Sei sulla pagina 1di 85

MSc (Petroleum Engineering)

Project Report 2012/2013

Obinna Nwafor
H00137989

History Matching and Uncertainty


Quantification with Produced Water
Chemistry

Heriot-Watt University
Institute of Petroleum Engineering

Supervisor -

Declaration:
IObinna Nwafor onfi m th t thi wo k bmitt d fo
m nt i
my own and is expressed in my own words. Any uses made within it of the
works of other authors in any form (e.g. ideas, equations, figures, text,
tables, programs) are properly acknowledged at the point of their use. A
list of the references employed is included.
Sign d..
D t 20August, 2013

ii

Acknowledgements
I would like expressly my profound gratitude to my Supervisor Oscar Vasquez for his tireless effort
in guiding me through this work. He has taken me from reservoir studies, and mixed in a little bit of
chemistry and lots mathematics. Without his kind reassurance, I would not have come this far.
My gratitude also go to th whol Un t inty research group at IPE, Heriot-Watt University, led
by Mike Christie. I am very grateful to Mike, especially for those solutions he throws my way at the
moments my wits were fully spent. Vasily Demyanova has been to me a great teacher and guide,
especially with the woolly concepts of uncertainty quantification. He never once got deterred by my
silly questions, which popped out every morning. I am truly thankful. Dan Arnolds has been
inv l bl to m in t hing nd t nding th t n w y. He takes time to lead them back-in,
ensuring they were securely tucked into their beds.
I would not have had the courag to t k on m th m ti l optimi tion witho t th
encouragement from Eric Mackay. I am truly thankful to his insight, especially those words about
f t
job p
nt tion . My g tit d l o go to th Ch pl in of H iot-Watt University,
Alastair Donald, every Sunday I found my way back to him for the spiritual fodder that kept me
going.
I appreciate Epistemy Ltd for the Raven software used for this study, and the Computer Support
Team who kept the systems running in spite of the mounting pressure.
I wish to acknowledge all my lectures and tutors at the IPE over the last one year, they gave me a
little bit of themselves, which I have put together in this work.
My parents, especially mother, my sisters and brother in-law have been a rock to me throughout
my one year stay in Edinburgh which culminated in this work. Thank you for being there, caring
and offering your words and resources. I thank the Almighty God for his grace every step of the
way.

iii

Abstract
History matching is used in reservoir calibration. Conventional history matching could be improved
by addition of more constraints to be matched. The injected sea water, produced as part of associated
water could have the potential of serving as an additional constraint. Such data can be obtained
cheaply by using ion in sea water as natural tracers.
This study aims to determine the extent of improvement to history matching and reduction of
uncertainty in forecasts brought about by the addition of injected sea water production data to the
history matching process. The study was carried out using the PUNQS3 reservoir model. It is a
synthetic reservoir model that has been used for similar studies and severally to test out methods in
history matching and uncertainty quantification. The uncertain parameters in the PUNQS3 model are
the porosity and permeability.
Two cases of automatic history matching were carried out. One involved the use of injected sea
water tracer production data as additional constraint. The other involved using only the conventional
production data. The automatic history matching was based on multi-objective particle swarm
optimization (PSO). Which is a nature inspired stochastic optimization technique. The uncertainty
quantification was done using Neighbourhood Approximation method (NA-Bayes) based on a
Bayesian framework. The result for the two cases was compared on the basis of advance of their
pareto front of models ensemble towards lower misfit values. The quality of history matching and
size of uncertainty were also considered. They all show indications that addition of injected sea water
production data improves the history matching process, reduced uncertainty in forecast, as well as
creates a more robust uncertainty quantification. However the improvements were not large, but
could be more significant for more complex reservoir history matching problems.

iv

Table of Contents
Acknowledgements ......................................................................................................................... ii
Abstract ........................................................................................................................................... iii
1

Aims......................................................................................................................................... 1

Introduction ............................................................................................................................. 2
2.1

Reservoir History Hatching .............................................................................................. 2

2.2

Automatic History Matching and Mathematical Optimization ........................................ 3

2.2.1

Construction of Inverse Problems and Mathematical Optimization Problems ......... 4

2.3

Value of improved history matching ................................................................................ 4

2.4

Tracing Injected Sea Water in Hydrocarbon .................................................................... 5

2.4.1

Origin of Oilfield Water ............................................................................................ 5

2.4.2

Composition of Oil Field Water ................................................................................ 6

2.4.3

Compatibility of Formation Water and Sea Water .................................................... 6

Review of Concepts and Studies ............................................................................................. 9


3.1

Stochastic Optimization Techniques for History Matching.............................................. 9

3.1.1

Genetic Algorithm ................................................................................................... 10

3.1.2

Differential Evolution .............................................................................................. 10

3.1.3

Ant Colony Optimisation......................................................................................... 11

3.1.4

Particle Swarm Optimisation ................................................................................... 12

3.1.5
Use of Particle Swarm Optimization for Reservoir History Matching Optimization
Problem 14
3.1.6

Multi-Objective Particle Swarm Optimization. ....................................................... 14

3.1.7

Optimal Solutions in Multi-Objective Optimization: Pareto Front ......................... 15

3.2

Uncertainty Quantification for Forecast ......................................................................... 16

3.2.1

Definition of Uncertainty......................................................................................... 16

3.3 Uncertainty Quantification Method: Neighbourhood Approximation Using Bayes


Theorem (NA-Bayes) ................................................................................................................ 17
3.3.1

Bayes Theorem ........................................................................................................ 18

3.3.2

Likelihood of Observation ....................................................................................... 19

3.3.3

Bayes Integral: Resampling by NA-Bayes .............................................................. 20

3.3.4

Neighbourhood Approximation and MCMC Walk ................................................. 20

3.3.5

Bayesian Credible Intervals ..................................................................................... 22

3.4 Review of Empirical Studies on Improving History Matching by Adding Data on


Injected Sea Water Production .................................................................................................. 22
3.5

Problem Statement .......................................................................................................... 24

Data Summary: PUNQS3 Reservoir Model .......................................................................... 27


4.1

Origin .............................................................................................................................. 27

4.2

Available Reservoir Description ..................................................................................... 27

Methods ................................................................................................................................. 29
5.1

Work Flow Diagram ....................................................................................................... 29

5.2

PUNQS3 Problem: Uncertain Parameters ...................................................................... 30

5.3

Modifications to the PUNQS3 and Historical Data ........................................................ 30

5.4

Parameterisation .............................................................................................................. 30

5.4.1

Regions and Justification/ Generation of Simulation Files ..................................... 31

5.4.2

Correlation of Porosity and Permeability ................................................................ 33

5.4.3

Parameter Distribution ............................................................................................. 34

5.5

Selection of Histories to Match and Objective Functions for Optimisation ................... 34

5.5.1

Case 1: ..................................................................................................................... 35

5.5.2

Case 2: ..................................................................................................................... 37

5.6

Generation of Truth Case Histories ................................................................................ 37

5.7

Variance for History Data ............................................................................................... 38

5.8

Optimisation Algorithm and Setup ................................................................................. 39

5.8.1

Setup of PSO run ..................................................................................................... 40

5.8.2

Ensemble of Models: convergence of ensemble; number of Ensembles ................ 41

5.9

Uncertainty Quantification: ............................................................................................ 41

5.9.1

Prior Probabilities

5.9.2

Likelihood of Models

5.9.3

Monte Carlos Integration of the Bayesian Integral ................................................. 43

5.9.4

Bayesian Credible intervals. .................................................................................... 44

5.10

................................................................................ 42

Analysis of Results ...................................................................................................... 44

5.10.1

Research Question 1 ................................................................................................ 44

5.10.2

Research Question 2 ................................................................................................ 44

5.10.3

Research Question 3 ................................................................................................ 45

Results ................................................................................................................................... 46
6.1

Research Question 1 ....................................................................................................... 46

6.1.1

.......................................................................................... 42

Identification of Pareto Solutions Models for Case 1 and Case 2 ........................... 46

6.2

Research Question 2 ....................................................................................................... 47

6.3

Research Question 3 ....................................................................................................... 49

Discussion and Conclusion .................................................................................................... 50

vi

7.1

Improvement to History Matching.................................................................................. 50

7.2

Economic Value .............................................................................................................. 52

7.3

Reduction of Uncertainty ................................................................................................ 52

7.4

Conclusion ...................................................................................................................... 53

References ............................................................................................................................. 55

Appendix ............................................................................................................................... 58
9.1

Detailed Comparison of Front Advance by Region ........................................................ 58

9.1.1

Case 1: Excluding Sea Water Tracer Production Data ............................................ 59

9.1.2

Case 2: Excluding Sea Water Tracer Production Data ............................................ 60

9.2

3 Dimensional Pareto Plot for Case 1 ............................................................................. 60

9.3

Field Oil Production Total (FOPT) Intervals for Case1, C-Case and Case 2 ................. 61

9.4

Comparison of History Match for Best Five Models of Case 1 and Case 2 ................... 64

Well Water Production Rate ...................................................................................................... 64


9.5

Well Bottom Hole Pressure ............................................................................................ 67

9.6 Comparison of History Match for Most Pareto Models for Case 1(Iteration 493 Run2)
Case 2(Iteration 262 Run3)-Well Bottom Hole Pressure .......................................................... 70
9.7 Comparison of History Match for Most Pareto Models for Case 1(Iteration 493 Run2)
Case 2(Iteration 262 Run3) - Well Water Production Rate ....................................................... 73
9.8

Field Oil Production Total .............................................................................................. 76

