Supplementary files are available for this work. For more information about accessing
these files, follow the link from the Table of Contents to "Reading the Supplementary Files”.
able of Contents
Alternative Modeling Approaches:
A Case Study in the Oil & Gas Industry
Scott Johnson
Bob Eberlein
Name: Scott Johnson
Organization: BP
Complete Postal A ddress: 3111 Winding Shore Lane, Katy TX 77450
Phone/Fax: 281.366.2889/281.366.3436
E-mail address: | ohnsoST@ bp.com
Name: Bob Eberlein
Organization: V entana Systems, Inc.
Complete Postal Address: 17 Loker Street, Wayland MA 01777
Phone/Fax: 508 651 0432/508 650 5422
E-mail address: bob@ vensim.com
Abstract
Simulation modeling is one technique that BP relies on to improve the quality of capital
investment decisions. To date these simulations have been highly detailed discrete event
simulations with the financial computations run in parallel using spreadsheets. In an
ongoing project, some of the tools and thinking of system dynamics are being introduced.
The goal is to understand how they can complement or substitute the existing tools and
analysis techniques and to gauge the reaction of people relative to the existing approach.
For analytic work in support of capital investment decisions the ultimate measure of
success is not always clear. Plants and distribution systems that perform as expected in
terms of throughput and uptime do validate the results of technical components of the
simulation. Making such comparisons in detail is, however, costly and time consuming.
In addition, changes in market and business conditions tend to have a very large impact
on the economic value of infrastructure and the uncertainty around those is extremely
high.
In this study, we will discuss a project in which we introduced a continuously formulated
model as a supplement to the discrete event model. For this case we found that the
continuously formulated model could be used to address the same issues as the discrete
model and could do so efficiently. We also found that the discussion generated by this
alternative formulation was more productive and tended to lead the project team in a
more interesting direction. While all of these results are tentative, we do believe this is a
worthwhile approach to developing models in support of investment decisions.
Overview of Paper
In this paper we will first give background on BP’s investment activity and the manner in
which models are used to help guide that. We will touch on the component models,
project organization, and workflow typically used to attack the problem and then outline
the reasons we thought system dynamics might help. Following this we will discuss the
traditional solution and then what we did to approach the problem in parallel. We will
then compare the reactions of the project teams to the two approaches and draw some
conclusions. The actual models used for the project and their results are proprietary. We
have, however, included a version of the continuous model that has been modified to
contain no proprietary data in the program CD.
Background
BP is a vertically integrated oil and gas company with operations spanning the globe. Its
business is organized into upstream, downstream, and chemicals sectors. The upstream
sector is responsible for exploration, development, and production of raw oil & gas
products. The downstream sector refines the raw upstream products into petroleum
products such as gasoline and lubricants and sells them to the public. The chemicals
sector is engaged in the production of feedstock for a variety of industrial manufacturing
processes.
In order to accomplish these activities, each sector routinely invests in the design and
construction of new production facilities. Due to the commodity nature of oil & gas,
there is intense competition among industry participants to efficiently build
technologically advanced, safe facilities while balancing capital investment, operating
costs, and availability. As stated in the annual report, during 2001 BP invested a total of
$13 billion in new facilities, 70% of which was spent in the upstream sector. The
magnitude of these upstream capital investments demands that appropriate tools and
techniques be used to continuously improve the quality of decisions and resulting
shareholder value.
BPs current approach to improving the quality of facility capital investment decisions
involves the preparation of one or more discrete event models to evaluate availability and
size facilities along with spreadsheet models to evaluate the economic consequences.
These models are normally built during the early phases of a project, well before
construction begins, to finalize the facility design. The WITNESS discrete-event
simulation environment and Excel spreadsheet software are used to construct these
models. The WITNESS models are built to calculate facility availability based on a
variety of equipment assumptions such as Mean Time Between Failure (MTBF), Mean
Time To Repair (MTTR), availability of spare equipment, frequency and duration of
scheduled maintenance and resource availability. Custom input and output Excel
spreadsheets insulate the user from the underlying complexity of the WITNESS model
and automate many analysis tasks. The separate economic model is typically created
using Excel spreadsheets and uses as inputs the results of the availability modeling
activity.
