You are on page 1of 12

Research article Petroleum Geoscience

Published online October 8, 2018 https://doi.org/10.1144/petgeo2018-022 | Vol. 25 | 2019 | pp. 207–218

Efficient brownfield optimization of a reservoir in west Siberia


O. S. Ushmaev1, V. M. Babin1, N. G. Glavnov1, R. R. Yaubatyrov1,
D. Echeverría Ciaurri2*, M. V. Golitsyna3, A. V. Pozdneev3 & A. S. Semenikhin3
1
Gazpromneft STC, 75-79 liter D Moika River emb., St Petersburg 190000, Russia
2
IBM Thomas J. Watson Research Center, 1101 Kitchawan Road, Yorktown Heights, NY 10598, USA
3
IBM East Europe/Asia, 10 Presnenskaya emb., Moscow, 123112, Russia
R.R.Y., 0000-0002-6484-1699; D.E.C., 0000-0001-7168-4233
* Correspondence: detxebeciau@gmail.com

Abstract: In this work we present a methodology for optimal management of brownfields that is illustrated on a real field. The
approach does not depend on the particular reservoir flow simulator used, although streamline-derived information is leveraged
to accelerate the optimization. The method allows one to include (non-linear) constraints (e.g. a recovery factor larger than a
given baseline value), which are very often challenging to address with optimization tools. We rely on derivative-free
optimization coupled with the filter method for non-linear constraints, although the methodology can also be combined with
approaches that utilize exact/approximate gradients. Performance in terms of wall-clock time can be improved further if
distributed-computing resources are available (the method is amenable to parallel implementation). The methodology is
showcased using a real field in west Siberia where net present value (NPV) is maximized subject to a constraint for the recovery
factor. The optimization variables represent a discrete time series for well bottom-hole pressure over a fraction of the production
time frame. An increase in NPV of 7.9% is obtained with respect to an existing baseline. The optimization methods studied
include local optimization algorithms (e.g. generalized pattern search) and global search procedures (e.g. particle swarm
optimization). The controls for one injection well in the real field were actually modified according to the solution determined in
this work. The results obtained suggest improvement for most economic scenarios.
Received 22 February 2018; revised 14 May 2018; accepted 2 August 2018

The importance of brownfields resurges in periods of relatively low injection rate). The cost function in these optimization problems
values of oil price. The management of this kind of field is principally has been often observed to be relatively smooth (unlike in well
based on the periodical setting of well controls (this may also include location optimization problems, where the cost function may be
shutting wells in and switching their functionality, e.g. from being a markedly non-linear due to significant spatial heterogeneities in the
production well to an injection well). Expensive capital investments reservoir rock properties), and, in cases of moderate or high number
such as the drilling of new wells are not contemplated in brownfields of variables, different optimized solutions with similar cost-function
very often, and, for this reason, these fields can be a very attractive values can be found (e.g. see Echeverría Ciaurri et al. 2011a; Van
component of a field portfolio of an oil and gas company. Essen et al. 2011). Consistent with that, the optimization techniques
The modification of well settings in brownfields can be an arduous of choice in the context of brownfield management are in most cases
task that cannot be performed with arbitrary frequency. Thus, it can of a local nature (in contrast to global search procedures that may be
be very important to elaborate optimized schedules of well settings used, at the expense of additional computation cost, in those cases
that lead to efficient production of the field. However, this can be a where the optimization cost function might present a higher degree
rather difficult enterprise because of the frequently large number of of non-linearity than expected; global search methods are less
control variables associated with the wells in a brownfield, the frequently employed in production optimization than local algo-
elevated computational cost of the numerical models used (which can rithms; the reader may consult Harding et al. 1998; Echeverría
render optimization problems challenging even for a moderate Ciaurri et al. 2011a for examples of applications of global search
number of variables), and the presence of multiple and complicated methods in production optimization).
operational constraints (finding feasible controls that satisfy all of Whenever derivative information can be computed efficiently,
these constraints is usually not trivial). On many occasions these local optimization is addressed by means of gradient-based
schedules are determined using available rules of thumb and educated optimization methods (e.g. see Nocedal & Wright 2006). Adjoint
guesses, which, due to their heuristic foundations, may lack accuracy formulations (e.g. see Sarma et al. 2008; Van Essen et al. 2010) can
and robustness, and even fail and not provide acceptable solutions. be considered in simulation-based optimization for a rapid gradient
Therefore, in search of techniques more solid than ad hoc computation of the cost function, provided one has complete access
solutions, brownfield management has been approached by means of to and detailed knowledge of the simulator used. This is not the case
formal optimization and accurate simulation. In the optimization in many situations in practice (e.g. when the simulation engine relies
problems commonly formulated for brownfields, one aims at significantly on commercial software) and consequently either
maximizing a performance metric such as net present value gradients are approximated (e.g. see Su & Oliver 2010; Dehdari &
(NPV), and the optimization variables are sequences of well settings Oliver 2012; Zhao et al. 2013 for schemes relevant to production
(e.g. well bottom-hole pressure (BHP) values or rates) that represent optimization) or not even computed. Derivative-free optimization
how each well is controlled as a function of time. Operational (non- (e.g. see Kolda et al. 2003; Conn et al. 2009; Kramer et al. 2011)
linear) constraints can be given, for example, by a minimum recovery provides a fairly efficient alternative in practice for a broad spectrum
factor (that may correspond to a baseline production plan) or through of optimization problems in oilfield operations (in this respect, the
pre-specified bounds for field quantities (e.g. maximum field reader may consult, for example, Echeverría Ciaurri et al. 2011b).

© 2018 The Author(s). Published by The Geological Society of London for GSL and EAGE. All rights reserved. For permissions: http://www.geolsoc.org.uk/
permissions. Publishing disclaimer: www.geolsoc.org.uk/pub_ethics
208 O. S. Ushmaev et al.

