SPE 110805

Use of Advanced Optimization Techniques To Manage a Complex Drilling Schedule

Morten Irgens, Actenum Corporation, and William Lashar Lavenue, Saudi Aramco

Copyright 2007, Society of Petroleum Engineers

Results, Observations and Conclusions
This paper was prepared for presentation at the 2007 SPE Annual Technical Conference Early results are convincing and show showed reduced
and Exhibition held in Anaheim, California, U.S.A., 11–14 November 2007.
planning time, reduced response time, improved work
This paper was selected for presentation by an SPE Program Committee following review
of information contained in an abstract submitted by the author(s). Contents of the paper, as
processes, reduction in scheduling errors, improved “what-if”
presented, have not been reviewed by the Society of Petroleum Engineers and are subject to abilities and decision-making, reduction in transportation cost,
correction by the author(s). The material, as presented, does not necessarily reflect any
position of the Society of Petroleum Engineers, its officers, or members. Papers presented at and increased expected production. Laboratory experiments
SPE meetings are subject to publication review by Editorial Committees of the Society of
Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this paper
indicated an improvement in travel cost of 35% and in
for commercial purposes without the written consent of the Society of Petroleum Engineers is production of 25% over common schedule construction
prohibited. Permission to reproduce in print is restricted to an abstract of not more than
300 words; illustrations may not be copied. The abstract must contain conspicuous methods.
acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O.
Box 833836, Richardson, Texas 75083-3836 U.S.A., fax 01-972-952-9435.
Significant New Contributions
The new application brings near real-time automation to a
Abstract previously manual process, and applies optimization
Description of the Material technology to dynamic, mission-critical operations, offering
Rig schedules are major components of operating plans tangible benefits.
and a main connection point between reservoir management
and operations. It is therefore of importance that rig 1 Introduction
scheduling is done well. In an environment of diminishing reserves, everyday is a
As the number of drilling and workover rigs in use at race to produce the maximum volume of oil and gas at the
Saudi Aramco has increased, the challenge of scheduling their lowest possible cost. This is forcing reservoir development
use and transportation has grown exponentially. Ensuring that planning and reservoir operations to become increasingly
the right rig is in the right place at the right time, while responsive, agile and lean. Achieving efficiency and
minimizing rig movements and costs, is practically responsiveness is hindered by a weak coupling (1) between
unmanageable using a manual approach. Consequently, Saudi long-term strategic planning and day-to-day operations, and
Aramco has implemented an automated schedule optimization (2) between reservoir development optimization and asset
application. This interactive application supports schedulers in management; removing these two obstacles becomes two
their daily work, and re-optimizes rig schedules as operational important challenges.
changes render previous plans invalid. There is no panacea and no single approach to address
Application these challenges – rather; they have to be addressed at several
Recent advances in optimization technology address levels and in several ways. This paper will give an example of
complex, dynamic situations, and show the tangible value of how we can address the two challenges through improved
operational decision-making in real time. Saudi Aramco has asset management – in this case scheduling and managing
developed a software application where the rig schedule is drilling rigs. The paper is divided into two parts.
optimized against multiple objectives, disruptions to the Part 1 of the paper is intended for the reader who wants to
schedule is quick and easily managed, the impact of learn about the problem and the solutions without going into
scheduling decisions on user-defined key performance very technical details. The first section in this part is Section 2,
indicators is immediately displayed, and schedule optimization which describes a simplified version of the problem,
is supported by a sophisticated user interface enabling direct scheduling rigs over a certain time period while taking into
schedule manipulation in response to changing needs. account rig transportation cost and expected production.
Saudi Aramco has used this project to investigate how The complexity of this problem makes it unsolvable
recent technological advances can turn rig scheduling into a without the use of advanced information technology. Section 3
continuous operational task that improves the alignment of describes a case study where a software application that
daily decisions with long-term strategic objectives. The incorporates the technological developments described in this
deployed rig scheduling system is thus designed for both paper is deployed to the drilling coordination group at Saudi
planning and operations and allows for managing the rig fleet Aramco. The application is called Aris, and it provides an
while optimizing on both operational and strategic objectives,. operational tool that enables rapid rescheduling in line with
It also allows for rapid rescheduling when changes and long-term strategy and balances asset management objectives
disruptions occur. with reservoir optimization objectives. Section 4 discusses
some future directions of the application, where the focus is on
2 SPE 110805