Table of Figure
FIGURE 3-1VORONOI CELLS FOR TEN RANDOM POINTS (MODELS) IS A SOLUTION SPACE. B. THE UPDATED
VORONOI CELLS AFTER 100 POINTS ARE SAMPLED AND INTERPOLATED USING GIBBS SAMPLER. .................. 21
FIGURE 4-1 EXPECTED FACIES WITH ESTIMATES FOR WIDTH AND SPACING OF MAJOR FLOW UNITS ...................... 28
FIGURE 5-1WORK FLOW DIAGRAM ............................................................................................................................. 29
FIGURE 5-2 INJECTORS AND PRODUCER WELLS ON THE PUNQS3 RESERVOIR MODEL .............................................. 30
FIGURE 5-3SAMPLE OPTIMIZATION RESULT-QUALITATIVE PERMEABILITY MAP-LIGHT BLUE SHOWS HIGHER PERM.
........................................................................................................................................................................... 32
FIGURE 5-4 3 LAYER 3 PARAMETERISATION ................................................................................................................ 32
FIGURE 5-5 LAYER 2 PARAMETERISATION .................................................................................................................. 32
FIGURE 5-6 LAYER 1 PARAMETERISATION ................................................................................................................... 32
FIGURE 5-7 LAYER 5 PARAMETERISATION ................................................................................................................... 32
FIGURE 5-8 LAYER4 PARAMETERISATION.................................................................................................................... 32
FIGURE 6-1 CASE 1: SWTP PARETO PLOT .................................................................................................................... 46
FIGURE 6-2 CASE 2: NSWTP PARETO PLOT .................................................................................................................. 46
FIGURE 6-3 COMPARISON OF PARETO BY TRADE-OFF REGIONS ................................................................................ 47
FIGURE 6-4 CASE2: NSWTP FOPT ................................................................................................................................ 48
FIGURE 6-5 CASE1: SWTP FOPT ................................................................................................................................... 48
FIGURE 6-6 CASE 2: NSWTP-FWPT .............................................................................................................................. 49
FIGURE 6-7 CASE 1: SWTP-FWPT ................................................................................................................................. 49
FIGURE 6-8 TERMINAL FWPT UNCERTAINTY LESS FOR CASE 2 ................................................................................... 50
FIGURE 6-9 TERMINAL FOPT UNCERTAINTY LESS FOR CASE 1 .................................................................................... 50
FIGURE 9-1 CASE 1: SWTP PARETO PLOT IN REGIONS ................................................................................................ 59
FIGURE 9-2 CASE 2: NSWTP PARETO PLOT IN REGIONS .............................................................................................. 60
FIGURE 9-3 3D PARETO PLOT CASE 1: INCLUDE SEA WATER TRACER PRODUCTION DATA ........................................ 60
FIGURE 9-4 CONTROL CASE -C FOPT-SWTP SHOWS SAME TREND AS CASE 1: SWTP ................................................. 61
FIGURE 9-5 CASE 1: SWTP FOPT FULL PLOT ................................................................................................................ 62
FIGURE 9-6 CASE 1: SWTP FOPT FULL PLOT ................................................................................................................ 62
FIGURE 9-7 WWPR PRO-1-BEST FIVE MODELS OF EACH CASE .................................................................................... 64
9-8 WWPR PRO-4 BEST FIVE MODELS OF EACH CASE ................................................................................................. 64
9-9 WWPR PRO-5-BEST FIVE MODELS OF EACH CASE................................................................................................. 65
9-10 WWPR PRO-11-BEST FIVE MODELS OF EACH CASE............................................................................................. 65
9-11 WWPR PRO-12-BEST FIVE MODELS OF EACH CASE............................................................................................. 66
9-12 WWPR PRO-15-BEST FIVE MODELS OF EACH CASE............................................................................................. 66
9-13 WBHP PRO-1-BEST FIVE MODELS OF EACH CASE ................................................................................................ 67
9-14WBHP PRO-4-BEST FIVE MODELS OF EACH CASE ................................................................................................ 67
9-15 WBHP PRO-5-BEST FIVE MODELS OF EACH CASE ................................................................................................ 68
9-16 WBHP PRO-11-BEST FIVE MODELS OF EACH CASE .............................................................................................. 68
9-17 WBHP PRO-12-BEST FIVE MODELS OF EACH CASE .............................................................................................. 69
9-18 WBHP PRO-15-BEST FIVE MODELS OF EACH CASE .............................................................................................. 69
9-19 WBHP PRO-1 CASE1 VS. CASE2 MOST PARETO OPTIMAL MODELS .................................................................... 70
9-20 WBHP PRO-5 CASE1 VS. CASE2 MOST PARETO OPTIMAL MODELS ................................................................... 70
9-21 WBHP PRO-4 CASE1 VS. CASE2 MOST PARETO OPTIMAL MODELS ................................................................... 71
9-22 WBHP PRO-11 CASE1 VS. CASE2 MOST PARETO OPTIMAL MODELS ................................................................. 71
9-23 WBHP PRO-12 CASE1 VS. CASE2 MOST PARETO OPTIMAL MODELS ................................................................. 72
9-24 WBHP PRO-15 CASE1 VS. CASE2 MOST PARETO OPTIMAL MODELS ................................................................. 72
9-25 WWPR PRO-1 CASE1 VS. CASE2 MOST PARETO OPTIMAL MODELS .................................................................. 73
9-26 WWPR PRO-5 CASE1 VS. CASE2 MOST PARETO OPTIMAL MODELS ................................................................... 73
9-27WWPR PRO-5 CASE1 VS. CASE2 MOST PARETO OPTIMAL MODELS .................................................................... 74
9-28 WWPR PRO-11 CASE1 VS. CASE2 MOST PARETO OPTIMAL MODELS ................................................................. 74
9-29 WWPR PRO-12 CASE1 VS. CASE2 MOST PARETO OPTIMAL MODELS ................................................................. 75
9-30 WWPR PRO-15 CASE1 VS. CASE2 MOST PARETO OPTIMAL MODELS ................................................................. 75

viii

9-31 FIELD OIL PRODUCTION TOTAL ........................................................................................................................... 76

Tables
TABLE 5-1 LIST OF PARAMETER FOR PUNQS3 RESERVOIR .......................................................................................... 33
TABLE 5-2 DISTRIBUTION OF PARAMETER .................................................................................................................. 34
TABLE 6-1 MISFIT OF FIELD OIL PRODUCTION TOTAL FOPT FROM TRUTH CASE ........................................................ 47
TABLE 6-2 MISFIT OF FIELD WATER PRODUCTION TOTAL FWPT FROM TRUTH CASE ................................................ 47
TABLE 6-3 MISFIT FOPT FROM P(50) AS A MEASURE OF SPAN OF UNCERTAINTY ENVELOPE .................................... 48
TABLE 6-4 MISFIT FWPT FROM P(50) AS A MEASURE OF SPAN OF UNCERTAINTY ENVELOPE ................................... 48
TABLE 9-1 COMPARISON OF FRONT ADVANCE BY REGION ........................................................................................ 58

1 Aims
Automatic history matching involves the use of mathematic optimization algorithms to determine
perform reservoir model calibration. However useful, it never yields unique single answers nor
eliminate totally the uncertainties associated with the reservoir model. The value of a well
calibrated reservoir model is the reduction in uncertainty or increase in reliability of the
performance data it provides to the decision process, involving the huge financial resources
invested to exploit hydrocarbon reservoirs. Hence, this study is a quest for improvement to the
history matching process.
It has been proposed by several previous studies that since injected sea water carried
complementary information on flow paths within the offshore reservoirs, its use as an additional
constraint in history matching could greatly improve the process, yielding highly reliable models
and reducing uncertainty in forecasts. In this view, the specific objectives of this study are:

To carry out a comparative history match of a synthetic reservoir model in two cases, one
constrained additionally by production data of injected sea water.

To effect the generation of sea water production data using the equivalent of natural water
tracers in the reservoir model.

To establish using an appropriate measure, if the addition of injected sea water production
data in one case has impacted in it, a better performance of the history matching process.

To establish using an appropriate measure, if the addition of injected sea water production
data in one case has impacted in it, a reduction in the size of uncertainty associated with
the forecasted performance of the reservoir.

To make conclusions on the potential of injected sea water production data in improving
reservoir history match results

2 Introduction
2.1 Reservoir History Hatching
It is common to use numerical simulators to predict the performance of a reservoir. The
reliability of a reservoir simulation result is dependent on the inputs to the reservoir model
description. This input can be classified as relating to static (geological) properties description, or
dynamic (fluid flow) properties. Such information is gathered in the course of exploration and
appraisal. Static data will include time independent information derived from cores, wire line
logs, seismic surveys, etc. Dynamic data are time dependent data derived from flow relations,
they relate to reservoir properties such as relative permeability, fluid saturations, viscosity, flow
rate, fractional flows, etc. (Cheng et al., 2004, p.1). It is impossible to eliminate all uncertainties
in a reservoir.
History Matching is a major technique for calibrating the reservoir model in order to
maximize reliability of simulation results. It is the fine tuning of estimated reservoir description
parameters to match known past performance of the reservoir such as fluid rates, well bottomhole pressures, field average pressures, etc. History matching is an inverse problem, we attempt
to use the observed data about a reservoir to predict its properties. As is typical of inverse
mathematical problem, the solutions are never unique (Cunha, Prais, & Rodrigues, 2002, p.1).
Conventional history matching involves manual variation of field description parameters.
Simulation runs a made for each variation of model parameters, and the simulation results are
compared with the historical values. This is expensive in terms of human labour and computing
time. It is also highly subjective as the iteration direction depends on experience and insight.

2.2 Automatic History Matching and Mathematical Optimization


Several History matching techniques have been studied and applied in the quest to
automate the process of finding solutions to history matching problems. They take the approach
of treating the inverse problem as a mathematical optimisation problem, in which a defined
objective function is either maximized or minimized. This objective function takes the form of a
function of the difference between observed history data and simulated result data (Cunha et al.,
2002).
Cunha et al. (2002, p.2) also indicates that automatic history matching can be broadly
classified into two groups, gradient based techniques and stochastic techniques. (Sarma,
Durlofsky, Aziz, & Chen, 2007, p.1) identified the streamline based history matching technique
as a class of its own. Each method has its own limitations and strengths.
The deterministic or gradient based techniques uses gradients of the mathematical model,
related to the parameterised properties of the model, to minimize the objective function which is
based on misfits between historical data and simulated results (Cunha et al., 2002, p.2). They are
known to converge very fast. However, they are poorly adapted to the multi-modal and nonunique nature of solutions to history matching problems. Sarma et al.(2007) and Cunha et
al.(2002) agree that gradient based minimization is easily trapped into local minima point.
Stochastic history matching techniques have the exact opposite properties to gradient
techniques. They require a large number of simulation, hence, convergence and computing time
is quite significant. However, they are not easily trapped in local minima point, rather they effect
a more efficient search of the solution space. Sarma et al. (2007, p.1) noted that stochastic
techniques more easily honour complex geological models as they treat the simulator as a black

box. The Streamlined based history matching techniques are limited by their inability to model
complex physics.
2.2.1 Construction of Inverse Problems and Mathematical Optimization Problems
As has been discussed earlier, history matching an inversion process where historical
production data of a reservoir is used to improve the estimates of parameters which characterize
Subject to the conditions:
the reservoir. Sarma et al.(2007, p.3) expressed the general construction of history matching
problem as mathematical optimization as follows:

Initial Conditions expressed as values X0

EQ. 2-1

Where

Y refers to model parameter to be estimated; Yprior refers to the initial parameter estimates
C refer to the covariance which with Yprior is determined from the initial geological model
X refers to the states of the reservoir at various time N in Simulations. Such that f n (Xn+1,
Y) simply refers to production data.
gn (Xn+1, Xn, Yn) represents the equations to which the simulator constrains the
reservoir model, linking the Parameters Y and reservoir states or results X.
The Lagrangian Ln (Xn, Yn) is the estimate of error between the observed data Dobs and the
simulated result fn (Xn+1, Y) also referred to as Misfit M.
is the variance of the data

2.3 Value of improved history matching


History matching is applied to calibrate a reservoir model. Calibration has the singular
purpose of ensuring that simulations results are very reliable. Simulations results are applied to

field development planning, field optimization, economic evaluation of fields, testing of solution
ideas for field exploitation. These activities all have an economic value in the chain relevant to
production of the hydrocarbon.
It is impossible to accurately model every aspect of a real reservoir, hence every
simulation results has an attached uncertainty. Where history matching has been effectively
applied to a reservoir model, the uncertainty in the simulation results can be significantly
reduced. Statistical techniques can be used to quantify uncertainty.
The economic value of reservoir history matching lies in the reduction of uncertainty.
Uncertainty reduction facilitates the decision making process and risk management. The value of
improved history matching will stem from having results with fewer solution models and a direct
reduction in the uncertainty. The economic value of history matching will vary for each reservoir.
It is possible to quantify this economic value if we can quantify the reduction in uncertainty of
economic parameters resulting from the application.