It is worthwhile noticing that global search methods proceed, in different well control settings. Although a wide variety of well
general, in a gradient-free fashion. Unlike in adjoint-based settings can be handled through these variables, here x will be BHP
optimization techniques, the computational cost of most deriva- values for the injection wells in the field of interest. The
tive-free methods is larger the higher the number of optimization optimization cost function, f (x), may refer to any performance
variables the problem has. However, many gradient-free algorithms metric. In this work, net present value (NPV) for a relatively short
are amenable to a distributed-computing implementation, and that time frame (1 and 3 years) will be considered in all cases as the cost
may yield acceptable performance in terms of wall-clock time. function:
Another family of efficient alternatives to adjoint formulations "
leverages techniques that borrow concepts from streamline-based X X
f (x) ¼ NPV(x) ¼ po qoj,k (x)
simulations (the reader may consult Batycky & Thiele 2006; Datta- k j
Gupta et al. 2010; Voznyuk et al. 2014; Wen et al. 2014 for some #
examples of interest in brownfield management). These techniques X X
þ clp qlpj,k (x)  cwi qi,k
wi (x) Dtk =(1 þ d)tk
are essentially devised to optimize well rates and, in principle, this j i
may not be the best option in real scenarios because a considerable
(1)
number of fields are controlled by means of procedures that do not
provide a constant rate for relatively long periods of time (e.g. where po , clp and cwi are the oil sale price (this price is adjusted so that
injection wells controlled through a selection of wellhead choke the separation cost and taxes are included), lifting cost and injection
sizes). Well controls based on piecewise constant BHP values are cost, respectively; qoj,k , qlpj,k and qi,k
wi are the rates for oil production,
much closer to existing field implementations and, since in this liquid production and water injection for each well j and kth time
work we eventually apply the results of the optimization to an actual interval of duration Dtk ; and d is the annual discount rate where time tk
brownfield, these controls are the ones that will be considered here is expressed in years. We will also include bound constraints: that is, x
(however, the method proposed in this research can also be used will be defined in X ¼ {x [ Rn ; xd  x  xu }, where xd and xu
with well rate controls). denote lower and upper bounds, respectively, and non-linear
Derivative-free optimization approaches, although applicable to a constraints. These latter constraints are mathematically formulated
wide variety of problems, may render formulations that, if there are a by means of a vector function k(x) [ Rnk , in which components
relatively high number of variables, can be extremely time- have to be non-positive for feasible solutions. The specific non-linear
consuming to solve (e.g. when models that integrate reservoir and constraint in our optimization problems refers to a measure of
surface facilities are considered). This might not be an attractive recovery factor (RF) computed for the same time frame as the NPV
option in many practical scenarios where solutions are required in a (this RF measure needs to be larger than that of a baseline plan). Since
short time frame (e.g. overnight or in few days). Hence, derivative- the methodology presented in this research can be applied to an
free optimization is very often combined with surrogate modelling arbitrary number of constraints, we will keep the notation k(x) [ Rnk
(e.g. see Koziel et al. 2011, and especially Jansen & Durlofsky (i.e. as vector function) throughout the paper. The optimization
2017, for applications of reduced-order models in production problem solved can be formulated as follows:
optimization) and with mechanisms that alleviate the computational
x [ argmax f (x) subject to k(x)  0 (2)
cost associated with an elevated number of optimization variables. x[X
In this work we reduce the number of variables in each iteration of
where x indicates any (locally) optimal solution, and the precise
the optimization algorithm by selecting the subset of wells which
definition of the RF non-linear constraint k(x) is adjusted as required
contribution to a measure related, for example, to the cost function
in order to translate feasibility into being non-positive.
or to the constraints is the highest for all subsets with the same
number of wells. It should be stressed that this selection takes into
account information related to streamline-based simulation, is
Derivative-free optimization
applicable to general well controls (not necessarily well rates) and is
updated as the optimal search progresses (since the contribution of For the reasons mentioned above, we opt for derivative-free (i.e.
each well to the measure chosen may depend on the respective well simulator-non-intrusive) optimization methods. Most of the experi-
controls). Although not included in the results presented in this ments performed are based on generalized pattern search (GPS: e.g.
paper, the use of decline models to approximate metrics based on see Torczon 1997; Audet & Dennis 2002; Echeverría Ciaurri et al.
long-term simulation runs (e.g. lifelong reservoir recovery factor) 2011b for more details of that method and applications to oilfield
may accelerate the optimization process further. We would like to operations), which is a local optimization technique (as noticed
note that geological uncertainty is not considered in this research earlier, well control optimization problems have been observed to
(this type of uncertainty may not be as severe for mature fields as for typically present multiple optima with similar cost-function values).
greenfields). In any event, the retrospective optimization framework Additionally, particle swarm optimization (PSO: e.g. see Eberhart
presented in Wang et al. (2012) could be readily used as an add-on & Kennedy 1995; Kennedy & Eberhart 1995; Onwunalu &
to the approach introduced here in order to deal with uncertainty. Durlofsky 2010 for a description of the algorithm and its potential in
This paper is structured as follows. In the next section we state the field development), a global search procedure, has also been tested
optimization problem of interest and describe the methodology out in one example in this work. In both algorithms, the exploration
proposed to solve this problem. Thereafter, in the following section, of the search space is performed by means of a collection of points
we apply this methodology to a numerical model for a field in west (stencil for GPS and swarm for PSO) that are modified in each
Siberia, and later we report on actual experiments performed on that iteration. We reiterate that the computational cost of many
field to validate the solution obtained with the methodology. We derivative-free methods scales with the number of optimization
end the paper with some conclusions and possible future research variables. Therefore, the optimization of mature fields with a
directions. relatively large number of wells is on some occasions a rather time-
consuming solution. GPS and PSO tend to provide fairly robust and
effective implementations (e.g. see Bellout et al. 2012; Humphries
Problem statement and methodology
et al. 2014; Isebor et al. 2014 in addition to the references given for
In this research we address optimization problems in which these two algorithms). Derivative-free alternatives to GPS and PSO
optimization variables x [ Rn represent a discrete time series for can be found, for example, in Kramer et al. (2011).
Efficient brownfield optimization 209