extending the operational tool towards reservoir management well has an expected production volume and each rig has a
and its integration with operations and asset management, transportation cost. The schedule as a whole has a target
particularly rigs and real time rig schedule optimization. production volume and the objective is to schedule the rigs in
We present the results and conclusions in Section 5. The order to minimize the transportation cost, while meeting or
results were strongly convincing on several levels. The case exceeding the target production volume.
study showed reduced planning time, reduced response time, The challenge in solving this problem is that the number of
improved work processes, reduction in scheduling errors, possible ways in which facts and decision points can connect
improved “what-if” abilities and decision making, significant to build a solution or make a reliable decision grows
reduction in transportation cost, and increased expected exponentially with small increases in the amount of datai
production brought online. Laboratory experiments indicated [Garey & Johnson 1979, Cook 1971]. As the problem also
an improvement in travel cost of 35% and in production of contains complex objectives and constraints, we are restricted
25% over common schedule construction methods. from the use of classical techniques to address this
Part 2 of the paper is intended for the more technically combinatorial explosion in problem difficulty.
interested reader. Section 6 describes how advances in Real world deployments create several additional
Operations Research and Artificial Intelligence – two separate challenges. For example, a scheduler must account for
fields focused on formalization and automation of decision- capacity limitations on well site preparation (leading to
making – allow us to begin addressing the challenges posed by constraints on well site readiness), and reservoir management
the case study. The section also describes the specific requirements (leading to constraints or preferences on the
techniques we have used, while Section 7 describes the drilling sequence of certain wells). As discussed below, there
problem model. Section 8 describes a set of experiments. are also non-trivial challenges arising from a reactive
environment and from the need to align with overall reservoir
Part 1: The Application development strategies.

2 Problem description 2.2 Reducing the Divisions between Long-term

Planning and Daily Operations
2.1 Rig Scheduling Drilling operations are becoming more time-critical,
Above we mentioned the increased pressure felt by oil dynamic, and response-based. This extends the challenge of
companies today to stay lean and agile while managing rig management beyond the creation of rig deployment
increasingly complicated operations. Drilling coordinators are schedules to include ongoing schedule optimization in
consequently under mounting pressure to optimize the drilling response to changing conditions.
schedules and respond quickly to operational changes. This is The moment a schedule is rolled out, unplanned events
becoming a bigger challenge as faster drilling technologies occur, rendering the schedule less useful. The weather does
have created shorter drilling cycles and companies operate an not cooperate. Special equipment is not available. Production
increasing number of rigs. volume from a well did not meet expectations. Reservoir
Saudi Aramco is a case in point. The number of rigs the management changes its opinion about a well drilling
company operates has grown considerably, outgrowing the sequence. When such things happen, even the most carefully-
previous method of scheduling and controlling cost. Drilling prepared schedules will drift from their intended course and
coordinators have many choices to make, and they often have often will become useless. Time pressure is intense and in
to reschedule their drilling programs several times per day. many situations, unplanned events cause cascading problems
The constant re-planning process is complicated by many that can overwhelm the resolution effort.
factors, which include limited resources, increasing rig rates, Traditional optimization approaches tend to assume that
long wait times for materials and equipment and variations in conditions in the real world are static and little consideration is
rig capabilities. given to the need for rescheduling. The techniques focus on
While Aris handles a large and complex problem, in the producing a “one-off” translation of long-term planning
interest of clarity we will present a simplified version of the objectives to short-term schedules. Experience shows that the
problem. The simple formulation of the problem is to schedule schedule and the reality on the ground will eventually have
rigs over a specified period of time while taking into account drifted enough apart for the schedule to need updating. A
the rig transportation cost and the expected production. naive repair of a schedule may produce results that are not
Within the time period, there is a certain number of wells to be aligned with the long-term planning objectives.
drilled and a certain set of rigs available. We will assume that There have been several approaches to improve the
there is only one type of rig activity to be scheduled (i.e. robustness of schedules (resulting in schedules that are useful
drilling), and that rig transportation has a cost but negligible for longer periods), including adding slack and incorporating
time duration. Figure 1 shows an example of an abstract Gantt uncertainty [Davenport et Al 2001, Berry 1993, Memik Et al
chart and transportation map of a schedule. 2002]. While these approaches can be very useful, they may
The drilling coordinators receive a list of wells to drill, only delay the inevitable schedule adjustments that need to be
requirements about how they should be drilled, and made. The challenge therefore is to furnish the drilling
precedence constraints between (some of) the wells. There are planners with a decision support tool that ensures that the
more wells to be drilled than the rig capacity in the given time schedules stay optimized and aligned with the strategic
period, and we will assume that the rigs are never idle. Each objectives, while minimizing disruptions to operations.
SPE 110805 3

