncertainty into the optimization
formulation. More importantly, from the cumulative cash
flow comparison from the three realizations, the expected
profit shows a 2-9% improvement in NPV from the SMIP
schedule when compared with the CM schedule (Table 3).
Even though the total metal content extracted by the SMIP
schedule is less, the higher NPV is as a result of the SMIP
model scheduling consistently higher grades earlier in the
Figure 7. Average ore grade processed in SMIP and CM schedules over mine life compared to the CM model.
18 periods.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2970480, IEEE Access
Figure 9. Risk profiles for cumulative cash flow for SMIP and CM schedules.
Figure 10. Average grade and cumulative NPV from the stochastic schedule (red) and the conventional schedule (gray) for 10th realization.
Figure 11. Average grade and cumulative NPV from the stochastic schedule (red) and the conventional schedule (gray) for 25th realization.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2970480, IEEE Access
Figure 12. Average grade and cumulative NPV from the stochastic schedule (red) and the conventional schedule (gray) for 40th realization.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2970480, IEEE Access
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2970480, IEEE Access
SHUWEI HUANG was born in Hebei, China. NAILIAN HU is currently a Professor with the
He received the B.S. and M.S. degrees in mining University of Science and Technology Beijing.
engineering from the University of Science and His work has been funded by many grants,
Technology Beijing (USTB). He is currently including the National Science Foundation of
pursuing the Ph.D. degree in mining system China and the Ministry of Science and
engineering with the School of Civil and Technology of the People's Republic of China.
Resource Engineering, University of Science and His research interests include mining system
Technology Beijing, and Laurentian University. engineering, mining economics, intelligent
His research interests include mining system mining, and mining safety. He is a member of a
engineering, mining economics, and minig council of the China Gold Association.
optimization.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.