2.4 Tracing Injected Sea Water in Hydrocarbon


2.4.1 Origin of Oilfield Water
The US Geological survey indicates that ground water constitute only about 1.7% of the
all the water on earth (U.S. Geological Survey, n.d.). Ground water in oil fields come from
various sources (Collins, 1975, p.194) classified as follows:

Meteoric Water: - Water th t h

Sea Water: - This refers to water from modern sea, used for water flooding.

Interstitial Water:- Water occupying the pore spaces in formation rock (aquifer water).

Connate Water: - Refers to interstitial water of syngeneic origin with the formation rock..

ntly b n in tmo ph i

l tion

Digenetic Water: -Water with chemical or physical change from rock sedimentation.

The injected sea water applies for offshore field, where sea water being the most available water
source, is used for water flooding. Produced water from an oil reservoir, for the purpose of this
study will be grouped into two, namely: Injected Sea Water and Formation Water. The formation
water is made up of connate water, aquifer or interstitial water and digenetic water.
2.4.2 Composition of Oil Field Water
The water in modern sea is generally saline. The salt or ion composition is mainly of
hlo id (Cl), odi m (N +), sulphate (SO24), m gn i m (Mg2+), calcium (Ca2+), and
potassium (K+). They constitute about 90 percent of all the salt in sea water. While inorganic
carbon, bromide, boron, strontium, and fluoride constitute the other major dissolved content of
seawater. MacKenzie (2013) gives a full list sea water composition.
Formation water cannot be ascribed a single composition as they come from various
sources and pass through various physical and chemical processes. In a study of water from
various geological aged rocks, Collins, (1975, p.216) concluded that the water were not of the
same chemical composition, and have evolved considerably compared to the modern sea water.
Any water in the reservoir can be modified by four major processes, dilution by meteoric
water or fresh water, reaction with minerals in the rock formation, clay membrane filtration and
ion exchange, mixing of sea water and aquifer water resulting in precipitations.
2.4.3 Compatibility of Formation Water and Sea Water
Vazquez, McCartney, et al.(2013,p.1) observed that injected sea water and formation
water
can be quite incompatible for mixing. Mixing of both waters could result in several possible
geochemical reactions which may lead to scale precipitation. Precipitation of insoluble

compounds result in formation damage through reduction of permeability and porosity of the
reservoir. Collins (1975, p.367) identified the ions responsible for formation of scale from water
mixing as Ca+2,Sr+2, Ba+2, Fe+2, SO4-2, HCO3- . The time and actual precipitation of the scale may
be subject to other environmental changes such as pressure and temperature changes or factors
that affect concentration of the brines. The most notorious of the scales is barium sulphate
BaSO4 which is highly insoluble and often impossible to remove once formed.
2.4.3.1 Conservation of Natural Tracers within Reservoir
Valestrand et al.(2008, p.2) defined tracers as inert chemical or radioactive compounds
used
to label fluids or track fluid movements. Artificial water tracers are used for inter-well tracer
tests. The interest of this study lies on natural water tracers. Even though ions sea water are
affected by chemical activities, Huseby et al., (2009, p.2) indicated that in most cases ions in sea
water only react moderately with the formation water. Such ions can be used as natural tracers of
sea water. Ions which may be used for such application include SO42-, Mg2+, K+, Ba2+, Sr2+, Ca2+,
Cl- (Huseby et al., 2009, p.2).
The second option for natural tracers of water are isotopes. Hydrogen isotopes are the
best being abundant in water. Another isotope is Strontium

87

Sr, a radiogenic isotope found in

high concentration in potassium rich rocks (Huseby et al., 2009, p.2). The high concentration is
transferred to formation waters with which such rocks have equilibrated. The ratio of 87Sr to the
more abundant 86Sr isotope can be used as tracers for formation water.
The choice of natural water tracers might be an economic decision rather than a choice
based on quality. Ion content data of produced water are routinely analysed as part of the flow
assurance, hence has little extra acquisition cost compared to isotopes. For this study, the

assumptions is that there are scale risks in our synthetic reservoir which exclude the use of SO42as a tracer. The alternative choice is the use of Cl- ions as tracers. These ions do not move in
between reservoir phases and are not subject to portioning effects (Valestrand et al., 2008, p.2).

3 Review of Concepts and Studies


3.1 Stochastic Optimization Techniques for History Matching
A major problem of history matching techniques is the possibility of existence of local
minima in the solution space (Hajizadeh, Demyanov, Mohamed, & Christie, 2011 p.211). It has
earlier been mentioned that gradient based history matching techniques lack the ability to
navigate through such a parameter space without being trapped in a local minima. Secondly,
accurate forecast of future reservoir performance is very important for decision making.
Hajizadeh et al.(2011, p.210) reported that simple optimization techniques have been
found to be inadequate for history matching problems. He further noted that even the Monte
Carlo Approaches were not intelligent enough for the optimization task. These problems speak to
a need for powerful optimization techniques which is able to navigate through multi-modal
parameter space to identify fitting solution models, as well as execute such task in feasible time.
Stochastic optimization algorithm have these characteristics and several have been developed
since the 1990s. Schulze-riegert & Ghedan (2007, p.1) lists some of them as follows:
Evolutionary Algorithms; Gradient techniques; Response surface modelling and optimisation on
the response surface; Hybrid schemes which couple different optimisation techniques; Ensemble
Kalman Filter Techniques. More recent algorithms were described by (Hajizadeh et al., 2011,
p.212) as follows:
Evolution Algorithms

Evolutionary Strategies (ES); Genetic Algorithm[6,7]; Differential Evolution (DE) [14]

Swarm Intelligence Algorithms

Ant Colony Optimisation (ACO); Particle Swarm Optimisation (PSO) ;Neighbourhood


Algorithm

10

3.1.1 Genetic Algorithm


The genetic optimisation algorithm is a form of evolutionary algorithm based on
techniques

or

concepts

of

natural

evolution

as inheritance, mutation, selection, and crossover.

and

the

genetic

system,

such

An optimisation problem is defined, its

objective function is defined as the fitness function and the solution models sought are encoded
with chromosome or a string of bits such as binary numbers (Obitko, 1998). Each bit is related to
a parameter of the solution model by a function (Tatiana Tambouratzis, 2013, p.163).
It starts with the initiation of a generation of randomly generated solutions to the problem.
The process of selection follows, in which the solution are ranked based on specified fitness
criteria such as the value of the objective function. The best solutions of the population are
selected to breed a new generation. Breeding of a new set of solutions is achieved through
genetic operations. Two or more parent solution models are randomly selected from the set of
bests for the generation. There chromosomes or encoding are treated as genetic identities on
which the genetic operations are performed. Popular operations are mutation and crossover,
others include regrouping, migration, extinction, roulette wheel selection, elitism, etc. (Tatiana
Tambouratzis, 2013, p.163). A generation ends when its population size is reached. The objective
function is observed to approach closer to the target solution or fitness value with each new
generation.
3.1.2 Differential Evolution
Differential Evolution is a form of Evolution Algorithm, hence it shares the same
description and procedures with Genetic Algorithm, but differs in method of evolution. Creation
of the new generation occurs before selection is done. It used uses a specified mutation and
recombination operation to create the new generation. Rather than chromosomes, differential

11

algorithm uses vectors of solution model parameters to represent solutions in the parameter space
(Hajizadeh, 2011, p.71). For mutation, three solutions from the population are randomly selected
(x1g, x2g, x3g where g denotes the generation). A mutation or mutant vector is calculated as
follows.
(Hajizadeh, 2011, p.72)

EQ. 3-1

The factor F controls the rate of evolution, ranging from 0 to 1. In recombination or cross over
operation, each member of the population
and offspring
for the parent
retained as

is crossed with the mutated vector

to produce

. The selection process is then implemented in which the objective function


and offspring

are compared, the one with lower objective function is

for a minimization problem (Hajizadeh et al., 2011, p.218).

EQ. 3-2

3.1.3 Ant Colony Optimisation


Swarm Intelligence algorithms are optimisation techniques based on the collective
behaviour organism such as insects, birds, etc. (Leonor Melo Francisco Pereira, 2013, p.179).
Swarm intelligence algorithms are the most recent of optimisation algorithms, they include the
Ant Colony Optimisation, Bee Colony Optimisation, and Particle Swarm Optimisation.
The Ant Colony Optimization algorithm is based on the behaviour of ants as they search
for food. Ants leave pheromones to mark their paths so other ants can follow them as they search
for food. The strength of the pheromone markers decay with time, the more frequently a path has
been taken, the stronger the accumulation of pheromones from various ants, and the more likely
this path will be taken by other ants. The longer the path taken is, the more time for the
pheromones to dissipate. Hence, the less likely other Ants will follow the path. The information

12

shared through pheromone distribution allows ants to translate from random exploration to the
shortest path to the nearest food source. This behaviours of ants are modelled in optimization as
solution searches by a number of ants in a colony. When one ant finds a better solution based on
defined criteria, a pheromone like property of its search influences the other ants to search more
in the vicinity of the solution found (Hajizadeh et al., 2011, p.214). This however risks trapping
the process in a local minima, hence some parameters or weighting are applied to the decision
criteria so that the ants at times deviate from the pheromone informed decisions in order to create
a better exploration of the solution space.

3.1.4 Particle Swarm Optimisation


Particle Swarm Optimization (PSO) was originated by Kennedy J, Eberhart R in 1995
Hajizadeh et al., 2011, p.219). It is an optimization techniques based on the movement of flock of
birds and school of fishes. A set of particles, each representing a solution to the defined problem,
is randomly initialized. Subsequently two factors influence the next movement of each particle as
they search for better solution:
1. pbest: - a particles best known position in terms of the optimization target.
2. gbest: - a global best known position from among the swarm of particles.

They are updated in each iteration and used to calculate the particles next movement as follows:
Updating:
{

EQ. 3-3

where

EQ. 3-4

Velocity:
(

EQ. 3-5

13

EQ. 3-6
(Hajizadeh et al., 2011, pp. 219-220)

Where
and
are as defined earlier
denotes a particular particle in the swarm
is the velocity of the particle of index in a specified dimension for the next
iteration in a multi-dimensional solution space.
and
are randomly generated real numbers between 0 and 1
and
are weighting used to control the focus of exploitation of a local area versus
exploration of the solution space.
is a weighting factor which controls the rate of convergence of the algorithm
represents the iteration count

The particle swarm optimization has various variants from this basic definition. They may differ
on the method of updating the velocity, choice of gbest, updating of particle position, etc. Below
is a flow diagram of the process.

Start

Initialize Z particles in
Solution space i=1

Do for Particle i, for


i=1 to Z
Yes i=Z

Update Particle i s
position

Calculate Particle i s
new position using
EQ.3-5

Evaluate Particle s
Position for fitness

Compare current
Particles position with
pBest

Next Iterations
Set i=1

No

Is stopping
criterion met?