Reduction in the number of variables using streamline- these solutions according to the flexibility available: that is, in some
derived information cases, constraint satisfaction can be relaxed if the improvement in
cost function is acceptably large. In this work we combine the filter
Here we partially alleviate the problem of having a moderate or high with GPS and with PSO (the reader may consult Isebor et al. 2014
number of variables by reducing the number of wells optimized in for more details on the use of PSO with the filter method to deal with
each control interval. This approach is based on constructing a linear computationally expensive non-linear constraints for which deriva-
model for reservoir fluid flow during a given period of time using tives are not available). Alternative approaches to handle general
streamline-derived information in the form of a number of volume constraints in a derivative-free manner within the context of
ratios and efficiencies defined for all injection–production well pairs production optimization can be found, for example, in Echeverría
(for more details, see, for example, Batycky & Thiele 2006; Wen Ciaurri et al. (2011a).
et al. 2014). This model (which essentially approximates the
reservoir thorough a network of pipes) allows one to express
production rates as a sum of contributions associated only with Real field case study
injection wells (and vice versa: i.e. injection rates can be written in
terms of contributions corresponding only to production wells).
Pilot field description
Thus, a performance metric such as NPV can be decomposed into a The optimization methodology has been tested on Field ‘X’ of
number of components linked just to injection wells. We use these Gazprom Neft, located in west Siberia. The formation is a layered,
constitutive components to rank wells and, in turn, to select a subset uplifted reservoir with the presence of structural trapping. The
of them that can be optimized. Metrics associated with non-linear reservoir, which is 12 × 4.3 km and 51 m thick, is mainly composed
constraints (e.g. RF in our case) can also be used for the ranking of sandy rocks interbedded with siltstones, dense clay and carbonate
(otherwise, selections based on a metric solely related to the rocks. The formation is divided into 9–28 permeable layers. The
optimization cost function may favour infeasible well configura- average thickness for the pay sand is 25 m and the initial thickness
tions). Note that a different subset of wells can be obtained for each for the net pay varies from 1.1 to 25.6 m. The average porosity,
of the control intervals considered, and that the subset selection is permeability and initial oil saturation are 18.8%, 23.6 mD and 61%,
updated in every iteration of the optimization method applied. respectively.
Although this streamline-based approximation is suited to wells Oil reserves were proved for Field ‘X’ through exploration
controlled by means of rates, the precision obtained with it in drilling in 1989. The development of the field, mainly via
brownfield management scenarios for other types of well controls waterflooding, started in 2006. Production started from the
can be often satisfactory (connections between injection and (vertical) exploration wells and additional horizontal wells were
production wells tend to be stable for relatively long periods of progressively drilled. The average production rate for the vertical
time in mature fields, provided well controls are not modified and for the horizontal wells was 723 and 3189 barrels (bbl) per day,
drastically). This reduction approach might be attractive in real respectively. The peak oil rate took place from 2007 to 2012 and was
implementations because field operators prefer production plan approximately equal to 17 000 bbl/day. The lack of abrupt injection
updates where only a small number of controls are changed (in breakthroughs and gradual increase in water-cut suggests high fluid
many cases, adjusting well controls implies travelling to remote displacement efficiency. Field rates and the number of active wells
locations of difficult access). It should be added that not all as a function of time are represented in Figures 1 and 2. The
production scenarios are amenable to being modelled efficiently by optimization period considered in this paper started on 1 May 2016,
streamline-based simulation but the management of brownfields, when the field was in the fourth stage of development. At that
among others, is one of them (e.g. see Wen et al. 2014). moment, there were 19 production wells (16 of them were
horizontal) and 15 injection wells (one of them was horizontal),
and the field oil production rate was 6440 bbl/day. Injection and
Filter method for non-linear constraints production rates for all the wells in the field at the beginning of the
The non-linear constraints in the optimization problem in equation optimization time frame are represented in Figure 3. The water-
(2) are challenging to deal with because, as is the case with the cost flooding optimization is addressed through the adjustment of
function, simulation is needed to evaluate k(x) and, consequently, controls for all the 15 injection wells. This optimization strategy
derivative information may not be available. The filter method (e.g. requires no capital expenditure and looks especially promising
see Fletcher et al. 2006; Nocedal & Wright 2006) has recently been under conditions of relatively low oil price. The simulation model
used in production optimization (e.g. see Echeverría Ciaurri et al. used in the optimization is based on a black-oil formulation and is
2011a) to handle successfully this kind of constraints. A filter is discretized by means of a 70 × 145 × 114 grid (the number of active
essentially a new acceptance criterion for each optimization iteration grid blocks is equal to 768 000). Each of these grid blocks has a
that is based on concepts borrowed from bi-objective optimization. length and width equal to 100 m, and an average height of 0.3 m.
By means of the filter method, points are accepted/rejected not only The history-matching quality of the model in most regions of the
based on the values of the cost function f (x) but also on a measure reservoir was deemed satisfactory for the optimization experiments
of constraint violation: performed. However, some wells were located in areas where this
quality was questionable and, as will be discussed below, control
h(x) ¼ max{0, k(x)} (3) changes affecting those wells in the optimized solutions were
where the maximum is determined for all components of k(x), disregarded.
which are usually scaled so that they have comparable values. Thus,
the optimal search with a filter is enriched using points that violate
Results of optimization
constraints and, thanks to that additional information, optimization
techniques that include a filter for non-linear constraints may In the optimization we aimed to maximize the NPV as defined
outperform other approaches that are based mainly on feasible earlier in equation (1) over a production time frame equal to 1 year
points. The filter method is really an add-on to many optimization (first numerical test) and to 3 years (the rest of the numerical tests).
algorithms. The output of an optimization method that incorporates The optimization constraint requires that the RF (or equivalently, the
a filter is a collection of solutions that represent a trade-off between volume of oil produced) during that same time frame be larger than
cost function and constraint violation. The user can select any of the RF associated with an existing baseline. The optimization
210 O. S. Ushmaev et al.

Fig. 1. Oil and liquid production rate, water injection rate, and water-cut (WCT) for Field ‘X’ since the start of production until April 2016.

variables describe sequences of BHP values that represent the The GPS algorithm terminates when the size of the compass stencil
controls for each injection well as a function of time. We have tested (e.g. see Echeverría Ciaurri et al. 2011b) used to explore the
two derivative-free optimization algorithms: generalized pattern optimization space is smaller than a predefined value (this smallest
search (GPS) and particle swarm optimization (PSO). In both stencil size can be selected through considerations around the
algorithms, the non-linear constraint for RF is handled by means of resolution deemed acceptable for the control variables), while PSO,
procedures based on the filter method (in the optimization problem which has a swarm size equal to 100, is run multiple times due to its
there are also bound constraints, which can be dealt with in a random nature and is always stopped after 1600 cost-function
relatively straightforward manner). The optimization starting point evaluations (in our experiments this number seemed to be high
for GPS is always the baseline plan for both time frames of 1 and enough for the algorithm to converge). This swarm size being
3 years (this plan is not included in the initial swarm of PSO in order somewhat larger than typical choices in optimal well location (e.g.
to better check how global the search performed by this method is). see Onwunalu & Durlofsky 2010; Echeverría Ciaurri et al. 2011b;