2.3 Improving the Coordination between Reservoir scheduler that excels at strategic decisions and has a cultural
Management and Asset Management knowledge that the system does not necessarily have.
Reservoir development encompasses two main entities: Impact on key performance variables (including estimated
The field production system and the geological reservoir production volume and transportation cost) of possible
[Awashti 2007]. Improving reservoir development is not only decisions are immediately displayed for any action and before
complex because decisions in each of these entities impact decisions are effectuated. This allows rapid, agile and
each other, but also because they are handled by various unlimited “what-if” scenario investigations. Using the
groups. For our purpose we can simplify this process and say application, the user always knows:
that the objective of reservoir management is to maximize the
 The impact of scheduling rigs and wells on production
production volume or meet production targets by selecting
output.
which wells to drill and how to drill them. In contrast, the
objective of the rig schedulers is to take the well selection  What the expected rig movements and costs will be.
from the reservoir management and create a schedule (by
assigning rigs to wells) while minimizing transportation cost  How to schedule and reschedule rig and well activities as
and meeting the production targets. operational conditions change and disruptions occur.
Commonly, reservoir development plans based on certain
depletion strategies are turned into operating plans with well  How various schedule alternatives influence the
selections. The operating plans provide operating envelopes operational risk associated with achieving production
for the rig scheduling. This operating envelope is defined by a targets.
number of constraints set by the operating plan, including
production targets and drilling sequence for some wells. If the Like most real world applications, we do not have the
operating envelope provided is too narrow for creating luxury of one single problem or managing one single criterion.
feasible schedules, we enter a cycle where the operating plan Aris therefore needed to provide an overall solution, balancing
gets adjusted until rig scheduling is feasible. This is a number of criteria, including transportation costs, production
problematic because the process is time consuming and error volume, reactivity and schedule disruption. Since these criteria
prone. It is also problematic because the operating plan is are partially mutually detrimental, the application needs to
often created with substantial slack to avoid a cycle of find a good compromise among them. The user can influence
adjustments during operations. In both cases this may result in this, as well as how much a rescheduling is allowed to change
highly suboptimal schedules. the existing schedule.
Providing decision support and optimization that assist this The user can select which performance variables should be
synchronization is not only likely to reduce effort, time and optimized, and to what extent this optimization can be
errors, but is also bringing the overall operation better in line constrained (“enveloped”) by other performance variables.
with key performance objectives. The challenge, therefore, is This is important because different performance variables are
to provide optimized strategies for operational rig scheduling relevant to different groups. While rig schedulers are
and well selectionii. The strategy must incorporate objectives, concerned with minimizing the schedule cost, the reservoir
targets, key performance variables and constraints for both engineers are concerned with maximizing the reservoir
reservoirs and rigs. The system should be able to both operate utilization, and these objectives are negatively correlated, the
within the envelope of the reservoir targets and negotiate these application needs to make it easy for the parties to quickly
targets when necessary. arrive at agreed solutions. Aris allows the two parties to limit
each otheriii. The user can choose to control to what extent
3 Case Study improving one performance variables may impact others
negatively. For instance, the rig scheduler can decide that
In this Section we describe the deployment of the Aris rig
optimizing on production should not increase the rig cost with
scheduling application to Saudi Aramco’s drilling and work-
more than 5%. Similarly, the user can decide that optimizing
over department. We describe the problem model and the
on rig cost should not reduce the production volume in the
technological solution underlying the application in Section. 6.
current schedule.
The application provides rig management and scheduling
optimization using a modern user interface, enabling
4 Future Developments
interactive rig movement maps and Gantt charts. The schedule
optimization occurs through an interactive collaboration This first release of Aris addressed our first challenge of
between the human operator and the underlying technology, bridging planning and operations by providing rig scheduling
where the user can affect the schedule optimization, lock parts as both a planning and operational activity. We are currently
of the schedule and ask the system to improve on other parts. working on extending this to also addressing our second
The result is what is known as a mixed initiative system, challenge, improving coordination between reservoir
where the scheduling decisions are partly made by the system development planning and asset management. The application
and partly by the user. In this manner, we believe that we are will include support for multiple user types and provide a
able to capture the benefit of having two decision makers with coordination channel between reservoir management and rig
vastly different capabilities. The two decision makers are the scheduling. It will thus provide functionalities to schedule rigs
software that is designed to be able to compare a vast amount and select wells in an optimized manner.
of possible schedules in a heuristic manner and the human The application takes several steps to improve the
4 SPE 110805