Yes
Update gBest

Update the Value of


PBest

End

Figure 3-1 Work Flow for Particle Swarm Optimization

14

3.1.5 Use of Particle Swarm Optimization for Reservoir History Matching Optimization
Problem
(Hajizadeh et al., 2011) conducted a comparative study of various stochastic optimization
algorithms. This study was conducted on two relatively well known reservoir models, the Teal
South reservoir and the PUNQS3 reservoir model with 45 parameters. The authors considered the
following algorithms: Differential Evolution- Best Variant; Differential Evolution; Particle
Swarm Optimization; Ant Colony Optimization; Neighbourhood Algorithm. The study reached
the following conclusions (Hajizadeh et al., 2011, p.238):
1. All the stochastic algorithms performed well compared to gradient based algorithms.
2. That for all the algorithms studied, Differential Evolution-Best and Particle Swarm
Optimization had the fastest convergence, as well as achieved the lowest misfit solutions.
3. That all the algorithms had uncertainty bounds which included the truth case
Ant Colony Optimization had the smallest span of uncertainty, followed by Particle Swarm
Optimization, these differences are very small. Several other studies have been conducted using
particle swarm optimisation (Lina Mohamed, Christie, & Demyanov, 2011), (Linah Mohamed,
Christie, & Demyanov, 2009), (Hajizadeh, 2011b), (Arnold, Vazquez, Demyanov, & Christie,
2012), (Vazquez, MacMillan, et al., 2013), (Vazquez, McCartney, et al., 2013).

3.1.6 Multi-Objective Particle Swarm Optimization.


History matching problems will often have to consider different kinds of data. For
example history matching of a reservoir may involve data from the well bottom-hole pressure,
gas oil ratio, production volumes or rates, water cuts, etc. These data come in different numerical
ranges, a summation of all into a misfit definition will result in those with high numerical range
overshadowing those with low numerical range. An inefficient optimization will result.

15

Where F is the objective function.


3-7

EQ.

The solution to this problem for single objective optimization is to apply weights c1, c2
and c3 (see EQ. 3-7) in order to control the contribution of misfit components. Determination of
values of these weight is debateable. An alternative approach to this problem is the use of multiobjective optimization. Data of different numerical range and type are separated different
objectives for concurrent optimization. It eliminates arbitrary combination of dissimilar data into
misfit functions. In a recent study, Lina Mohamed et al.(2011) conducted a comparative study of
Single Objective Particle Swarm Optimization (SOPSO) and Multi-Objective Particle Swarm
Optimization (MOPSO). The study was conducted as a history matching task on IC Fault
Reservoir model from Imperial University, UK. The authors concluded as follows:

That the MOPSO was twice faster than SOPSO in convergence and obtained good fitting
models to the history matching problem.

That MOPSO obtained a more diverse set of solution models compared with SOPSO.

That while SOPSO gave a narrower uncertainty range, MOPSO resulted in a more robust
and more accurate uncertainty definition which included the truth case.

This study will use Multi-objective Particle Swarm Optimization for the mentioned benefits.
3.1.7 Optimal Solutions in Multi-Objective Optimization: Pareto Front
In single objective optimization problems, an optimal solution is selected based on the
value of the misfit function. In multi-objective optimization the task of selecting optimal
solutions becomes a bit more complex as the solutions represent different trade-off between
objectives in terms of dominance (Lina Mohamed et al., 2011,p.2). Veldhuizen & Lamont (1997,

16

p.2) indicated that one of the key ideas for dealing with this problem of selecting optimal
solutions is the concept of Pareto Optimality. Pareto optimal solutions stems from the definition
of Pareto Dominance.
Let us view two solutions of a multi objective optimization problem as vectors

and

, where

there are p objective functions.


and
is said to dominate

if and only if

is partially less than

(for minimization problems).

i.e
{

EQ. 3-8

Based on this definition we now define the pareto optimal solution set as solutions that are not
dominated by any other solution. This means that for N number of evaluated solutions, solution
is pareto optimal if
{
{

}
EQ. 3-9

The pareto set when plotted graphically is called the pareto front. It physically represents the
range of optimal trade-off between the objectives.

3.2 Uncertainty Quantification for Forecast


3.2.1 Definition of Uncertainty
Uncertainty is the lack of assurance about the truth of a statement or the exact magnitude
of an unknown measurement or number (Schulze-Riegert & Ghedan, 2007, p.2). Uncertainty
may result from an actual lack of knowledge, a difficulty in measurement or errors in
measurement.

17

For a reservoir, the major driver of uncertainty is heterogeneity of the reservoir. Most
reservoir data are only reflective of measurements at the well location. Inferences have to be
made about locations in between wells. Heterogeneities occur at all scales of the reservoir, from
microscopic pore scale to the megascopic properties.
Heterogeneity in reservoir characterisation translates to uncertainty in simulation outputs
and forecasts. Uncertainties in reservoir characterisation were grouped by Schulze-riegert &
Ghedan (2007, p.7) as follows.
Uncertainty in Geological Data: Uncertainties due to measurement errors, selection or
interpretation of geological data of the reservoirs
Uncertainty of Geological Data: Uncertainties inherent from the complexity of reservoir
geology or lithology. Issues and interpretation from of sedimentation, lithology, and mapping.
Uncertainty in Dynamic Reservoir Data: Uncertainty in properties that affect the flow of
fluids.
Uncertainty in Reservoir Fluids Data: Composition of reservoir fluids retains some
uncertainties as to the extent to which obtained samples are representative of the whole field.
The many sources of uncertainties means that it is impossible to totally eliminate uncertainty
from reservoir model. It is imperative that we quantify the uncertainty in the results of reservoir
models and simulation to better inform of the limits of their applicability.

3.3 Uncertainty Quantification Method: Neighbourhood Approximation Using Bayes


Theorem (NA-Bayes)
Bayesian framework was used for uncertainty quantification in this study. It has been
applied severally in in similar studies by several scholars (Christie, Demyanov, & Erbas, 2006),
(Christie, Subbey, & Sambridge, 2002), (Sambridge, 1999), (Hajizadeh et al., 2011).

18

It is important to note that uncertainty quantification is also necessitated by the


inadequacy of misfit values as a measure of accuracy of a solution model. By definition misfit is
a function of the difference between observed values and simulated values of reservoir
performance.
EQ. 3-10
However it is not just simulated values that are subject to inaccuracies, observed values are
subject to measurement errors. Accounting for this error, we refine the misfit definition as
] (Christie et al., 2006, p.145)
EQ. 3-11
Hence in actual fact, misfit values are the difference in errors associated with observed data and
simulated data. Misfit may not reflect direct relations to the truth value of a reservoirs
performance.
3.3.1 Bayes Theorem
Bayes theorem is a method of inference which allows us to update the probability
estimate for a hypothesis as added evidence is acquired about the hypothesis (Christie et al.,
2006, p.4).
EQ. 3-12

Bayes rule stated above can be summarised as stating that


P(H) - Prior
probability of the hypothesis before any Evidence; P(H|E) - Posterior probability of
hypothesis given the evidence; P(E|H)-Likelihood of the evidence given the hypothesis or
Probability of observing the evidence in the event of the hypothesis being true; P(E) Marginal Probability of the evidence independent of any particular hypothesis.

19

Relating Bayes rule to reservoir history matching problems, the models M under investigation are
the hypothesis H, the historic production data or Observation O is the Evidence. We can rewrite
the Bayes Rule using the new notation as follows:
EQ. 3-13

(Christie et al., 2006, p.3)


is the posterior probability, while
probability of the evidence

is the prior probability. The marginal

which is the denominator is now expressed as the Bayesian

integral. The Evaluation of the posterior probabilities entails the evaluation of three elements
1. Prior Probabilities

2. Likelihood of
Observations

3. Bayesian integral expressing


Probability of the observation

The prior probabilities are evaluated from the prior information obtained from the reservoir on
the variability of the parameters describing the uncertain properties of the reservoir.
3.3.2 Likelihood of Observation
The evaluation of the likelihood is based on Gaussian error statistics. The likelihood is defined as
the negative exponent of the misfit between observations and simulation values. This is
expressed in mean squares form below:
(Christie et al., 2006, p.4) or
(Sambridge, 1999, p.3)

EQ. 3-14
EQ.

3-15
(Sambridge, 1999, p.3). Where C is the covariance matrix of the observation data. For a single
parameter the definition of misfit simplifies as

EQ. 3-16

20

Where do is the observed data, g (m) is a function of the model, or simulation result for the
model.
3.3.3 Bayes Integral: Resampling by NA-Bayes
A Bayesian integral is contained in the definition of the normalizing factor

in the

expression for Bayes rule. Bayesian integrals are evaluated using Monte Carlos integration:

EQ. 3-17

Using Monte Carlos Integration

Where

EQ. 3-18

is the density distribution of the sampled models drawn from the solution space in

the forward solution earlier described? Difficulties arise in evaluating the relation as the density
distribution with which the models space is sampled in the optimization setups are usually
unknown. The solution to this problem is to re-sample the solution space in such a way that the
density distribution of the samples equals the probability distribution, hence they cancel out.

EQ. 3-19

This can be executed using Neighbourhood Approximation Algorithm


3.3.4 Neighbourhood Approximation and MCMC Walk
Makov Chain Monte Carlos random walk have the unique property that when used to
sample a given probability distribution based on set rejection criteria, the Markov Chain will
have a distribution

that is equal to the probability distribution of the sampled ensemble.

This property is what we need to evaluate the Bayesian integrals using Monte Carlo integration.
However, one of the impediments to using this approach is that we do not have a full
detailed description of the probability distribution in the solution space. The ensemble of models

21

are only representative. This is where neighbourhood approximation of the solution space comes
in. Using voronoi cells, the entire volume of the solution space is described by an approximation
of the actual probability. Voronoi cells are nearest neighbour regions in the solution space
defined around each model in the solution ensemble (see figure 3-1), they have the properties of
being space filling polyhedral, with their size, shape and volume automatically adapted to the
distribution of the models in the ensemble (Sambridge, 1999a, p.4). The spacing filling attribute
allows the points within a voronoi cell to be assigned a probability equal to the probability of the
model around which it is defined.

Figure 3-2Voronoi cells for ten random points (models) is a solution space. b. The updated
Voronoi cells after 100 points are sampled and interpolated using Gibbs sampler.
The sampling of the approximate probability distribution formed using voronoi cells is now
accomplished using the MCMC random walk. The MCMC variant used is the Gibbs Sampler
Algorithm. The Gibbs Sampler selects a model by taking random sized step in the direction of
each dimension of the solution space in turn (Sambridge, 1999). When it has stepped in all
dimensions, a parameter vector representing a model results. The Gibbs sampler implement

22

typical MCMC rejection check on each step, this turns the sampling to select models of high
probability, hence following probability distribution in the solution space (Sambridge, 1999a,
p.5).
3.3.5 Bayesian Credible Intervals
A

Bayesian credible interval is an interval in which the probability of find a truth case

or say the solution model is

(Levy, 2007). It offers a convenient way of expressing

uncertainty.