The completion of these models requires a multidisciplinary team with skills in
engineering, commercial analysis and technical modeling support. Most modeling team
members come from the business unit recommending the investment, which will also
operate the resulting facilities and be held accountable for the operational and financial
performance of the facilities. BP also has a group that supports modeling across business
units called the Upstream Technology Group, Integrated Business Modeling Team (UTG
- IBM) of which one of the authors is a member. The team members work together,
often with contract consulting support, to develop the models to support the investment
analysis. For the technical models flexibility and model processing time are two very
important characteristics since the process of selecting the final design is highly iterative,
involving the identification and evaluation of multiple cases.
A typical development effort starts with a discussion between an engineer and an internal
modeling support person. After a review of past projects, a starting template is found that
most closely matches the current project. A process flow diagram and reliability block
diagram are then developed and used as input for the discrete event simulation model.
The development of the discrete event simulation model is typically done by a contractor
with the results and assumptions refined based on discussions with the BP personnel.
Once the engineers are comfortable with the model results the commercial analysts use
them as a basis for developing cost models. The technical models give an indication of
expected throughput and operating expense which, combined with investment costs and
assumptions on market conditions, allow the construction of a financial model. There is
then some iteration with the financial and technical models making changes and
rerunning them both for the purposes of plant design and understanding uncertainty. This
iterative process can span multiple days and involve several meetings between
stakeholders to complete.
Problem Statement
Given this background, we wanted to introduce elements of the system dynamics method
and also make use of the Vensim® software to see what kind of difference this would
make. This was done in the context of real projects which are still ongoing as of the
writing of this paper. The approach we used was not to replace the methods and models
in use, but to build additional model in parallel with the standard development stream and
present these to the project teams. We were looking for answers to, or at least insights
into, the questions:
1. What is the best way to use system dynamics methods to improve the quality of
capital investment decisions?
2. What are the benefits associated with a causal view?
3. What are the advantages of having a single model that integrates both facility
availability and economic consequences?
4. Does an integrated model approach allow for faster turnaround in terms of
simulation speeds and understanding of results?
5. Can the alternative models replicate the results produced by the discrete event
WITNESS models?
6. What do the project members like and dislike about the alternative models relative
to the standard models?
7. How do the alternative models differ from a more standard system dynamics
model?
8. How do the alternative models differ from the WITNESS discrete event models?
9. How can the alternative models position the project team to investigate different
problems that they could not address using the WITNESS models and
spreadsheets?
To address these questions we are using this parallel development approach on two
projects. The first project is a plant to produce Liquefied Natural Gas (LNG), and the
second is a transportation system (storage tanks, boats, and docks) to move LNG from a
plant to the destination distribution facilities. In this paper we will focus on the first of
these as that is the one for which we were able to do a substantial amount of work and
interact thoroughly with the project team. Longer term, we hope to link the models from
the two projects into a single model to understand the value of bringing these projects
together.
LNG Production Problem
BP is continually researching the development of new LNG plants to service markets
around the globe. The major questions that need to be answered are: how big to make the
plant, what should be the balance between plant equipment reliability, redundancy and
capital cost and what are the bottom line implications?
In order to address these questions a project team was formed following the process
described under Background. The project began without any emphasis on the alternative
modeling approach. During the development of the discrete event model the authors
developed the alternative model behind the scenes. There was no attempt to engage
participants in the system dynamics process or change the early problem clarification
process. The value of such early intervention is not being investigated here, though it is
certainly an interesting research problem.
The WITNESS Discrete Event Model
The WITNESS model uses an excel interface to both enter inputs and display results. A
portion of the model itself is shown in Figure 1. It is a very detail oriented model that
includes each major piece of equipment. For each piece of equipment there are from one
to three failure modes and each of these has an associated mean time to failure, mean
time to repair, and effect of plant throughput. The failure times are assumed to have a
negative exponential distribution, the repair times a log normal distribution, and the effect
on plant throughput a constant.