coordination between reservoir management and rig schedules meet up. The application combined the
scheduling. First, the application provides a way for reservoir schedules into a single view that reduced the time for
managers to manage a repository of well types and wells. It updates and reduced errors.
also allows the system to select and optimize on the selection
of wells and drilling methods (well types) according to  Improved experimentation and decision-making: The
selected criteria. Reservoir managers can add and remove application allows schedulers to evaluate “what-if”
wells and add and manipulate sequence constraints between scenarios more easily and includes comparison features.
wells. The map view of the rig movements helps users identify
Second, the application provides an opportunity for potential concerns more quickly.
reservoir managers to add targets for collections of wells (for  Reduced training time for rig schedulers: Good
instance all the wells in a reservoir) directly into the system. schedulers that know the rig business are hard to find. At
These additional targets could be annual, weekly, monthly, the same time, the oil and gas sector is seeing a growing
quarterly or daily. The users can now optimize on production vacancy rate due to retirement. This application not only
until the target production is met or surpassed, providing a empowers schedulers today, but it also makes the entry
production profile for the various wells and clusters of wells. level for starting scheduling much lower.
Figure 2 shows a screen shot of the scheduling system with
target production and scheduled production. Given a schedule  Reduced transportation cost: The case study showed an
with production profiles, rig schedulers are able to optimize on immediate significant transportation cost reduction for
rig cost, investigating various ways of scheduling the rigs 2007 for the long-term planning alone.
while respecting the production profiles. Hence, as events that It should be noted that the savings in transportation were
invalidate the rig schedule happen, the objective of the done with no reduction in the expected production volume.
schedulers is not only to reschedule in order to keep the rigs When optimizing on an objective, users can allow the
busy and minimize rig cost, but also to meet the production optimizer to decrease the value of other objectives within any
targets in the operating plan. limits. If we allow the optimizer to make any reduction in the
Third, the application provides a way for reservoir expected production volume of the schedule, the
managers to optimize on production volume respecting the transportation cost can be reduced even further. It should also
constraints of the rig schedule. Since the rig scheduling system be noted that the scheduling was done with a low tolerance for
has all the knowledge of the operational assets (e.g., upsetting the current schedule.
availability and allocation of the assets, the transportation In order to better understand the nature of the savings, we
costs, and the schedule disruption levels the operations can ran a set of experiments that confirmed a significant reduction
take), the reservoir managers can play various “what-if” in rig transportation cost. We describe the experiments in
scenarios and have the application optimize on production detail in Section 8 below. Figure 3 shows how transportation
volume within the possibilities and constraints of the cost and production loss improves with run time. In general
operations. The results from this optimization will then we get significant improvements over straight forwardly
provide an envelope for the rig scheduling. constructed schedules in less than ten seconds.
There are a significant number of technologies available to
5 Results and Conclusions provide decision makers with more information from the
The Aris application has brought a number of benefits to drilling operations, either in real time or post drilling. More
the end user organization. data and more information do not necessarily lead to better
decisions. The challenge is that the problems of well selection
 Reduced planning time: The annual rig scheduling and rig scheduling are both very complex and that there is a
optimization process is reduced from approximately 10 growing demand for responsiveness and agility of these tasks.
days to a few hours. In this paper we have discussed two challenges to
 Reduced response time: The time it takes to handle responsiveness and agility in reservoir and asset management.
schedule revisions is dependent on the type of schedule The first challenge is to reduce the division between long-term
revisions that are requested. The benefit from the strategic planning and the day-to-day operations. The second
application here is that revisions result in improved challenge is to increase the coupling between reservoir
schedules compared to before. development optimization and rig scheduling optimization.
The objective of Aris was to meet these challenges through an
 Improved work process: The work process has been efficient rig scheduling application that supports daily
improved on several levels. For example, the end users rescheduling application. The innovations described in this
are now associating Universal Transverse Mercator paper make it possible to improve the flexibility and reactivity
[UTM] coordinates to wells earlier in their work process of the scheduling process and to remove the gap between
to support the application, but the early coordinates are schedule preparation and execution. As discussed, the
also of value for other users. application is also able to optimize on production volume and
 Reduction in scheduling errors: The end user had meet production targets. Importantly, the application is also
multiple schedules that shared rigs. This meant that a able to optimize on rig transportation cost while not
change had to be made to multiple schedules and it was compromising production targets or production volumes
difficult to find overlaps or gaps at the times when the achieved through production optimization. The result is likely
SPE 110805 5

to change the way the end user will manage the relationship of efficient algorithms for scheduling problems. They are
between his/her longer term planning and day-to-day similar to genetic algorithms [Goldberg 1989] and simulated
operations. annealing [Aarts 2003] in the sense that they operate by
Early experiments have shown that the system is able to manipulating candidate solutions and can be seen as exploring
reduce rig transportation costs and increase potential the extremely large space of possible solutions. In general,
production, even when one of these performance variables is SLS algorithms for scheduling problems will:
constraining the other. In addition, the system has improved
the quality of schedules and the collaboration among work  Work on a complete schedule, suitable for presentation to
groups. users, and so can be interrupted at any time.