3.4 Review of Empirical Studies on Improving History Matching by Adding Data on


Injected Sea Water Production
One of the more recent studies on using tracer information for improved history matching
of a reservoir was carried out by Valestrand et al.(2008). The study investigated the effect of the
use of partitioning gas tracer data in the history match of a reservoir for the determination of the
permeability and transmissibility Multipliers. This study was carried out on the synthetic
reservoir with various realistic features coupled. Ensemble Kalman filter (EnKF) was used in
updating the reservoir properties permeability and transmissibility. The estimation using the gas
tracer data was successful in estimating transmissibility and permeability that provide as good
match to history data, the control case without the gas tracer data was not successful in this
regard. Valestrand et al.(2008) concluded that partitioning gas tracer data was of crucial
importance in successfully estimating the reservoir properties.
Huseby et al.(2009) carried out a similar study using the Ensemble Kalman filter method.
Their study however was focused on water tracers and considers ordinary and natural water
tracers. The study was conducted using the model 2 of the 10th SPE Comparative Solution
Project. The model is derived from a part of the North Sea Brent Sequence. The inversion

23

problem was to estimate the reservoir permeability and porosity. Three cases of history matching
were considered: using oil rate and water cut production data only; using the natural tracer SO42with oil rate and water cut production data; and using ordinary inter-well tracer data. The three
cases were compared based on mean square error of their estimation the true porosity and
permeability of the reservoir. The result showed that for both porosity and permeability, there
were only slight differences between estimations done without tracers, and with ordinary tracers.
However, the estimations done using natural tracers showed a marked improvement in quality
with much lower error values.. The study also noted the lack of explanation for the better
performance of natural tracers than ordinary tracers, since the former do not carry
complementary information on water injection sources. The authors concluded that tracer data
were underexploited as a source of knowledge about reservoirs.
Arnold et al.(2012) carried out a study on the value of adding produced water chemistry
as tracer of injected sea water, to further constrain the history matching of the PUNQS3 reservoir
model. The study was carried out using single objective Particle Swarm Optimisation. It
considered well bottom-hole pressure, oil production rate, gas oil ration, water production rate
and well tracer data. The cases with tracer data and without tracer data were compared. Based on
misfit calculated as mean square error from the history date, it was found that out of five trials of
each case, only one case with tracer data achieved very low misfit. A second comparison was
made on the bases of clustering of the solution models in parameter space, and did not find any
considerable improvement due to tracer data. However the study found that tracers reduced the
number of acceptably matched minima points from the parameter space. The study concluded
that adding tracer data did not harm nor greatly improve the quality of the history match, but
made significant improvements to forecast.

24

Vazquez et al.(2013) extended the methods of the study by Arnold et al.(2012) to an


actual real field, the Janice Field. The reservoir model provided for the study was history match
by the operator. The aim of the study was not strictly comparative, the study analysis found that
the conventionally matched model was mismatched to produced sea water fraction of the three
wells considered. Based on this, it identified new geological uncertainties in the reservoir model
and a new history match was carried out using Particle Swarm Optimisation and sea water
production tracer data. In terms of sea water fraction, it achieved a significantly better match for
one well, and slightly better match for the remaining two well. Using the Bayesian framework,
the study made uncertainty quantifications for the field forecasts, and developed scale risk
assessment for the field.
Vazquez et al.(2013) conducted the most recent study on effect of the composition of
produced water on history matching a one dimensional reactive reservoir model. The study was
not comparative of the non-tracer, rather it investigated connectivity and reactions between the
producer and injector in two cases which differed based on the amount of produced water
chemistry data available. It concluded that produced water chemistry had the potential to provide
the information sought on dispersivity of the connection flow paths between the wells, it also
considered it a successful application of Particle Swarm Optimisation to a reactive reservoir
model.

3.5 Problem Statement


The problem to be investigated by this study is the determination of the extent of improvements
to reservoir history match and uncertainty quantification induced by the used of produced water
chemistry data to specify historic production of injected sea water fractions. It has been proposed
by several studies reviewed in the preceding section that since injected sea water carried

25

complementary information on flow paths within the reservoir, its addition as a constraint to
reservoir calibration should improve the quality of history match and also reduce the amount of
uncertainty in forecast. Specifically this study aims to answer the following questions.
Question 1: Does adding natural tracer data reduce the mean square error misfit achieved by
sampled models with reference to oil rate and bottom-hole historical data?
Question 2: Does adding natural tracer data generally reduce the range of uncertainties specified
by Bayesian credibility intervals over the history match and forecast period?
Question 3: Does adding natural tracer data reduce the range of uncertainties specified by
Bayesian credibility at the terminal point of the forecast period?
While none of the earlier works has sought to compare the effect of adding natural sea water
tracer data to history matching by measuring the range of uncertainties, Arnold et al.(2012) had
observed that it made improvements to forecast generally. The comparison by misfit of sampled
models has been a bit more complicated due to earlier studies use us single optimization
techniques, which requires addition or removal of tracer data points before misfit values could be
compared (Arnold et al., 2012, p.6). Also, the combination and weighting of the objective
function for oil rate, water rate and natural chemical tracers for sea water in earlier studies
prefixes the combinational relationship between these objectives. While this does not hamper the
optimisation schemes from finding good solution models, it does limit the extent of exploration
in the search for good fitting models. This study will be carried out using multi-objective particle
swarm optimisation to allow a free comparison of misfit values and also maximise the space
searched by the optimisation algorithm. The choice of optimisation algorithm has also been
informed by the earlier reported works on comparative study of optimisation techniques (Linah

26

Mohamed et al., 2009) and multi-objective particle swarm optimisation (Lina Mohamed et al.,
2011). This value of improving history matching has been earlier discussed in the introduction.

27

4 Data Summary: PUNQS3 Reservoir Model


4.1 Origin
The PUNQS3 is a synthetic reservoir model based on an actual reservoir developed by Elf
petroleum. The synthetic case was initially developed for the PUNQ (production forecasting with
uncertainty quantification) project sponsored by the European Community. In the PUNQ project
ten partners from industry, research institutes and universities are collaborating on research on
uncertainty quantification methods for oil production forecasting (Soleng, 1999, p.1). It has
however become a benchmark for testing methods in history matching and uncertainty
quantification (Arnold et al., 2012, p.2).
The reservoir model consist of 19x28x5 grid blocks, of which 1761 blocks are active. It is
bounded to the east and south by a fault, the north and west are linked to a strong aquifer. It also
includes a gas cap, while six well are located around the gas oil contact. There were no injector
wells since the reservoir had a strong aquifer support.
The production scheduling is based on the real reservoir. Wells are under production constraint
based on flow. The scheduled flow periods are for a first year of extended well testing, followed
by a three year shut-in period, before field production commences. During field production, two
weeks shut-in period for each year is included for each well to collect shut-in pressure data. Total
production period is for 16.5years or 6025 days.

4.2 Available Reservoir Description


The geological description of the reservoir as provided by imperial college is given below
(G ologi l D

iption fo PUNQS3 R

voi Mod l, n.d.).

The layer thickness is of the order of 5meters in thickness, it played a major role in geological
interpretation. The sediments were deposited in a deltaic, coastal plain environment. Layers 1, 3,

28

and 5 consist of fluvial channel fills encased in floodplain mudstone. Layer 2 represents marine
or lagoonal clay with some distal mouthbar deposits; and layer 4 represents a mouthbar or
lagoonal delta encased in lagoonal clays.
L y

1, 3, nd 5 h v lin

k of high-porous sands ( > 20 %), with an azimuth

somewhere between 110 and 170 degrees SE. These sand streaks of about 800 m wide are
embedded in a low porous shale matrix ( < 5 %). The width and the spacing of the streaks vary
somewhat between the layers. A summary is given in the table below.
In layer 2 is a marine or lagoonal shale in which distal mouthbar or distal lagoonal delta occur.
They translate into a low-porous ( < 5%), shaly sediment, with some irregular patches of
somewhat higher porosity ( > 5%).
L y

4 ont in mo thb

o l goon l d lt

within l goon l l y , o

flow nit i expected

which consists of an intermediate porosity region ( ~ 15%) with an approximate lobate shape
embedded in a low-porosity matrix ( < 5%). The lobate shape is usually expressed as an ellipse
(ratio of the axes= 3:2) with the longest axis perpendicular to the paleocurrent (which is between
110 and 170 degrees SE).

Layer Facies
Width
Spacing
Channel Fill
800 m
2-5 km
1
Lagoonal Shale

2
Channel
Fill
1000
m
2-5 km
3
Mouthbar
500-5000 m
10 km
4
Channel Fill
2000 m
4-10 km
5
Figure 4-1 Expected facies with estimates for width and spacing of major flow units

29

5 Methods
5.1 Work Flow Diagram
Start

Defin e Histo ry
Data to be
Matched

Defin ition of Study


Objectives-PUNQS3
Rese rvoir Model

Defin e Uncertain
Parameter

Parameterisatio n of
PUNQS3 Rese rvoir
Model

Generat ion of
Simulation Files

Determination of
Range of Uncertain
Parameters

Generat ion of
Distribution Files
for Optimisation

Modification of
PUNQS3 Model to
in clud e Water
Injectors an d Water
Tracers

Generat ion of History


Data from Truth
Case: Blind to
Parameterisatio n

Determine I nitia l
Values o f
Varia nce for
History Data

Determination of
Optimisation
Objective Function s

Case1 PWC:
Selection of History
Data to Mat ch

Include Sea
Wate r Production
Rate History

Execute Three Runs


of PSO Optimisation
at Converging
Iteration Nu mber

Exclude Sea
Wate r Production
Rate History

Case2 No PWC:
Selection of History
Data to mat ch

Setup 3 ensembles
jo intly for forecast and
Unce rtainty
Quantification:
RAVEN

Setup Optimisation
Software: RAVEN

Execute MultiObjection
Optimisation : PSO
for 3000 Iteration

EXECUTE: Calculate
Bayesian Credibility
Intervals P10, P50 &
P90

Determine
Convergence Point:
Number of Iteratio ns

Repeat PSO
Optimisation for
Converged Num ber
of Iterations

Review Variance
Values t o Modify
Ensemble Density

Setup the 3
ensembles jointly for
forecast and
Unce rtainty
Quantification:
RAVEN

Setup Ense mble for


Forecast and
Unce rtainty
Quantification:
RAVEN

EXECUTE:
PPD Approx.
MCMC Walk

EXECUTE
PPD Approx.;
MCMC Walk;
Simulations for
forecast;

Is No. of Sa mpled
Models Adequate?

Execute Two
additional Runs of
PSO Optimisation

Figure 5-1Work Flow Diagram

EXECUTE
PPD Approx.;
MCMC Walk;
Simulations for
forecast;

EXECUTE: Calculate
Bayesian Credibility
Intervals P10, P50 &
P90

Resu lt Analysis and


Conclusions

End

30

5.2 PUNQS3 Problem: Uncertain Parameters


The PUNQS3 reservoir model was generated synthetically. The reservoirs description (see
Data Summary) was developed to match the synthetic reservoir. The reservoir model without the
porosity and permeability data, was distributed to researchers who were asked to invert the
production data to estimate the permeability and porosity. This study will be working with the
same problem configuration to determine the porosity and permeability distributions for the 5
layers on the PUNQS3 model. However our focus is to assess impact of adding injected sea
water production data as an additional constraint to the history matching method.

5.3 Modifications to the PUNQS3 and Historical Data


The PUNQS3 was initially designed without any
injection wells due to the strong aquifer support
modelled. It is imperative that sea water is injected into
the reservoir for this study, hence four Injections wells
have been added to the reservoir as shown in the figure
above. With this modification, the original history data
distributed with the reservoir model can no longer be
used for this study. These modifications necessitate the
generation of a new history data using the truth case data Figure 5-2 Injectors and Producer
Wells on the PUNQS3 Reservoir
provided at the PUNQS3 website of Imperial College Model
(G ologi l

iption

fo

PUNQS3

voi

Mod l, n.d.).