Fig. 2. Total number of active production and injection wells as a function of time for Field ‘X’ since the start of production until April 2016.
Efficient brownfield optimization 211

with a considerable number of variables, which can be a time-


consuming solution when derivative-free methods have to be used
(we reiterate than in many cases, in practice, efficient gradient-based
techniques cannot be applied because the user has no access to the
simulation source code). Additionally, from the point of view of the
field operator, frequent modifications of the well controls may not
be desirable since these changes usually imply the execution of
costly processes (e.g. travel to remote locations and technically
complicated procedures). Here we have approached the optimiza-
tion over a time frame of 1 year using well controls that are constant
over intervals of either 3 months or the entire year. Five injection
wells (selected by an expert) have been considered for optimization
in this test (note that all wells in the field are taken into account in the
simulations). Thus, the number of optimization variables for each of
the two settings just described is equal to 20 and 5. The results
obtained with the GPS algorithm are shown in Figure 4. In this
figure we represent the evolution of the best feasible solution for the
two optimization runs based on well controls that are constant over
intervals of 3 months and 1 year. The difference between the
optimized NPVs is relatively small (i.e. 0.96 and 1.13%) but the
computational cost associated with the use of controls for 1 year is
around 10 times smaller. Hence, in the rest of numerical tests (all of
them for a total production time frame of 3 years) the piecewise
constant representation of BHP controls for injection wells will be
based on intervals of 1 year.
Next, we compare GPS with PSO when applied to the
optimization of the same five wells as in the previous case but
over a production time of 3 years (divided in three intervals of 1 year
during which BHP is constant for each of the five wells). Therefore,
this optimization problem has 15 variables. The results obtained are
Fig. 3. Injection and production rates for all of the 19 production wells illustrated in Figure 5. Again, the plots represent the evolution of the
and 15 injection wells in Field ‘X’ on 30 April 2016. The area of each best feasible solution in each optimization process. The PSO
bubble is proportional to the respective total well rate. Net pay and oil– algorithm was run four times and the average of all these runs is
water contacts are also shown in the figure. shown in the plot (it should be noted that all the runs yielded the
same optimized NPV; this result suggests that no more runs may be
Wang et al. 2012), where, as mentioned above, cost functions may be needed to assess performance for this configuration of PSO in this
considerably non-linear, may enhance the global nature of PSO in example). Both GPS and PSO reach values very close to the
our application. converged NPV (i.e. increase of 2.26%) after around 400 function
First, we need to decide on the size of the time interval where evaluations. The PSO algorithm seems to take advantage of a large
BHP is constant for each well. In principle, one would like to use as swarm size (equal to 100) to explore the space initially in a more
many intervals as possible and exploit that flexibility in the thorough fashion and improve the cost function faster in spite of not
optimization to possibly obtain higher NPV. However, a large including the baseline plan as a particle in the first swarm (the
number of intervals will be translated into optimization problems stochastic component of PSO may have contributed during the

Fig. 4. Evolution of the best feasible


solution computed with GPS for two
optimization runs over a total time frame of
1 year. The optimizations are based on
intervals of 3 months and 1 year,
respectively, for each of the five wells
considered (which were selected by an
expert). fev, function evaluations.
212 O. S. Ushmaev et al.

Fig. 5. Evolution of the best feasible


solution computed with GPS and PSO for
optimization over a total time frame of
3 years. The optimizations are based on
intervals of 1 year for each of the five wells
considered (which were selected by an
expert). fev, function evaluations.

earliest stages of this optimization problem to deal with the non- of distinct subsets of wells for each of the (three, in this example)
linear constraint in more efficient manner). The systematic search control intervals and for every iteration of the optimization algorithm.
performed by GPS (supported by mathematical convergence Different selections can be obtained depending on the performance
theory) is translated into more rapid convergence in the neighbour- metric that is used to rank the wells involved in the optimization. Here
hood of the optimal solution. we consider two different performance metrics to choose five wells:
It is important to notice that the optimized BHP controls for all NPV and the overall operational expenditures (OPEX) associated
solutions obtained with both optimization techniques are almost with all the wells in Field ‘X’. The selection of five wells instead of
identical (in Fig. 6 we show the baseline, the solution found by GPS another number did not follow any particular guideline in this
and one of the solutions determined with PSO; we can see that the work. Research on criteria to choose that number could be a
two optimized solutions differ by a small margin in the third interval useful enhancement to the methodology. It is important to notice
of injection well INJ-13; the differences between the solutions and emphasize that, in spite of selecting only five out of the 15
computed in PSO are negligible). This fact may indicate that this wells, we may still explore the original 45-dimensional space
problem has a unique (global) optimum (as indicated above, well because the well selection could be modified in each iteration of
control optimization problems with a higher number of variables the optimization algorithm (in other words, the overall method
may present multiple solutions but with similar cost-function contemplates changes in the controls for all 15 wells). We
values). These results suggest as well that hybridization of global additionally compare the two mentioned selection procedures
and local techniques (as proposed, for example, in Isebor et al. with the optimization of the controls corresponding to five wells
2014) may somewhat accelerate convergence. chosen by an expert (these five wells are the same for every
Finally, we address the control optimization for all the 15 injection control interval and each iteration of GPS).
wells in Field ‘X’ during three intervals of 1 year. The attendant The results of the four optimization runs are shown in Figure 7.
optimization problem has 45 variables. As explained above, in this In this figure, the runs that correspond to the selection of wells by
work we have introduced a procedure to accelerate the optimal search means of an expert and through NPV and OPEX as performance
by means of streamline-derived information and the iterative selection metric are denoted by C1, C2 and C3, respectively. As expected,