 Make small incremental changes such as moving a single

Part 2: Technical Solution activity in a schedule, similar to how a human scheduler
would operate.
6 Technological Context
There are two fields of research focused on addressing  Use highly efficient methods to evaluate the benefits of
issues surrounding scheduling, planning and optimization; each change, which also allows the user to monitor the
Operations Research and Artificial Intelligence. Operations progress of the algorithm.
Research (OR) developed as a field during World War II,
when military planners looked for ways to make better SLS algorithms are able to support more “intuitive”
decisions in such areas as logistics and training schedules problem models than those seen in Operations Research (OR).
[Anderson 2005]. The field uses computers, mathematical OR algorithms are specialized for solving specific problem
tools and algorithms to deal with analyses of complex real- models, and may require counter-intuitive descriptions of the
world systems, aid in decision-making and to efficiently solve scheduling problem. In contrast, SLS algorithms operate on
complex management issues – typically with analytical problem models that closely match the thinking of human
models of the tradeoffs involved with the goal of improving or schedulers. This allows application developers to exploit
optimizing performanceiv. domain-specific knowledge and patterns that result in a
The field of Artificial Intelligence (AI) that emerged performance benefit and in schedules that are more acceptable
during the 1950s as researchers started to use computers to to end users. While SLS algorithms will often find good
mechanize formal logic. AI employs computers, symbolic solutions in a reasonable amount of time using models that are
logic and mathematical tools and algorithms to solve not easily represented as linear programs, they are not
challenging problemsv. An important part of the work in AI is guaranteed to find a globally optimal solution
to find techniques that aid decision-making and to efficiently SLS algorithms are particularly appropriate for user
solve complex management issues, including logistics, interaction and so are ideal choices for situations where human
planning, and scheduling. Recently the technologies from the schedulers are already in place. Also, they have been
two fields have progressively intermixed to move forward the successfully applied to other optimization problems in the oil
“state of the art” in decision management. and gas industry [Aloise et Al 2006, Tupac et Al 2007,
Methods known as constraint programming, ant system Beckner et Al 1995].
optimization, evolutionary algorithms and meta-heuristic The algorithm developed for Aris assumes the existence of
search are now becoming integrated with traditional OR an initial schedule. This may be produced as a result of the
approaches. The combination of the mathematics of OR and normal planning process, or may be a previous schedule
the modeling flexibility and solution methods of AIvi provide a generated by the algorithm itself. It then considers possible
qualitative and quantitative leap and has lately created fast and alterations to the schedule, including:
flexible systems for solving difficult practical problems. The
results are technologies that can be used to work on situations  Reinsertion of any single activity to another point in the
that have been inaccessible to previous approaches, but which schedule. Activities are considered in order of a heuristic,
are critical to most organizations in the real world. They and insertion points are considered in order of their
include problems with nonlinear constraints, mixes of proximity to the selected activity.
constraints and preferences, hard and soft constraints, ill-  Swapping of any two contiguous clusters of activities. A
defined objective functions, a preference for practical and “contiguous cluster” of activities is defined as “a sequence
flexible solutions rather than solutions that are optimal with of activities in the schedule, no more than a specified
respect to some monolithic objective, preferences for rapidly distance apart.” The appropriate distance is chosen
finding practical solutions rather than optimal solutions later, according to domain knowledge.
and preferences for heuristic methods.
Of special interest for us are the developments of Each possible alteration is evaluated and the first
automated reasoning, constraint networks, constraint acceptable iteration will be chosen and applied to the
programming and local search. For our rig scheduling schedule. An alteration is acceptable if it improves a user-
challenge, our approach has been a variation of Stochastic specified weighted sum of the objectives we described and
Local Search (SLS) combined with very efficient underlying does not worsen any single objective by a user-specified
data structures, including an invariant network. The field of threshold.
Stochastic Local Search [Hoos 2004] includes a broad range To handle our particular problem, we have designed our
 6 SPE 110805

search algorithm in combination with highly efficient Decision Variables

mechanisms to evaluate the benefits of any modifications to The decision variables are the start time start(a) and rig
the schedule. To compute these objectives efficiently we make rig(a) for each activity a.
use of incremental invariants [van Hentenryck 2005]. An
invariant, in this context, is any mathematical expression that Constraints
we use to evaluate a schedule. Invariants can be used to model There are no overlaps between any two activities ai and aj
regular algebraic expressions, but can include more complex assigned to the same rig r: [start(ai),start(ai) + dur(ai)] ∩
mathematics such as sets, functions, and models of distance. [start(aj),start(aj) + dur(aj)] = Ø. Each activity a must start
In the case of this rig scheduling problem, we need highly
efficient schedule evaluation mechanisms because in addition after its release date, start(a)  release(a). There are no time
to non-trivial objectives, we also have some preferences, such periods when a rig is not used (except at the end of the
as well precedence and drilling time windows. Each schedule schedule). For each consecutive pair of activities (ai ,aj), on rig
is also evaluated according to these preferences and scored r, start(ai) + duration(ai) = start(aj). Each activity a executes
appropriately. For example, the algorithm adds a penalty for only when the rig is available, [start(a),start(a) + duration(a)]
the number of days a well is drilled too early or too late.  available(rig(ai) = Ø. Each activity is assigned to a
Disruption of other schedules and rig idle time are also compatible rig: rig(a)  compatible(a).
penalized. In all, a schedule is evaluated on 10 different
criteria and given a score separate from the production and Objectives
transportation cost. The algorithm is designed to be simple, This problem is multi-objective. In this simplified version,
fast and to use the underlying invariant network, which has we have two objectives: To maximize expected production
resulted in an application that is able to very quickly improve and to minimize transportation cost. Transportation cost for
the quality of schedules, taking into account the multiple each rig is a fixed value for short distances, but it grows
objectives, preferences and constraints found in the real world. linearly for large distances. Due to future uncertainties, the
cost for rig transportation between two points is weighted for
7 Model the time at which it occurs; near-term transportation incurs a
It is a scheduling problem where the goal is to assign start higher penalty than far-term transportation. Production values
times and rigs to drilling activities so that the constraints are for each well are determined by scaling the expected
satisfied while maximizing production and minimizing production volume according to the time of drilling. A well
transportation cost. The problem is similar to multi-mode which is drilled earlier will be able to contribute more to short-
resource constrained scheduling [Jozefowska 2001, Hartmann term production, and so it is of greater value to the company.
2001] in that each activity may be considered to have one The production objective is measured by the production loss,
mode for each different rig that could be used to execute the which is the opportunity cost of not scheduling an activity as
activity. On the other hand, the transportation cost aspects early as possible:
make it similar to a vehicle routing problem [Johnson &
McGeoch 1997, Toth & Vigo 2001] or scheduling with loss(a) = max(0,start(a) – release(a)) ∙ production(a)
changeover costs [Beck].
The transportation cost is based on the distance between
Data each consecutive pair of activities on the same rig r and the rig
There is a set of activities a1, a2,…an corresponding to the cost for r.
wells that should be drilled. Each activity has a fixed duration
duration(a), a geographic UTM coordinate, an expected TransportationCost(ai,aj) = fixed(r)
production volume production(a), and a release date + max(0,d(ai,aj) – K) ∙ perKmCost(r)
release(a). An activity should not be scheduled to start before
its release date. The release date may be used to ensure that where d(ai, aj) is the distance between the activity ai and
the drilling activity is not scheduled to start before the well activity aj as calculated from their UTM coordinates.
site has been prepared. To service the application requirements, we focus on one
For execution of the activities, a set of rigs r1, r2,…rm is objective at a time and turn the other objective into a
available. Each rig has an available time interval avail(r) that constraint or goal as is done in goal programming [Eilon
limits when the rig can be used for drilling activities. A set of 1970]. For example, the objective when optimizing on
costs are associated with each rig: The daily drilling rate, a production is:
fixed cost per move, fixed(r), and a variable transportation cost n
per km moved over K km, perKmCost(r ). minimize  max(0,start(ai) –release(ai)) ∙ production(ai)
A drilling activity requires a single rig for its entire i 1
duration and the activities cannot be pre-empted. The rig that such that transportation ≤ goaltransportation ∙tolerance.
is used to execute an activity is selected from a subset of the
rigs. For example, an oil rig cannot be used for drilling a gas The constant goaltransportation allows the users to define the goal
well. We denote this subset of rigs for an activity a as or target transportation cost. When re-optimizing or starting
compatible(a). from an existing solution, the goaltransportation is set to the
transportation cost in the exiting solution times a tolerance
SPE 110805 7