The WITNESS model is run with for 20 years with 100 replications. This is done to get
average values for throughput and repair activity and is a standard thing to do with
models having a great deal of randomness. Although the 20 years is a reasonable time to
think of a plant running without a major refit, there is no attempt to capture the lifecycle
of the plant or startup and leaming effects. The statistics that come out of the simulation
are steady state statistics.
mi OD Mm IS #5) | se [ro por |BAR2|ba\er
th Selection
ppseart | 0 |) toe Moomre cuncck | [FE wriness Cwitress ING. PD BOSE FRE A aOBRL oem
Figure 1. A partial view of the WITNESS discrete event model.
The Vensim Discrete Event Model
We built a Vensim model that used the same conceptual modeling framework as the
WITNESS model. An overview of that model is shown in Figure 2 (the model itself is
available on the proceedings CD'). This model is also highly detailed, but the detail is
relegated to the model subscripts rather than the diagram. For example, the variable
Equipment is Functional represents 263 different equipment/failure mode
combinations.
At first glance, this looks like a fairly typical causal loop diagram. However, on closer
inspection it is clear that many of the variables such as Equipment is Functional
can only take on two values. Although there is feedback in this model, there is not any
structure that can generate surprising dynamics.
* Please note that we had to make some adjustments to the Vensim model to be included with the paper.
Most importantly, the different failure and repair times have been changed because this information is
considered proprietary.
-tenspreent ty Functional traapmentt yee]
Fl Bat Medst Took we Ip 18)x)
S68 (9 £26 Fives Ht Se vagy SOOM
AM AMPH EVALER BMS lH 3
of rs
@ average LNG Somme Pararaiose
be production
= i ‘
“s in
a product
a slantindot eas
> plet avnisbdty
ca =
on m@ throughput
turd wn marcia
a
aeeeage ple Rewer Gemerason
avalatdy (ilies
Sassen
ecqupronnt fans
‘Mesa Time T al
Fabre
“ a
Per Tea eon oS
Raat] (OE & a || Bie tess ramcatanel, | veeernron ore wee QEOLIPREL LAO som
Figure 2. A overview of the V ensim model.
The Vensim model is different from the WITNESS model in that we use a constant Time
Step of 1 hour, rather than stepping through time to the next event. We also use a random
realization of failures with a Poisson distribution, instead of distributing the failure times
with a negative exponential distribution. These two representations can be shown to be
the equivalent, though our use of the Poisson distribution on repair completions is
different from the log normal distribution on repair times in the WITNESS model. There
is also one element of the WITNESS model (the Plant cold box) that uses a custom
distribution that was not implemented in the Vensim model.
Nonetheless the two models give results within 2% and respond in the same way to
changes in key assumptions on reliability.
The Vensim model also has built in some rudimentary financial and operational
computations as shown in Figure 3. These are similar to those used in the financial
models done in Excel, though they are not yet complete. This makes the financial results
available alongside the technical results without having to run things through another
model. Because the current model has incomplete financial modeling we have not
presented this to the project team as yet.
Opx t w\
before tax
<$ per MM *\ . operating income
“i Fixed ai,
vo opx Spend Rae
So ne A io oveane Rate
Opx per a.
Production
<LNG production> er MM:
proauco LNG Price S ‘
Figure 3. Example financial computations.
Model Comparison
As we have noted, the two models produce substantially the same results. It is
interesting, as well as delightful from our perspective, that while the WITNESS model
takes about 20 hours to complete 100 simulations of 20 years each the Vensim model can
do the same thing in about 1 hour. This result is important because it does have
implications on how the model can be used to investigate problems.
In any situation where the project stakeholders are not directly involved in developing the
model it is important that they have techniques for building confidence in its correctness.