Fig. 6. (Left) Baseline of BHP controls for the optimization problem with three intervals of 1 year and five wells selected by an expert. (Centre) BHP
controls optimized with generalized pattern search (GPS). (Right) BHP controls optimized with particle swarm optimization (PSO). The optimized solutions
(slightly) differ only in the third interval of injection well INJ-13. This difference will very probably become smaller if the optimization processes are run
over a longer time. In any event, the cost-function values associated with these solutions are essentially the same (in both cases they represent an increase of
2.26% in NPV with respect to the baseline).
Efficient brownfield optimization 213

the explicit optimization of the 15 wells yields the highest optimized solutions are located in the interior of the search
optimized cost-function value (i.e. 3.91% increase in NPV with domain (i.e. for each solution k(x) is strictly less than zero or,
respect to the baseline) and demands almost 5000 cost-function equivalently, the RF is higher than that of the baseline) and
evaluations, noticeably more than the other three approaches. infeasible points only appear in the optimization process during its
Although the optimization problem was formulated for a first stages (note that the initial guess in the optimization is at the
production time frame of 3 years, it is important to verify that the boundary of the search space, i.e. k(x) is equal to zero). None of the
optimized strategy does not underperform the baseline also in the infeasible points visited in the optimization runs were regarded as
long term. To this end, we have computed its NPV and RF during being superior to the feasible solutions found.
the period starting on 1 May 2016 and ending when cash flow The solution with highest optimized NPV (i.e. the solution where
in equation (1) is equal to zero. In this period, the well controls at BHP controls for all 15 injection wells were considered as
the end of the first 3 years (i.e. on 30 April 2019) are kept optimization variables; in practice, one may not have computing
untouched until production stops. Nothing guarantees that the time to optimize all 15 wells and may resort to selecting a subset of
overall strategy will be optimal for the entire reservoir life but, since them using the procedure presented in this paper) was scrutinized
both NPV and RF were the cost function and constraint of an before proposing the attendant experiment on the real field.
optimization problem defined over a relatively long period of time, Although this solution is based on adjustment of 14 out of the 15
we can expect that long-term performance metrics may be injection wells, only changes in three of these wells were initially
acceptable. The optimized strategy (followed by unchanged deemed acceptable from a practical point of view (the contribution
controls) yields around 6 additional years of profitable production to the increase in NPV from many of these wells is relatively small
and the overall increase in NPV with respect to the lifelong and, consequently, difficult to measure experimentally; in addition,
performance of the baseline plan is 7.89%. It is worthwhile some of the wells are located in regions of the reservoir where the
commenting that, even though in the optimization RF is introduced quality of the history matching is questionable and, in consequence,
as a constraint, the RF obtained for the complete reservoir life is modifications of their controls may not have the effects on the real
1.14% higher than that of the baseline (note that this is an absolute field predicted through simulation). The difference in this study
and not a relative increase, i.e. 1.14% additional recoverable between the number of wells optimized and those eventually
reserves are produced compared to the baseline). However, we considered practically relevant is noticeable. Further research could
stress the fact that the degree of uncertainty in long-term predictions be directed towards reducing the gap for these two numbers (e.g. by
is so high that conclusions based on this type of forecast need to be considering new formulations of the optimization problem that
drawn carefully. promote certain types of solutions).
The procedure based on NPV (OPEX) for well selection yields an The three wells chosen and the respective relative rate changes
increase in NPV of 3.73% (3.35%) and is around 59% (66%) faster that correspond to the new BHP controls are shown in Figure 8. (It is
than the direct optimization of all wells. When the five wells are important to note that the two solutions obtained using the reduction
chosen by an expert, the optimized solution is computed in slightly in the number of wells via streamline-derived information also
less than 1000 cost-function evaluations but the corresponding included relatively major changes for these three wells.) The most
increase in NPV is only of 2.26%. We can clearly see a trade-off drastic recommendation is related to injection well INJ-2: the
between accuracy in the optimization and computing cost that we optimized solution requires shutting the well in (i.e. relative change
can adjust according to the precision needed in the solution and the of −100%). The other two recommendations refer to an increase in
resources available (e.g. time). It is worthwhile mentioning that, as BHP for wells INJ-1 and INJ-3. The high injection rate of INJ-2 in
pointed out earlier when describing the filter method, if constraint the baseline plan might hinder effective fluid displacement from
satisfaction can be relaxed, slightly infeasible solutions with cost- INJ-1 to the five nearby production wells indicated in Figure 8.
function values higher than the optimized one (which corresponds, Therefore, the shutting-in of INJ-2 combined with a modest
in general, to a feasible solution) can also be outputted. In this injection rate increase of INJ-1 may distribute better fluid flow and,
particular optimization problem we have observed that the in turn, produce the reserves in that area in a more effective manner.

Fig. 7. Evolution of the best feasible


solution computed with GPS for the
optimization of 15 injection wells over a
total time frame of 3 years. In three of the
runs, five wells are selected in each
control interval and iteration of GPS. In
the run with the first selection procedure
(denoted in the figure by ‘5 wells C1’)
wells are chosen by means of an expert. In
the other two runs (denoted in the figure
by ‘5 wells C2’ and ‘5 wells C3’) the
selection procedure leverages streamline-
derived information to approximate the
performance metric NPV and OPEX,
respectively. fev, function evaluations.
214 O. S. Ushmaev et al.

Fig. 8. The three injection wells from the


optimized solution selected initially as
candidates for implementation in Field ‘X’
(eventually only changes in the controls
for INJ-2 were tested out). The
corresponding relative changes in
injection rate are also shown in the figure.
Note that, according to the optimized
solution, INJ-2 has to be shut-in (i.e. the
relative rate change is −100%) and that
the actual well settings to be modified in
the real field are choke sizes.

Production wells around INJ-3 may also benefit from a higher rate at the three wells identified above relevant in that solution, only one
this injection well. of them, INJ-2, was selected for actual implementation in the real
Taking all the selected changes of the well controls in total, we field. In addition, this well, for which shutting-in was recom-
expect that a reduction in the overall injection rate of 23% for INJ-1, mended in the optimization, would be still active although with
INJ-2 and INJ-3 will not be translated into a decrease in oil −50% relative injection change. There are two main reasons why
production (due to a decline in reservoir pressure) because other the complete optimized solution cannot be tested out in practice.
remaining reserves will be produced; additionally, the savings in First, there are important infrastructural constraints that were not
injection and lifting costs will contribute further to an NPV increase. taken into account in the numerical model used in the optimization
We reiterate that the well settings considered in the optimization are (e.g. technological limitations for the injection redistribution in
BHP values (the actual settings that can be modified for injection the current surface network). The use of a comprehensive model
wells in the real field are choke sizes, which are easier to relate to that integrates the reservoir and facilities may allow one to deal
BHP values than to rates). with these constraints at the expense of a considerably higher
computational cost than the already time-consuming reservoir
flow models. This more complex optimization process may be a
Results of field experiment possible direction for future research. In the second place, the high
degree of uncertainty typically present in real-field applications is
The experts responsible for direct operation and control of Field very often translated into the questioning of new plans that depart
‘X’ examined the optimized solution with the highest NPV. From noticeably from the status quo. Hence, rather than modifying