factor. As an example, this allows the user to optimize the current schedule. Aris does not worsen any objective,
production in an existing schedule while keeping the regardless of the current selection, unless the user allows.
transportation cost within 5% of its current value. This gives the user a strictly improving schedule. In addition,
even if the current focus is on improving production, Aris take
8 Experiments opportunities to improve on travel cost.
As Aris makes use of a random seed, we will make
8.1 Methodology multiple runs of each with the same problem instance and
obtain averaged performance results.
The focus of this application was on finding good and
workable schedules fast, (rather than chasing optimality in
8.2 Solution Quality
long computational runs) and to balance several objectives.
This makes it challenging to evaluate the overall benefit of the Figure 3 shows the improvement over time of
algorithmic design and to design experiments to illustrate transportation cost and production volume when optimizing on
these benefits. We have chosen to look at only two of the transportation cost. Figure 4 shows the improvements over
benefits provided by our algorithm: the improvements made to time of the same objectives when optimizing on production
transportation cost and production volume. volume. Each plot includes the mean and upper/lower standard
To evaluate the solver, we have constructed a random deviations. The horizontal axis displays algorithm runtime in
problem instance with similar attributes to what is observed in seconds, while the vertical axis describes the percentage
industry. The problem instance contains 20 rigs separated into improvement for each objective.
four drilling types. Each rig has a fixed cost for each transport, Optimization on any objective function did not allow any
with a mean of $140,000 and standard deviation of $40,000. reduction in other objective values, e.g., optimization on
Each rig has also a cost per kilometer it is moved, with a mean transportation cost did not allow any reduction in production
of $1,200 and standard deviation of $300, taking effect after volume. In these experiments, Aris does not focus exclusively
the rig travels more than 20 km. We then assume there are 15 on increasing production or reducing travel cost, but instead
drilling fields, each containing 10 to 30 wells. The fields form will provide greater weight to one or the other. As described in
natural clusters of various sizes: 2-16 km at the widest point, the previous section, production improvement is measured by
and 1-6 km at the narrowest point. Each well has a drilling production loss, which is the opportunity cost of not
duration estimate, taking between 10 and 100 days to drill. scheduling an activity as early as possible.
For the purposes of this experiment, we used a variant of First, we note that there is substantial improvement in both
the H1 heuristic to construct an initial schedule [Allahverdi & objectives, regardless of which we select for optimization. For
Aldowaisan 2001]. We will briefly describe the H1 heuristic, example, when optimizing for production, we see an 18%
and then comment on our variation of this heuristic. improvement in production within the first 20 seconds of
The H1 heuristic, has been used for scheduling rigs by algorithm execution, but also see a 25% improvement in travel
successively appending the "most valuable" unscheduled wells cost. Similarly, when optimizing for travel cost, we see a 37%
to the ends of rig routes [Aloise et al. 2006]. "Most valuable" improvement in travel cost, and a 4% improvement in
is in that application of the heuristic defined as the downtime production. Allowing a longer runtime resulted in a 38%
of the well multiplied by the potential production rate, which improvement for travel cost and an 8% improvement in
is equivalent to the “production loss” measure given before. production.
The H1 heuristic approximates a schedule construction Using the H1 heuristic to produce an initial schedule gave
technique observed in industry, where higher-performing us more opportunity to improve travel cost than to improve
wells are scheduled as early as possible and each rig is production. Given a schedule with good travel cost but poor
assigned a roughly equal number of wells. initial production, Aris would still find significant
Unfortunately, this scheduling approach sometimes improvement within a short period.
produces poor rig routes. If the travel time between wells is We can still observe the effect of selecting an optimization
large (for example, if it takes three times as long to move a rig focus despite that the initial schedules favor production. If we
as to drill a well), then this heuristic will result in schedules inspect the standard deviation curves in Figs. 3 and 4, we can
that maximize the distance traveled by rigs. The result is that see that there is a lower standard deviation for the objective
in extreme cases, not only will this increase travel costs, but that we selected for optimization. A lower standard deviation
also will decrease production. This worst-case behavior is not over multiple algorithm executions is indicative of a near-
observed in [Aloise et al. 2006] as, for their data set, it takes optimal solution. By selecting an optimization focus, Aris is
three to 48 times as long to drill a well as to move a rig. able to drive a single objective to near-optimality, while taking
Indeed, in our experience, transportation time is almost any opportunistic improvements available in other objectives.
negligible when compared to activity durations. It therefore
seems appropriate to eliminate travel time from the H1 8.3 Practical Benefits
heuristic. We therefore use the resulting “H1 sans travel Beyond the clear financial benefits of schedule
times” to generate initial schedules for our experiments. optimization, we can also use these graphs to analyze some
After establishing an initial schedule, we use Aris to practical aspects of the Aris solver. As previously mentioned,
optimize for either production or transportation cost. As wse all objectives improve rapidly within the first 20 seconds of
described earlier, this selection is normally made by a user, our solver, indicated by a steep descent in the graph. This is
and allows them to consider alternative improvements to the followed by a long "tail" or "plateau" period, where only small
8 SPE 110805