5.4 Parameterisation
The two uncertain reservoir properties were identified as porosity and permeability.

31

5.4.1 Regions and Justification/ Generation of Simulation Files


The reservoir description for the PUNQS3 indicates the existence three layers (1, 3 and 5)
which have fluvial sand channels embedded in a flood plain. The description also highlights the
direction of these sand channels as between 110 to 170 degrees azimuth. The sand channels had a
minimum width of 800m and spacing of 2 to 5 km.
For these sand channels we assume that the channels direction is mid-way between the
specified range at 145 degrees azimuth. This is to reduce the number of required parameters,
variations of 35 degrees in azimuth is not expected to have a major impact on the reservoir
performance. The parameterisation will be based on regions to capture the heterogeneity across
layers of the reservoir. Since the position of the sand channels is unknown, as region scheme has
been adopted which allows flexibility in the location and width of the sand channels, while also
minimizing the resulting of geometrically unrealistic models.
As illustrated in the Figures 5-4 to 5-8 below, the parameter regions have been defined
diagonally in the approximate direction of 145 degrees azimuth. The Larger regions whose
width(448m) is about half the minimum sand channel width(800m) described is alternated with
two smaller regions of width (256m) in order to build in reasonable flexibility on the location of
the sand channels. This is at the expense of having some poor models with sand width less than
800m. The same parameter regions is used in layers 3 and 5, being mindful that the sand
channels in in layer 5 have widths of about 2km. This arrangement was adopted with the
expectation that the best models will identify several adjacent regions to be of similar high
permeability to form the required sand body. This effect was observed for one of the low misfit
models, its first layer is visualized in Figure 5-3. There are a total of 17 parameter regions in
layer 1, 14 parameter regions in layers 3 and 5 respectively, and 18 in layer 4.

32

Figure 5-6 Layer 1


Parameterisation

Figure 5-5 Layer 2


Parameterisation

Figure 5-8 Layer4


Parameterisation

Figure 5-7 Layer 5


Parameterisation

Figure 5-4 3 Layer 3


parameterisation

Figure 5-3Sample
Optimization ResultQualitative Permeability
Map-light blue shows higher
perm.

Note that similar colours are not related. They all mark different parameterisation regions.

33

Layers 2 and 4 from the reservoir description are low permeability shale or clay. Layer 2
has no prominent flow unit, while layer 4 has a lobate shaped flow unit embedded in the lagoon
clay. Layer 2 has been parameterised as a single parameter region (See Figure. 5-4). An
arrangement of several block region has been made, with small block interconnecting them. The
idea is to be able to form an approximation of a lobate shape in many ways with several
selections. On the total, the PUNQS3 reservoir model has been parameterised into 66 parameters.
The table below lists the parameters by layer and shows the naming convention.
Table 5-1 List of Parameter for PUNQS3 Reservoir
Layer1

Layer2

Layer 3

Layer4

Layer5

$L1P
$L1P1
$L1P2
$L1P3
$L1P4
$L1P5
$L1P6
$L1P7
$L1P8
$L1P9
$L1P10
$L1P11
$L1P12
$L1P13
$L1P14
$L1P15
$L1P16

$L2P

$L3P2
$L3P3
$L3P4
$L3P5
$L3P6
$L3P7
$L3P8
$L3P9
$L3P10
$L3P11
$L3P12
$L3P13
$L3P14
$L3P15

$L4P1
$L4P2
$L4P3
$L4P4
$L4P5
$L4P6
$L4P7
$L4P8
$L4P9
$L4P10
$L4P11
$L4P12
$L4P13
$L4P19
$L4PA
$L4PB
$L4PC
$L4PD
$L4PE
$L4PF

$L5P2
$L5P3
$L5P4
$L5P5
$L5P6
$L5P7
$L5P8
$L5P9
$L5P10
$L5P11
$L5P12
$L5P13
$L5P14
$L5P15

5.4.2 Correlation of Porosity and Permeability


Hajizadeh et al.(2011) in his study indicated the existence of a correlation between
porosity
and permeability for the PUNQS3 reservoir model. The report also included a correlation
between vertical and horizontal permeability. The relations are given below.

34

These relations are used for this study to minimize the number of parameters required to fully
capture the uncertain permeability and porosity of the reservoir. The parameters will be used to
define porosity , while the porosity will be used to calculate the horizontal permeability
according to above relations. Vertical permeability

is calculated from the relations with

horizontal permeability. Permeability is assumed to be equal in all horizontal directions.


5.4.3 Parameter Distribution
The prior distribution of the parameters has been decided based on the PUNQS3 reservoir
description. The porosity variation are assumed to be uniform over the range. The range or
distribution were assigned based on the layers. Summary of distributions by layers is given
below.
Table 5-2 Distribution of Parameter
Layer
1
2
3
4
5

Distribution
Uniform
Uniform
Uniform
Uniform
Uniform

Porosity Range (%)


15-30
5-10
15-30
5-15
15 -30

5.5 Selection of Histories to Match and Objective Functions for Optimisation


This study is comparative of two cases of history match. The first case use the additional
constraint of injected sea water produced data and is denotes Case 1:SWTP (Sea Water Tracer
Production). The second case excludes the water tracer data from the history match, and is
denoted Case2: NSWTP (No Sea Water Tracer Production). Case-C is a repeate of Case 1 made
as a control.

35

5.5.1 Case 1:
5.5.1.1 Tracer Data
For this case tracers have been added to the injector wells. The same tracer is used for all
four injection wells to model the nature of natural water tracers which do not have
complimentary information on the specific injector positions. This study also assumes the
likelihood of scale formation, hence likely reactions of Ba2+ and SO42- to form BaSO4 or Barium
Sulfate. For this reason Cl- has been chosen as the better natural tracer for injected sea water. The
nature of the tracer is specified as non-partitioning between fluid phases and also exhibit the
required level of conservation from reactions in the reservoir. Based on these there is no need to
specially model the tracer transport separate from fluid flows calculated by the reservoir
simulator.
5.5.1.2 History Match Data
The history match and optimizations for this case considers combination of conventional
production data and tracer production data as listed below.

Producer Well Bottom Hole Pressure

Water Production Rate

Oil Production Rate

Sew Water Tracer Production Rate

The history match uses data for six producer wells PRO-1, PRO-4, PRO-5, PRO-11 and PRO-15.
5.5.1.3 Objective Functions
The basic definition of objective function for this study is simply the misfit of the
simulation values to the observed or history values for any given production data.

The

optimization work will minimize the misfit to achieve a match with the history data. The misfit
M i d fin d

th m n quare error as follows:

36

EQ.

5-1
is the number of data points in the history;
simulated values.

is the observed or history value,

is the

is the measurement error or variance.

5.5.1.3.1 Oil and BHP:


This first objective function combined two components, Producer Well Bottom Hole
Pressure (BHP) and Oil Production Rates. The two components are equally weighted. Each
component combined the misfit data from the six producer wells together into the single
objective function.

EQ.

5-2
The choice of combining the Producer Well Bottom Hole Pressure (BHP) and Oil Production
Rates is simply born out of their similar numerical value in the history data. They both ranged
between 100 and 250, hence equal weighting will not hamper any of them from being significant
in the optimization process.
5.5.1.3.2 Water
The second objective was defined for water production rate (WPR). It was chosen as a
Separate objective due to it being a different data from others, and having a different numerical
range from oil rate based on history. The definition combines data from the six producer wells.

37

EQ.

5-3
5.5.1.3.3 Sea Water Tracer (SWT)
The sea water tracer production rate has been defined as a separate objective with the
particular aim of making it easy to uncouple its misfit values from the data. This is to aid
comparison of this case where it is included and the Case 2, where it is not include.

EQ. 5-4

5.5.2 Case 2:
This is a repeat of the setup of Case1, without the Sea Water Tracer Production Data.
Only two objective functions have been defined for the case as follows.
5.5.2.1 Oil and BHP:
This first objective function is a replica of same objective used in Case 1.
[

]
EQ. 5-5

5.5.2.2 Water
The second objective is also an exact replica of the same objective used for Case 1.

5-6

5.6 Generation of Truth Case Histories

EQ.

38

Due to modifications to the PUNQS3 reservoir model, there was need to regenerate the
historical data. This was done using the truth case porosity and permeability data provide with
the PUNQS3 data set at the imperial college online site. The data set was setup with injectors I1,
I2, I3 and I4.
A second challenge to the reservoir set up was encounters with the water tracer data. The
PUNQS3 problem had been setup as having a history match period of 8 years, while the
remaining 8.5 years is used for forecast. The tracers were just breaking through to the producers
at the end of the first eight years. Only nine data points were nonzero for the six wells combined.
This was useful in studying effects of sea water tracer production on history match. Hence,
additional modifications have been made to the PUNQS3 model. The history match period was
extended to 11 years or 4018 days, the remaining 5.5 years was then used as the forecast period.

5.7 Variance for History Data


The variance ( ) is representative of measurement error. This was expressed as a
percentage of the maximum value for each history data type
were taken as 10% of

. Initial Estimates used

. However, the values of variance ( ) play a very important role

in the successful re-sampling of the Posterior Probability distribution using Markov chain Monte
Carlo (MCMC) methods in uncertainty quantification. For this reason, in this study, variance ( )
serves as both a factor for measurement errors and a factors for adjusting the density of sampled
models over the misfit surface for uncertainty quantifications.
The value of variance ( ) was determined by conducting sensitivities between
and

on the re-sampling using MCMC walks, a part of the uncertainity

quantification methods. The value

was found to give adequate number of

39

resampled models, equal to or above the number of iterations each in ensemble of models
generated by optimization runs. In summary, variance ( ) has been optimized for the uncertainty
quantification methods, rather than just measurement error.

5.8 Optimisation Algorithm and Setup


The use of particle swarm optimization has earlier been justified in the literature section. It
is based on the studies carried out by Hajizadeh et al.(2011) in which he conclude that Particle
Swarm optimization and Differential Evolution Best were the two best optimization techniques
among those compared. This comparison was based on the speed of convergence and the
minimization of misfit values. The adoption of multi-object optimization as earlier described is
based on the works of Lina Mohamed et al.(2011) who concluded that multi-objective
optimizations converged twice as fast as single objective optimization, gave a more diverse set of
solution models and more robust uncertainty quantification. These benefits inform the choice of
multi-objective particle swarm optimization technique for this study.
The descriptions of both particle swarm optimization (PSO) and the multi-objective
optimization have been given in earlier in Section 3. The software Raven from Epistemy Ltd was
used running the optimizations. Schlumberger Eclipse simulator was used with Raven for
running forwards solutions during the optimization process. The equation for the updating of
PSO particle position is repeated below to explain the specific setup used for the optimization
runs.
Updating:

40

EQ. 5-7

where

EQ. 5-8

Velocity:
(

EQ. 5-9
EQ. 5-10

Hajizadeh et al.(2011, p.220)

Where

is the last best solution or position of the particle i


is the global best known solution or position in the solution space
denotes a particular particle in the swarm
is the velocity of the particle of index in a specified dimension
iteration in a multi-dimensional solution space.
represents the iteration count

for the next

5.8.1 Setup of PSO run

and

are randomly generated real numbers between 0 and 1

the cognitive component is a weighting that controls the extent of exploitative local
search for solution by a given particles, it factors contribution of the particles last best
known position to next velocity (and resultant direction) of the particles.

the social component is a weighting that controls the extent of social exploration by a
particle. It factors the global best known position, hence encourages the particles to move
towards the best known global solution.