Fig. 9. Monthly operational rates for the


region of the experiment in Field ‘X’
before and after the optimized plan is
applied. The region of the experiment
comprises injection wells INJ-1, INJ-2
and INJ-3, and production wells PRO-1,
PRO-2, PRO-3, PRO-4 and PRO-5, as
shown in Figure 8. WCT, water-cut.
Efficient brownfield optimization 215

Fig. 10. Water rate reduction for the


injection wells in the region of the
experiment (i.e. wells INJ-1, INJ-2 and
INJ-3) and for the remaining 12 injection
wells in Field ‘X’. Different vertical axes
are used for each region.

significantly the settings for a given well, it may be a reasonable the optimized plan is applied. As can be seen in that figure, the
strategy to introduce relatively small changes in the current optimized plan yields in the region of the experiment a significant
controls that are representative of those associated with the drop in water injected together with a noticeable water-cut
drastically new settings (e.g. −50% relative injection change instead reduction and slight decrease in oil produced (note that this
of shutting the well in): that is, perform some kind of sensitivity decrease in oil production, typical for a brownfield, is clearly more
analysis in the real field in order to validate the solution proposed. pronounced for the rest of the reservoir, as will be illustrated
Consistent with these limitations, the choke for INJ-2 was replaced below). In order to better quantify this improvement in production
by a smaller one, which provides, approximately, a relative 50% performance, these results have to be compared with a reference.
injection reduction. Simulation results indicated that this recom- Since during a given time frame only one production plan can be
mendation yields around 40% of the increase in NPV given by the implemented, an ideal comparison can never be made in practice.
complete (optimized) solution, together with satisfaction of the A reasonable approach would be based on using a baseline
short-term RF constraint. The results of the corresponding field study determined with the trends for the existing plan just before the
are presented next. optimized strategy started. However, and this is especially
In this analysis, Field ‘X’ will be divided in two regions: the area common in the management of mature reservoirs, field develop-
affected by the optimized policy, which will be referred to as the ment implies continuous monitoring and corresponding adjust-
region of the experiment, and comprises injection wells INJ-1, ments of the production plan anywhere in the field. This is
INJ-2 and INJ-3 and production wells PRO-1, PRO-2, PRO-3, translated into frequent engineering activities, which may happen
PRO-4 and PRO-5, as shown in Figure 8; and the rest of the simultaneously and in the same region as the optimized strategy.
reservoir. The selection of the wells that belong to the region of the In our case, during the time frame of the test, the operator
experiment was made, as in the reduction stage described earlier, performed a number of treatments that affected the entire field
using streamline-derived information (i.e. the wells chosen were (including the region of the experiment). The effects of these
those with a relatively strong connection with the injection wells activities should be somehow taken into account in the analysis
INJ-1, INJ-2 and INJ-3). We represent in Figure 9 the monthly and that can be a rather challenging enterprise when real data are
operational rates for the region of the experiment before and after considered. To this end we proceed as follows. First, we compare

Fig. 11. Liquid rate reduction and


corresponding water-cut (WCT) for the
production wells in the region of the
experiment (i.e. wells PRO-1, PRO-2,
PRO-3, PRO-4 and PRO-5) and for the
remaining 14 production wells in Field ‘X’.
216 O. S. Ushmaev et al.

Fig. 12. Evolution of the oil rate for the


production wells in the region of the
experiment (i.e. wells PRO-1, PRO-2,
PRO-3, PRO-4 and PRO-5) and for the
remaining 14 production wells in Field
‘X’. Linear decline models, which are a
reasonable approximation for relatively
short time frames, are shown for each
region.

production performance before and after 1 May 2016 for the two between oil volumes we resort to a linear decline model, which is
regions into which the field is divided. This comparison, when often a reasonable approximation for short time frames. We thus find
analyzed outside of the region of the experiment, may allow us to that the average increase in the volume of oil produced with respect
roughly quantify the effects associated with the treatments to the baseline for the region of the experiment and for the remaining
performed by the operator and, with adequate scaling, estimate part of the field are 25 300 and 24 700 bbl, respectively. If we note
their contribution and deduct it within the region of the that the oil production rate in the second region is about 1.9 times the
experiment. Although this procedure implicitly assumes that the rate in the first region, we can conclude that our methodology may
regions have similar flow behaviour and does not include cross- yield around 12 300 bbl of oil more than the reference.
effects between them, it could provide useful information as far as The overall results from the experiment indicate that the
the impact of our methodology on performance metrics of interest application of the approach presented in this work seems to be
in the management of Field ‘X’. beneficial as far as NPV is concerned. The increase in revenue and
Field measured data are analyzed next according to the three the reduction in water injection expenses outweighs in our analysis
different components in equation (1) for the computation of NPV: (for most logical economic scenarios) the slightly higher lifting
water injection, and oil and liquid production. In Figure 10, we costs. The optimized strategy appears to produce more oil than the
illustrate the reduction in water injection rate for the wells in the baseline even with less water injected and at the expense of a
region of the experiment and also for the rest of the reservoir (note that relatively small amount of additional volume of liquid production.
a different vertical axis is used for each region). The average water We would like to add that the methodology presented could be
injection rate in the region of the experiment decreased from 7090 to enhanced when implemented in a closed-loop optimization fashion:
5430 bbl/day. We estimate that the reduction in the total amount of that is, measurements could be leveraged to update the existing flow
water injected during the 8 months (245 days) of the experiment was, models and with that provide recommendations based on improved
on average, 407 000 bbl. Separately, we quantify the mean decrease production forecasts.
in water injected for the rest of the reservoir as 172 000 bbl. If this
reduction is scaled using the average baseline injection rates for
both regions of the reservoir (i.e. 7090/21 400) and then is subtracted
Conclusions
from the estimation above for the region of the experiment, we can In this paper we have presented a methodology that can be used as a
conclude that in this region the average savings in water injected decision-support system in the optimal waterflood management of
during the 8 months of the experiment could be approximately brownfields. Optimal waterflooding may be an appealing solution
350 000 bbl. It is apparent that these savings have a noticeable (especially in periods of relatively low values of oil price) because it
effect in the operational expenditure component of the NPV. usually requires low capital investments, it generates recommenda-
Figure 11 represents the reduction in fluid production rate for both tions that are in most cases reversible actions (unlike, for example,
regions of Field ‘X’ (the respective water-cut for each region is also fracturing or sidetracking operations) and it allows relatively fast
plotted in the figure). If we repeat the process followed above, we verification of its viability (in contrast to, for example, many
now find that the average reduction in the volume of liquid produced treatments in enhanced oil recovery).
for the region of the experiment and for the rest of the field are 19 000 The approach described relies on simulation-based, derivative-
and 241 000 bbl, respectively. Since the liquid production rate in the free optimization techniques that are non-intrusive regarding the
latter region is much larger than the corresponding rate in the former particular simulator(s) used (i.e. in principle, any simulator can be
region (even when we consider the scaling factor, which is taken incorporated into the system). Gradient-based methods, when
equal to 2.6 in this case), we conclude that our methodology may derivatives are computed efficiently, often provide rapid solutions
actually yield a relatively small increase in liquid production volume, in production optimization scenarios (the computational cost of
and we estimate this increase to be around 73 700 bbl. This result, derivative-free optimization scales, in general, with the number of
together with the fact that the water-cut in the region of the control variables). However, fast computation of accurate gradients
experiment decreases in somewhat clearer form than in the rest of the is typically a simulator-intrusive procedure that cannot be applied in
field, might indicate a higher relative oil production for our approach many situations (e.g. when some commercial simulators are used).
with respect to the reference. Oil production rates for the two regions Simulation-non-intrusive optimization can readily handle a wide
of Field ‘X’ are shown in Figure 12. In order to compute differences range of problem formulations, including most types of well
Efficient brownfield optimization 217