additional improvements are found. Such a plateau period may presented at the 1995 SPE Annual Technical Conference &
continue for up to an additional 500 seconds with only small Exhibition, Dallas, Texas, October 22-25.
change in objectives or, indeed, the schedule itself. [Cook 1971] S. A. Cook, The Complexity of Theorem Proving
This behavior reflects the underlying design of the Procedures, Proceedings, Third Annual ACM Symposium on
the Theory of Computing, ACM, New York, 1971, 151-158
algorithm, which is focused on providing good solutions fast -
[Davenport et al. 2001] A.J. Davenport, C. Gefflot and J.C. Beck,
Aris makes the majority of improvements within a very short "Slack-based Techniques for Robust Schedules," Proceedings of
time frame. The reason for this design is that we designed Aris the Sixth European Conference on Planning (ECP-2001), 2001.
to be used in an operational environment. Such short [Eilon 1970] Samuel Eilon, Goals and Constraints in Decision-
turnarounds are necessary for using the application in a real- Making, Operational Research Quarterly (1970-1977), Vol. 23,
time production environment or when interacting with the user No. 1 (Mar., 1972), pp. 3-15.
for conducting “what-if” analysis. Indeed, the user can very [Garey & Johnson 1979] Garey, M. and D. Johnson, Computers and
quickly establish and present a lower bound on possible Intractability; A Guide to the Theory of NP-Completeness,
schedule improvements. 1979. ISBN 0-7167-1045-5 (This book is a classic, developing
the theory, then cataloging many NP-Complete problems).
We also note that, despite the inclusion of a random
[Goldberg 1989] Genetic Algorithms in Search, Optimization, and
number generator in the optimization algorithm, there appears Machine Learning, Addison-Wesley Publishing Company,
to be a very small deviation from the mean value between 1989.
runs. It is sometimes claimed that local search is [Hartmann 2001] Project Scheduling with Multiple Modes: A
unpredictable, and is prone to becoming "trapped" far from the Genetic Algorithm. Annals of Operations Research, Volume
optimal schedule. Our extensive tests, and hinted in the graphs 102, Numbers 1-4, February, 2001.
presented above, lead us to conclude that for Aris’ intended [Hoos 2004] Holger H. Hoos and Thomas Stützle, Stochastic Local
use, the local search algorithm we designed has highly Search: Foundations and Applications, Morgan Kaufmann/
predictable behavior and repeatedly produce high-quality Elsevier, 2004.
[Johnson & McGeoch 1997] D. S. Johnson and L. A. McGeoch, The
solutions.
Traveling Salesman Problem: A Case Study in Local
Optimization, Local Search in Combinatorial Optimization, E.
9 Acknowledgments H. L. Aarts and J.K. Lenstra (ed), John Wiley and Sons Ltd,
We would like to thank Dr. Kwaku M. Temeng, Aramco 1997, pp. 215-310.
Services Corporation, Jacqueline Stamatopolous, Saudi [Jozefowska 2001] Joanna Józefowska, Marek Mika, Rafa Rózycki,
Aramco, and Dr. Ken Jackson, Peter Harvey and James Styles, Grzegorz Waligóra and Jan Weglarz, Simulated Annealing for
Multi-Mode Resource-Constrained Project Scheduling, Annals
Actenum Corporation.
of Operations Research, Volume 102, Numbers 1-4, February,
2001.
10 References [Toth & Vigo 2001] Paolo Toth (Editor) and Daniele Vigo (Editor),
[Aarts 2003] E.H.L. Aarts, J.H.M. Korst, P.J.M. van Laarhoven, The Vehicle Routing Problem, Siam Monographs on Discrete
Simulated Annealing in Local Search in Combinatorial Mathematics and Applications, 2001.
Optimization, Princeton University Press, 2003. [Tupac et al. 2007] Y.J. Tupac, L. Faletti, M.A.C. Pacheco and
[Allahverdi & Aldowaisan 2001] Allahverdi A, Aldowaisan, T, No- M.M.B.R. Vellasco, Evolutionary Optimization of Oilfield
Wait and Separate Setup Three-Machine Flowshop with Total Development/Evolutionary Optimization of Oilfield
Completion Time Criterion, International Transactions in Development, SPE 107552, SPE Digital Energy Conference,
operations Research, 7(3), 245-264. 2007.
[Aloise et al. 2006] Aloise, D. J., Aloise, D., Rocha, C. T., Ribeiro, [UTM] TM8358.2: Defense Mapping Agency Technical Manual
C. C., Filho, J. C. and Moura, L. S. 2006. Scheduling Workover 8358.2. The Universal Grids: Universal Transverse Mercator
Rigs for Onshore Oil Production. Discrete Appl. Math. 154, 5, (UTM) and Universal Polar Stereographic (UPS), http://earth-
2006, pp. 695-702. info.nga.mil/GandG/publications/tm8358.2/TM8358_2.pdf.
[Anderson 2005] David R. Anderson, Dennis J. Sweeney and [van Hentenryck 2005] Pascal Van Hentenryck and Laurent Michel,
Thomas A. Williams, An Introduction to Management Science - Constraint-Based Local Search, MIT Press, 2005.
Quantitative Approaches to Decision Making (with CD-ROM
and InfoTrac®) 11th Edition ©2005, ISBN: 0324202318. 11 Figures
[Awashti 2007] Ankur Awashti, Sathish Sankaran, Michael
Nikolaou, Luigi Saputelli and Gerardo Mijares, “Closing the
Gap between Reservoir Simulation and Production
Optimization,” The Society for Petroleum Engineer Digitial
Energy Conference, Houston, 2007.
[Backer et Al] Backer, B.D., Furnon, V., Kilby, P., Prosser, P. and
Shaw, P., Solving Vehicle Routing Problems using Constraint
Programming and Metaheuristics, Technical Report, GreenTrip
Project.
[Beck] J.C. Beck, P. Prosser and E. Selensky, On the Reformulation
of Vehicle Routing Problems and Scheduling Problems, Tech.
Report APES-44.
[Beckner et al. 1995] Beckner, B.L. and Song, X.: "Field
Development Planning Using Simulated Annealing - Optimal Figure 1. The left part of the picture shows which wells the
Economic Well Scheduling and Placement," paper SPE 30650 rigs have to drill at what times. The right side of the screen
shows the rig movements.
SPE 110805 9