Both

and

have been set to 1.494, indicating an equal balance between social

exploration and local exploitative behaviour of the particles during optimization.

a weighting factor also called the initial inertia controls the rate of convergence of the
algorithm and has been set to 0.729

41

5.8.2 Ensemble of Models: convergence of ensemble; number of Ensembles


The number of iterations or number of models in the model ensemble affects the successful
application of Markov chain Monte Carlo (MCMC) methods in uncertainty quantification. For
this reason and to minimize computation resources the iterations are best set to the value where
full convergence of the optimization has been achieved. This point was determined by first
running an optimization of Case 1 using 3000 iterations. Secondly given that a 66 dimensions
parameter space is really big, three optimization runs have been scheduled for each study case.
This is to ensure that the solution used were obtained by searching a good portion of the solution
space.

5.9 Uncertainty Quantification:


Uncertainty quantification has been executed using the Bayesian frame work and the
Neighbourhood Approximation Bayes method. The detailed description of the process has been
reviewed in the Section 3. For each case three ensembles of models, each with a 500 models was
used for uncertainty quantification. The variance and iteration numbers used for optimisation
runs have been chosen to fit the uncertainty quantification process. The Software Raven was use
for the following process:

Determination of Prior probability of models

Approximation of the Posterior Probability Distribution (PPD) of the ensembles

Markov chain Monte Carlo (MCMC) Walk or re-sampling

Determination of Bayesian Credibility interval.

Bayesian Framework:

42

Ordinarily after resampling there is no need to conduct forward solutions of the models.
However, we are considering forecast periods which were not covered by the initial simulations
done during PSO optimisation, it is then necessary to simulate the models selected during the
MCMC walk for forecast. These simulations were executed using the Schlumberger Eclipse
Simulator (Black Oil) which was incorporated by the Raven Software.
5.9.1 Prior Probabilities
The prior probabilities were calculated based on the ranges and distributions specified for the
parameters established for the PUNQS3 model (see section on parameterisation in this chapter).
Under parameterisation, a uniform distribution was specified for all parameters. This means that
any within the range of a parameter has the same prior probability as another values in the range.
Hence, the prior probability of all the models within an ensembles will be the same.
5.9.2 Likelihood of Models
The model likelihood as reviewed in chapter two is defined based Gaussian error statistics. The
core of its evaluation is the misfit function M

(Christie et al., 2006)

The misfit functions are the same used for Particle Swarm Optimisation and have been define
within this chapter in the section on Optimisation Algorithms. The observation data used is the
same as specified for optimisation runs.

43

5.9.3 Monte Carlos Integration of the Bayesian Integral

Central to evaluation of posterior probability distribution of models is the evaluation of the


Bayesian Integral J. As described under concept reviews in Section 3, Monte Carlos integration
was used for the evaluations. The Raven software implements a Gibbs Samplers for the MCMC
walk which results in setting the density distribution

of the resampled models

approximately equal to the probability distribution density

. Effectively the Bayesian

integral is evaluated as a normed sum of the misfit function

5.9.3.1 The setup for the Markov chain Monte Carlo (MCMC) walk:
Number of Chains = 20

Burn-in Period = 10000

Length of Chain = 70000

Refresh Frequency = 10000

Note that the burn-in period is the number of initial random steps to be forgotten in order to
relieve the effects of the starting point on the sampling result of the MCMC walk. The refresh
frequency refers to the steps before the voronoi cell are updated.
5.9.3.2 Criteria for accepting the MCMC Samples
MCMC walks were considered successful when the number of selected samples was no less than
the input number of models in the input ensemble. Several trials were carried out to tune variance
for the history data and the number of models in the input ensemble for MCMC methods

44

5.9.4 Bayesian Credible intervals.


The result of the uncertainty quantification will be expressed as Bayesian credible intervals of
P10, P50 and P90.
5.10 Analysis of Results
The results analysis is based on the earlier established research questions as follows.
5.10.1 Research Question 1
Does adding natural sea water tracer data reduce the mean square error misfit achieved by
sampled models with reference to oil rate and bottom-hole pressure historical data?
One of the three model ensembles from each optimisation case is arbitrarily chosen for this
analysis. The use of multi-objective optimisation means that misfit values can only be compared
for the two cases across the pareto front. Comparison of the two cases is made on the bases of the
lowest misfit values for Oil and & BHP and Water respective, across three sections of the pareto
front. The expectation is that Case 1: SWPT will show pareto front with lower misfit values for
both Oil& BHP, and Water to indicate improvements due to inclusion of produced sea water
tracer data. Further the most pareto optimal models are identified for each case and the quality of
the history match is compared.
5.10.2 Research Question 2
Does adding natural sea water tracer data reduce the range of uncertainties specified by
Bayesian credibility at the terminal point of the forecast period?
The Bayesian credibility intervals P10, P50 and P90 is calculated as part of uncertainty
quantification. These intervals for the two cases will be compared by using two measures.

Means square errors for the three intervals relative to truth case for Field Oil Production
Total and the Field Water Production Total respectively.

Means square errors for the P10 and P90 intervals relative to P50 interval for Field Oil
Production Total and the Field Water Production Total respectively

45

This first measure captures how close the intervals are to the truth case, the second measure
captures the width of the P10-P90 interval as a measure of how large the uncertainty estimated is
over the whole production period of the reservoir. A reduction in uncertainty is expected to pull
P10, P50 and P90 closer together.
5.10.3 Research Question 3
Does adding natural sea water tracer data reduce the range of uncertainties specified by
Bayesian credibility at the terminal point of the forecast period?
It has also been considered that the effect of produced sea water history in reducing uncertainty
could be more pronounced only at the end of the production period. Hence, the uncertainty
interval P10, P50 and P90 at the production terminal point of 16.5 years is compared to the truth
values for the two cases (1 and 2) using Japanese candle stick plots. The expectation is that Case
1: SWTP will have a reduced span of uncertainty due to inclusion of produced sea water tracer
data in the history match.

46

6 Results
The results of the comparative history match study (for the Case 1: SWTP and Case 2: NSWTP)

6.1 Research Question 1


Does adding natural sea water tracer data reduce the mean square error misfit achieved by
sampled models with reference to oil rate and bottom hole pressure historical data?
6.1.1 Identification of Pareto Solutions Models for Case 1 and Case 2

Case 1: SWTP Pareto Plot


25

Water

20
15
Remainder

10

Pareto

5
0
0

10

15

20

OIL & BHP

Figure 6-1 Case 1: SWTP Pareto Plot

Case 2: NSWTP Pareto Plot


18
16
14

Water

12
10
8

Remainder

Pareto

4
2
0
0

10

15

20

OIL & BHP

Figure 6-2 Case 2: NSWTP Pareto Plot


Nine pareto models were identified for Case 1, and Eight were identified for Case 2.

47

Comparison of Paretos for Oil &BHP vs. Water Objective for Case 1 and Case 2

Comparison of Paretos by Trade-Off Regions


3

N
2.5

0.3929

Water

D
1.5

C
1
0.2903

0.5

0.3448
1.2285

0
0

X-

0.2

0.4

0.6

0.8

4
Case 1 Pareto

1.2

1.4

1.6

1.8

OIL & BHP

Case2 Pareto

C-Case Pareto

Figure 6-3 Comparison of Pareto by Trade-Off Regions


The Pareto front for Case 1 has advance ahead of case 2, hence has lower misfits data

6.2 Research Question 2


Does adding natural sea water tracer data generally reduce the range of uncertainties specified
by Bayesian credibility intervals over the history match and forecast period?
Table 6-1 Misfit of Field Oil Production Total FOPT from Truth Case
Table 6.2.1 Misfit of Field Oil Production Total FOPT from Truth Case
P(10)
P(50)
P(90)
Case 2: No Sea Water Tracer Data
22.00
4.09
152.30
Case 1: Sea Water Tracer Data
12.33
4.67
141.19
Control Case Case 1
11.00
7.11
148.78

P(10)+P(50)
26.09
17.01
18.11

Table 6-2 Misfit of Field Water Production Total FWPT from Truth Case
Table 6.2.2 Misfit of Field Water Production Total FWPT from Truth Case
P(10)
P(50)
P(90)
Case 2: No Sea Water Tracer Data
24.26
116.48
272.25
Case 1: Sea Water Tracer Data
13.42
87.32
312.84
Control Case Case 1
16.10
108.37
267.23

P(10)+P(50)
140.74
100.74
124.46

48

Case1:SWTP FOPT

Case2:NSWTP FOPT

4300000

4300000

4100000

4100000

3900000

3900000

3700000

3700000

3500000

3500000

3300000

3300000

3100000

3100000

2900000

2900000

2700000

2700000

2500000
4000

4500

5000

5500

6000

6500

2500000
4000

4500

5000

5500

6000

TRUTH FOPT

FOPT (P10)

TRUTH FOPT

FOPT (P10)

FOPT (P50)

FOPT (P90)

FOPT (P50)

FOPT (P90)

Figure 6-5 Case1: SWTP FOPT

6500

Figure 6-4 Case2: NSWTP FOPT

The span of uncertainty, measured as mean square error or misfit from the truth case, is less for
Case 1: including Sea Water Tracer Production Data than for Case 2: Excluding Sea Water
Production Data.
Table 6-3 Misfit FOPT from P(50) as a Measure of Span of Uncertainty Envelope
Misfit FOPT from P(50) as a Measure of Span of Uncertainty Envelope
Case
P(10)-P(50)
P(90)-P(50)
Total
Case 2: No Sea Water Tracer Data
44.53
109.14
153.67
Case 1: Sea Water Tracer Data
31.88
96.68
128.56
Control Case Case 1
35.49
93.09
128.57
Table 6-4 Misfit FWPT from P(50) as a Measure of Span of Uncertainty Envelope
Misfit FWPT from P(50) as a Measure of Span of Uncertainty Envelope
P(10)
P(50)
Total
Case 2: No Sea Water Tracer Data
45.55
33.77
79.32
Case 1: Sea Water Tracer Data
41.67
73.24
114.91
Control Case Case 1
50.47
38.10
88.57

49

Case 1:SWTP-FWPT

Case 2:NSWTP-FWPT

1800000

1800000

1600000

1600000

1400000

1400000

1200000

1200000

1000000

1000000

800000

800000

600000

600000

400000

400000

200000

200000

0
3000
-200000

4000

5000

6000

7000

0
3000
-200000

4000

5000

6000

TRUTH FWPT

FWPT (P10)

TRUTH FWPT

FWPT (P10)

FWPT (P50)

FWPT (P90)

FWPT (P50)

FWPT (P90)

Figure 6-7 Case 1: SWTP-FWPT

6.3 Research Question 3

Figure 6-6 Case 2: NSWTP-FWPT

7000

50

Does adding natural sea water tracer data reduce the range of uncertainties specified by
Bayesian credibility at the terminal point of
TERMINAL FOPT
UNCERTAINTY AT 16.5
YEARS

TERMINAL FWPT
UNCERTAINTY AT 16.5
YEARS
1,800,000.00

4150000

1,600,000.00

4100000

1,400,000.00

4050000

1,200,000.00

4000000

1,000,000.00
3950000
800,000.00
3900000

NPWC
NPWC

PWC
P10

P90

CPWC
P50

PWC

CPWC

Truth Case

Truth Case

Figure 6-9 Terminal FOPT Uncertainty less for


Case 1
the forecast period?