controls (assuming the underlying simulation incorporates these costs clearly outweigh in most economic scenarios the lifting costs
controls as input). Performance metrics such as net present value related to the additional volume of liquid produced (firmer
(NPV) and recovery factor (RF) can be considered as the conclusions could be drawn for longer field experiments). In any
optimization cost function and constraints (the approach can event, we can expect these promising results to extrapolate to other
accommodate flexibility in the satisfaction of the constraints by brownfields scenarios.
outputting slightly infeasible solutions but with satisfactory cost- This research can be extended in several directions. In the first
function values, as demonstrated, for example, in Echeverría Ciaurri place, besides testing alternative optimization techniques (e.g. the
et al. 2011a). use of approximate derivatives via ensemble-based and stochastic-
The number of well controls in many production optimization gradient methods), the optimal search process can be accelerated if
problems may be too high in order for derivative-free optimization to surrogates are considered at different stages of the workflow. The
provide solutions efficiently. Rate controls can be optimized in a selection of the number of wells in the reduced set could be
relatively fast manner by means of procedures that leverage supported by a number of guidelines related, for example, to the
streamline-based simulation. Nevertheless, these procedures are not computational resources available. Formulations of the optimal
applicable to other kinds of well controls such as, for example, BHP search problem may include measures of the plausibility of the
values that are easier to translate into real-field implementations than solutions found in order to make these solutions more readily
rates. In this research we have introduced a procedure that by using implementable on the real field. The underlying simulation-based
streamline-derived information selects the most important wells in the model in the optimization may integrate reservoir and surface
optimization according to a previously chosen metric and, in turn, facilities, and in that manner capture technical constraints that are in
accelerates the entire computing process. Although the approach has many cases disregarded. This type of integrated model might
been illustrated with the optimization of BHP values, any type of well facilitate the optimization of well controls that are more realistic than
controls can be considered. It should be noted that streamline-based bottom-hole pressure (BHP) and rate (e.g. choke size and pump
simulation is a modelling technique especially suited to certain frequency). However, the computational cost of the optimization
production scenarios – brownfield management being one of them. problems thus obtained may in many cases be prohibitive, and
The overall approach was applied to a real brownfield located in strategies to speed up computations significantly would then be
west Siberia that currently has 15 injection wells and 19 production essential. Uncertainty considerations may be included in the process
wells. The optimization problem relied on a history-matched model to endow solutions with additional robustness and flexibility
of the field and aimed at maximizing the NPV over a production (different future actions can be recommended based on upcoming
time frame of 3 years subject to a constraint for the RF defined for information and on the risk attitude of the decision-maker). Finally,
the same period. cognitive computing (e.g. see Kelly & Hamm 2013) may be
Multiple solutions that represent a tradeoff between precision leveraged in the whole system to strengthen the interaction between
and computational cost were computed. If only five wells are human and computational components.
optimized and these wells are manually chosen by an expert, the
increase in NPV with respect to the baseline is 2.26%. When the Acknowledgements The authors would like to thank Arseniy P. Gruzdev
procedure based on streamline-derived information was used to from IBM for his implementation of the particle swarm optimization.
also pick five wells, the NPV obtained was 3.73% higher than the
baseline (nonetheless, the approach was computationally more Funding This work was supported by a funded joint research project between
expensive than resorting to an expert to reduce the number of Gazprom Neft and IBM to develop robust technology for Gazprom Neft’s mature
wells). The optimized strategy with the highest NPV (which also assets in west Siberia. The theoretical outcome of this project formed a basis for a
corresponded to the most time-consuming optimization run and software prototype that was applied to a real oilfield.
entailed considering controls for all 15 injection wells) yielded an
increase of 3.91% with respect to the baseline (the attendant RF References
was also higher than that of the reference but, since this metric was Audet, C. & Dennis, J.E., Jr. 2002. Analysis of generalized pattern searches.
introduced only as a constraint, the observed increase cannot be SIAM Journal on Optimization, 13, 889–903.
Batycky, R.P. & Thiele, M.R. 2006. Using streamline-derived injection
ascribed to the optimization process). This solution is equally efficiencies for improved waterflood management. SPE Reservoir
satisfactory during the complete reservoir life (i.e. while cash flow Evaluation & Engineering, 9, 187–196.
is not negative), even though the optimization problem was only Bellout, M.C., Echeverría Ciaurri, D., Durlofsky, L.J., Foss, B. & Kleppe, J.
formulated over 3 years. The increase in lifelong NPV and RF 2012. Joint optimization of oil well placement and controls. Computational
Geosciences, 16, 1061–1079.
(regarding the baseline) associated with that strategy is 7.89 and Conn, A.R., Scheinberg, K. & Vicente, L.N. 2009. Introduction to Derivative-
1.14%, respectively. Free Optimization. MPS-SIAM Series on Optimization. Society for Industrial
The controls for one injection well were modified over about and Applied Mathematics, Philadelphia, PA.
Datta-Gupta, A., Alhuthali, A.H.H., Yuen, B. & Fontanilla, J. 2010. Field
8 months according to the optimized plan. However, other field applications of waterflood optimization via optimal rate control with smart
operations than this change in the controls for one well took place wells. SPE Reservoir Evaluation & Engineering, 13, 406–422.
concurrently (this could be a frequent situation in real scenarios Dehdari, V. & Oliver, D.S. 2012. Sequential quadratic programming for solving
constrained production optimization – Case study from Brugge field. SPE
because operators may not be able to cancel all previously planned Journal, 17, 874–884.
actions). In order to deal with this situation, we developed a procedure Eberhart, R.C. & Kennedy, J. 1995. A new optimizer using particle swarm
to approximately quantify the impact of our strategy, in the presence of theory. In: Proceedings of the Sixth International Symposium on
others, on the management of the field. This procedure could be useful Micromachine and Human Science. Institute of Electrical and Electronics
Engineers (IEEE), New York, 39–43.
for the operator to better assess the performance of multiple recovery Echeverría Ciaurri, D., Isebor, O.J. & Durlofsky, L.J. 2011a. Application of
mechanisms that are applied simultaneously which respective effects derivative-free methodologies to generally constrained oil production
may be difficult to tell apart. optimisation problems. International Journal of Mathematical Modelling
and Numerical Optimisation, 2, 134–161.
During the 8 months of the experiment, we estimated that the Echeverría Ciaurri, D., Mukerji, T. & Durlofsky, L.J. 2011b. Derivative-free
savings in water injected according to our solution were around optimization for oil field operations. In: Koziel, S. & Yang, S.-X. (eds)
350 000 bbl. Our estimation for the increase in liquid and oil Computational Optimization, Methods and Algorithms. Studies in
production was 73 700 and 12 300 bbl, respectively. Consistent Computational Intelligence, 356. Springer, Berlin, 19–55.
Fletcher, R., Leyffer, S. & Toint, P.L. 2006. A Brief History of Filter Methods.
with the numerical results obtained in the optimization, the extra Technical Report ANL/MCS/JA-58300. Argonne National Laboratory,
volume of oil produced together with the reduction in injection Lemont, IL.
218 O. S. Ushmaev et al.