Figure 2. An abstract screen capture of a scheduling

system with target production (dotted curve) and estimated
production (solid curve) for a collection of wells (for
instance for a reservoir).

Figure 4. Percentage improvement in transportation cost

and production loss as a function of run-time when
optimizing on production loss alone. At the 20 second
mark, the top dotted curve shows the average improvement
in transportation cost at each time point over 40
experiments, surrounded by the curves showing the
standard deviations. The bottom curves similarly show the
average and standard deviations for the improvement in
production loss.

i
From computational complexity theory, scheduling is a problem that is
known to be NP hard [Garey & Johnson 1979], i.e., at least as hard as NP. It
is strongly believed that NP hard problems cannot be solved in polynomial
time.
ii
The challenges to achieving cross-level decision support systems are
many, including different timescales, different objectives, different expertise
and different technologies.
Figure 3. Percentage improvement in transportation cost iii
and production loss as a function of run-time when The problem could be handled by providing a multi-objective optimizer
that investigates the Pareto border in question. However, such a solution
optimizing on transportation cost alone. The top dotted would mean a substantial change in the current work process in the
curve shows the average improvement in transportation organization.
cost at each time point over 40 experiments, surrounded by iv
The name “Operations Research” or “OR” is somewhat of a misnomer,
the curves showing the standard deviations. The bottom since OR is no longer concerned only with operations, and its application does
curves similarly show the average and standard deviations not involve any research in the traditional sense (though OR research is still
for the improvement in production loss. carried out to find new or better techniques). The terms management science
or decision science are also often used to describe the same field of expertise.
v
The name “Artificial Intelligence” or “AI” is somewhat unfortunate,
since AI is generally not concerned with creating artificial awareness or
consciousness. Instead, AI is playing an important role in businesses,
government, and the defense sector, performing tasks like online tutoring,
fraud and money laundering detection, data mining, biometric evaluation,
essay grading, speech recognition, self-adapting software virus protection, risk
management, civilian crisis and evacuation support, and disaster recovery.