With the WITNESS model the intended mechanism for doing this was to recreate and
animate the Reliability Block Diagram used as a basis for developing the model. This
diagram names the different pieces of equipment and the animation simply involves
showing which piece of equipment is currently broken along with the current plant status.
In practice, there were just too many elements in the model to be able to watch this, map
to the plant life time scale and form any judgment about the correctness of the model. In
the end the WITNESS model was treated as a black box with changes made to inputs
checked against the resulting performance changes.
The Vensim model is presented in a different format. We did not reproduce the
Reliability Block Diagram, though some of the project members though that we should.
Confidence in the V ensim model was built based on our ability to explain the model logic
and demonstrate how it was done, along with the fact that it did the same thing the
WITNESS model did. Due to the very technical nature of this model it was not the case
that the model could easily be made to tell a story. We had to rely on our ability to
describe how things worked. When compared with the WITNESS model this was murky
compared to muddy - an improvement but only modestly. It is possible that the
incorporation of Reality Check® logic into the model could be used to improve
confidence and this is something we are looking into.
While most of the confidence building through parameter changes was done using the
WITNESS model, it is true that the faster simulation time for the Vensim model would
have made this easier. Ironically, the experienced WITNESS modeler working on the
project was concerned that the V ensim model was not giving correct results because it
ran too fast. This concem faded after this individual reflected on some of the things that
might be causing the WITNESS model to run slowly. It was, however, observed by
several team members that individuals experienced with WITNESS models might be
uncomfortable changing to a Vensim model. Nonetheless, additional research on using
Vensim to address availability problems was recommended for selected projects.
Model Attractiveness
The ability to show causality and explain logic in the Vensim model, while modest for
the availability computations, was outstanding for the financial computations. Not only
were the relationships clearer than the spreadsheet cell references people were used to,
the fact that they were directly tied back to the availability model was considered
outstanding. The team quickly moved to the idea of integrating not only the economic
models but also transport and other models. Because of this extensibility Vensim was
perceived as an appropriate tool for exploring new areas.
The Vensim model was also perceived as a good tool for educating others about the
impact of various plant equipment configurations and specific equipment selection.
Ultimately this would lead to stakeholder buy-in and confidence in the plant design. The
Vensim model would also be useful for helping operations and maintenance personnel
learn the logic of the plant.
Conclusions and Extensions
While this is ongoing work and it seems likely that much of the learning remains ahead of
us, there are some relatively solid conclusions based on what we have done so far.
First it is both possible and fruitful to develop models typically done as discrete event
models using a continuous approach. Both approaches have the same theoretical
foundation and both yield substantially the same results. While the continuous approach
is not uniformly better than the discrete approach it does have some advantages. For this
case, based on the implementations, speed was one advantage. More importantly,
however, is the added ability to extend the model with more continuous additions.
Second, the continuous model seems to engage people in discussion in a direction that is
more general. Rather than digging into details of failure modes and the implications of
specific combinations of failures there was a tendency to move toward discussion of the
business implications of different design decisions. The discussions were also more fully
engaging for more team members.
It is, of course, important to do a good job on the availability and throughput modeling.
Our original development in this area combined a flow oriented approach to plant
operation with the failure mode modeling described in this paper. We did not pursue this
because we felt it was important to develop a model that could replicate the WITNESS
model results. Nonetheless, this combination of the stochastic failure modes with the
deterministic process flows would yield an improved ability to experiment with plant
design. This is something that can be most easily pursued in a continuous modeling
environment.
Finally, a few closing comments on the relationship of the model developed to more
traditional system dynamics models. The Vensim model might be best cartooned as a
very simple, limited feedback, system dynamics model agglomerated onto a complex
discrete event machine. It is a brute force marriage of two paradigms that actually works.
Not that many system dynamics models are run hourly for 20 years, but they can be.
While elegant ways to incorporate altemative modeling approaches into system dynamics
models do seem desirable, there is a lot that can be done with the tools at hand.
Back to the Top
CC BY-NC-SA 4.0