Harding, T.J., Radcliffe, N.J. & King, P.R. 1998. Hydrocarbon production Onwunalu, J. & Durlofsky, L.J. 2010. Application of a particle swarm
scheduling with genetic algorithms. SPE Journal, 3, 99–107. optimization algorithm for determining optimum well location and type.
Humphries, T.D., Haynes, R.D. & James, L.A. 2014. Simultaneous and Computational Geosciences, 14, 183–198.
sequential approaches to joint optimization of well placement and control. Sarma, P., Chen, W.H., Durlofsky, L.J. & Aziz, K. 2008. Production optimization
Computational Geosciences, 18, 433–448. with adjoint models under nonlinear control-state path inequality constraints.
Isebor, O.J., Durlofsky, L.J. & Echeverría Ciaurri, D. 2014. A derivative- SPE Reservoir Evaluation & Engineering, 11, 326–339.
free methodology with local and global search for the constrained joint Su, H.J. & Oliver, D.S. 2010. Smart well production optimization using an
optimization of well locations and controls. Computational Geosciences, 18, ensemble-based method. SPE Reservoir Evaluation & Engineering, 13,
463–482. 884–892.
Jansen, J.D. & Durlofsky, L.J. 2017. Use of reduced-order models in well control Torczon, V. 1997. On the convergence of pattern search algorithms. SIAM
optimization. Optimization and Engineering, 18, 105–132. Journal on Optimization, 7, 1–25.
Kelly, J.E., III & Hamm, S. 2013. Smart Machines: IBM’s Watson and the Era of Van Essen, G., Jansen, J.D., Brouwer, R., Douma, S.G., Zandvliet, M., Rollett,
Cognitive Computing. Columbia University Press, New York. K.I. & Harris, D. 2010. Optimization of smart wells in the St. Joseph field.
Kennedy, J. & Eberhart, R.C. 1995. Particle swarm optimization. In: Proceedings SPE Reservoir Evaluation & Engineering, 13, 588–595.
of ICNN ’95. International Joint Conference on Neural Networks. Institute of Van Essen, G., Van den Hof, P.M.J. & Jansen, J.D. 2011. Hierarchical long-term
Electrical and Electronics Engineers (IEEE), New York, 1942–1948. and short-term production optimization. SPE Journal, 16, 191–199.
Kolda, T.G., Lewis, R.M. & Torczon, V. 2003. Optimization by direct search: Voznyuk, A.S., Rykov, A.I. & Kotov, V.S. 2014. Multi-criteria analysis and
New perspectives on some classical and modern methods. SIAM Review, 45, optimization of waterflood systems in brown fields. Paper SPE 171229
385–482. presented at the SPE Russian Oil and Gas Exploration and Technical
Koziel, S., Echeverría Ciaurri, D. & Leifsson, L. 2011. Surrogate-based methods. Conference and Exhibition, 14–16 October 2014, Moscow, Russia.
In: Koziel, S. & Yang, S.-X. (eds) Computational Optimization, Methods and Wang, H., Echeverría Ciaurri, D., Durlofsky, L.J. & Cominelli, A. 2012. Optimal
Algorithms. Studies in Computational Intelligence, 356. Springer, Berlin, well placement under uncertainty using a retrospective optimization
33–59. framework. SPE Journal, 17, 112–121.
Kramer, O., Echeverría Ciaurri, D. & Koziel, S. 2011. Derivative-free optimization. Wen, T., Thiele, M.R., Echeverría Ciaurri, D., Aziz, K. & Ye, Y. 2014.
In: Koziel, S. & Yang, S.-X. (eds) Computational Optimization, Methods Waterflood management using two-stage optimization with streamline
and Algorithms. Studies in Computational Intelligence, 356. Springer, Berlin, simulation. Computational Geosciences, 18, 483–504.
61–83. Zhao, H., Chen, C., Do, S., Oliveira, D., Li, G. & Reynolds, A. 2013.
Nocedal, J. & Wright, S. 2006. Numerical Optimization. Springer Series in Maximization of a dynamic quadratic interpolation model for production
Operations Research and Financial Engineering. Springer, New York. optimization. SPE Journal, 18, 1012–1025.

You might also like