P10

P90

P50

Truth
Case

Truth Case

Figure 6-8 Terminal FWPT Uncertainty less


for Case 2

7 Discussion and Conclusion


7.1 Improvement to History Matching
A comparison of the two cases was carried out by considering the advance of the pareto
front in the two cases. The identified pareto front are shown in Figure 6-1, Figure 6-2 in the
Section 6. (A close up plot of the paretos are shown in Figure 9-1 and Figure 9-2 in Section 9Appendix). The combined plot of Case 1: SWTP and Case 2: NSWTP is shown in the Figure 6-3,
it shows that Case 1 front is clearly ahead of Case 2 front in achieving lower misfit for both
objectives function (Oil Production & Well Bottom Hole Pressure and Water Production Rate).
A more detailed comparison was made by marking the plot area of Figure 6-3 into grid
based regions of approximately similar objective trade-off. The square grid regions are addressed
as columns X= 1, 2, 3, 4, 5, 6, 7 and Rows N= A, B, C, D, E. Table 9-1 in the Section 9Appendix report the result of the comparison. The objective trade-off regions can be classified

51

into three on the basis of dominance of objectives. Regions C-3, D-3, and E-3 mark portions of
the two fronts where Oil & BHP Objective dominated Water Objective. Region B-3 has the most
Pareto optimal models for all the cases. A-4, A-5, A-6 are regions where Water Objective
dominates the Oil & BHP. In all three regions Case 1: SWTP has the lower misfit as shown in
Table. 9-1. The difference in misfit achieved by Case 1: SWTP relative to Case 2: NSWTP has
been expressed as a percentage of the mean of the misfit achieved by Case: 2 NSWTP for the
dominating objective in region. This is to quantify the extent to which the Pareto in Case 2 has
advance ahead of Case 1. This difference in objective is a maximum of 4.5% in Water Objective
dominated regions and 4.6% where oil objective dominates. We hence state that Case 1: SWTP
has achieved lower misfit values across the whole pareto front, compared to Case 2: NSWTP.
However, the relative improvement is very small at a maximum of 4.9%.
Having established that adding injected sea water tracer production data improved history
matching in case 1, it becomes necessary to consider the quality of the history match. The use of
multi-objective optimisation necessitates that we only compare models with similar objective
trade-offs (however, plots of the match of the lowest total misfit models have been provided in
the Section 9-Appendix. Figure 9-7 to 9-18). The most pareto-optimal models for both cases are
compared. Iteration model 493 for Case 1 and iteration model 262 for Case 2. The plots for well
bottom-hole pressure (WBHP) are reported in Figure 9-19 to 9-24 in Section 9-Appendix, and
well water production rate (WWPR) in Figure 9-24 to 9-30 also within Section 9-Appendix. For
WBHP, Case 1:SWTP model achieved a better match for wells Pro-1, Pro-5 and Pro 12
compared to Case 2:NSWTP. It achieved an approximately equal match as Case 2: SWTP for
wells Pro-11 and Pro-15. Only Pro-4 shows a slightly better match for well Pro-4. For WWPR
Case 1 model had a better match in wells Pro-4 and Pro-12 than Case 2. An approximately equal

52

match was achieved in wells Pro-1, Pro-11 and Pro-15 where both cases missed the water
breakthrough point. Only in Well Pro-5 was Case 2 model clearly a better match. Based on these
we can state that adding injected sea water tracer production data in Case 1 produced a slightly
better history match quality than in Case 2 where it was excluded.

7.2 Economic Value


The total oil production (FOPT) was compared for the two pareto-neutral models with reference
to the truth case. Case 1 model FOPT was 9,390 bbls less than truth case, while case 2 model
FOPT was 43,7200bbl less than truth case (see Figure 9-31). The give the value of the addition of
injection sea water tracer data as 34,330bbl of oil. Assuming constant price of oil into the future,
and no depreciation in oil value due to future production, this represents $3.82 million (2013)
based on the price of Brent on 17th August 2013( il P i , 2013). This is significant given that
no new cost for the acquisition of data on natural sea tracers is acquired. Produce water
composition are usually analysed are part of flow assurance in most reservoirs with scale risk.

7.3 Reduction of Uncertainty


The reduction in uncertainty was judged based on the size of mean square error or misfit of the
Bayesian credibility intervals from the truth case reservoir performance (see Figure 6-4 and
Figure 6-5 for FOPT and Figure 6-6 and Figure 6-7 for FWPT). For FOPT, the Case 1: SWTP
credible intervals P10 and then P90 were closer to the truth than those of Case 2: NSWTP with
misfit values of 12.33 against 22 for P10 (see Table 6-1 in Section 6-Results). The same trend
was observed for FWPT, the Case 1: SWTP credible intervals P10 and P90 were closer to the
truth case than those of Case 2: NSWTP with misfit values of 13.4 against 24.26 (see Table 6-2
in Section 6-Results).

53

Consideration of the size of the uncertainty intervals was made based on mean square
error of P10 and P90 from P50, as an indicator of P10 to P90 span of uncertainty. The values of
the span of the uncertainty in FOPT for Case 1: SWTP is lower at 128.56 compared to 153.67 for
Case 2: NSWTP (see Table 6-3 in Section 6-Results). However, for FWPT Case 2: NSWPT has a
lower span at 79.32 against 114.91 for Case 1: SWTP (see Table 6-4 in Section 6-Results). This
can be explained seeing that for FWPT, the truth case is outside the P10- P90 interval (See fig. 66 and 6-7 in Section 9) for both cases. The uncertainty estimate for Case 1: SWTP has expanded
in response to lower reliability of the estimate.
The terminal span of uncertainty was also consider using a Japanese candle stick plot.
Figure 6-9 shows that for FOPT the higher credible point P10 is closer to the truth case in Case 1:
SWTP than in Case 2: NSWTP. The span of the uncertainty is also lower for Case 1. The Figure
6-8 shows that for FWPT, the higher credible point P10 for Case 1: SWTP is again closer to the
truth, which lies outside the intervals. However, the span of the uncertainty is lower for Case 2
than for Case 1. This is the same effect of the truth case lying outside the intervals.
The trends for Case 1: SWTP has been replicated by the control Case C: SWPT in all
comparisons, hence this is not a one off occurrence.

7.4 Conclusion
A comparative study of history matching was carried out for two cases, where one used the
additional constraint of injected sea water produced, which is acquired using natural sea water
tracer. The study was carried out by using a synthetic reservoir model, the PUNQS3 model.
Multi-objective Particle Swarm Optimization was used for automatic history matching of both
case with 66 parameters in the model for the estimation of porosity and permeability. The
uncertainty quantification was done using NA-Bayes method which is defined on a Bayesian

54

probability framework. The results shows that adding natural sea water tracers production data
results in slightly better history matching results. The uncertainty in the performance forecast was
also reduced by the additional data, as well as made more robust or reliable. Similar results
showing only slightly better performance and better forecast were obtained by Arnold et
al.(2012). The estimated economic value of the improvement, about $3.8 million (2013), is quite
significant, but very small compared to the value of oil in the reservoir.
In the course of this work, some unreported results shows that the addition of injected sea
water tracer production data did not significantly impact the history matching problem with a
lower number of parameter (33). It is proposed that the value of adding injected sea water tracer
production data is more significant for higher complexity history matching problems. Future
work will aimed at verifying the value of injected sea water production data to history matching
for reservoirs that are significantly more complex than the PUNQS3.

55

8 References

Arnold, D., Vazquez, O., Demyanov, V., & Christie, M. (2012, June). Use of Water Chemistry
Data in History Matching of a Reservoir Model. Proceedings of SPE Europec/EAGE
Annual Conference. Copenhagen: Society of Petroleum Engineers. doi:10.2118/154471-MS
Cheng, H., Wen, X., Milliken, W. J., & Datta-gupta, A. (2004). Field Experiences with Assisted
and Automatic History Matching Using Streamline Models. Houston: Society of Petroleum
Engineers.
Christie, M., Demyanov, V., & Erbas, D. (2006). Uncertainty quantification for porous media
flows. Journal of Computational Physiscs, 217, 143158. doi:10.1016/j.jcp.2006.01.026
Christie, M., Subbey, S., & Sambridge, M. (2002). Prediction Modeling under Uncertainty in
Reservoir. Frieberg: 8th European Conference on the Mathematics of Oil Recovery.
Collins, A. G. (1975). Geochemistry of Oilfiled Waters (pp. 1503). Amsterdam: Elsevier
Scientific Publishing Company.
Cunha, L. B., Prais, J. R. P., & Rodrigues, F. (2002). Improved Reservoir Description T hrough
Hi to y M t hing: A N w T hni
W ith Appli tion to Gi nt D pw t
ilfi ld.
Calgary: Petroleum Society.
Geological Description for PUNQS3 Reservoir Model. (n.d.). Retrieved August 16, 2013, from
http://www3.imperial.ac.uk/earthscienceandengineering/research/perm/punqs3model/onlinedataset/geologicaldescription
Hajizadeh, Y. (2011). Population-Based Algorithms for Improved History Matching and
Uncertainty Quantification of Petroleum Reservoirs. Heriot-Watt University, Edinburgh.
Hajizadeh, Y., Demyanov, V., Mohamed, L., & Christie, M. (2011). Comparison of Evolutionary
and Swarm Intelligence Methods for History Matching and Uncertainty Quantification in
Petroleum Reservoir Models *. Berlin: Springer-Verlag.
Huseby, O., Valestrand, R., Naevdal, G., & Sagen, J. (2009, June). Natural and Conventional
Tracers for Improving Reservoir Models Using the EnKF Approach. Proceedings of
EUROPEC/EAGE Conference and Exhibition. Amsterdam: Society of Petroleum Engineers.
doi:10.2118/121190-MS
Leonor Melo Francisco Pereira, E. C. (2013). Multi-caste Ant Colony Algorithm for the
Dynamic Traveling Salesperson Problem. In M. Tomassini, A. Antonioni, F. Daolio, & P.
Buesser (Eds.), Adaptive and Natural Computing 11th International Conference
Proceedings, ICANNGA 2013 (pp. 179188). Lausanne.

56

L vy, R. (2007). L t
6: B y i n P m t E tim to ; Confid n int v l . R t i v d
August 17, 2013, from http://idiom.ucsd.edu/~rlevy/lign251/fall2007/lecture_6.pdf
MacKenzie, F. T. (2013). Chemical and physical properties of seawater. Encyclopedia
Britannica. Retrieved August 16, 2013, from
http://www.britannica.com/EBchecked/topic/531121/seawater
Mohamed, Lina, Christie, M., & Demyanov, V. (2011, May). History Matching and Uncertainty
Quantification: Multiobjective Particle Swarm Optimisation Approach. Proceedings of SPE
EUROPEC/EAGE Annual Conference and Exhibition. Vienna: Society of Petroleum
Engineers. doi:10.2118/143067-MS
Mohamed, Linah, Christie, M., & Demyanov, V. (2009). Comparison of Stochastic Sampling
Algorithms for Uncertainty Quantification. The Woodlands: Society of Petroleum
Engineers. doi:10.2118/119139-MS
Obitko, M. (1998). Genetic Algorithm. Retrieved August 16, 2013, from
http://www.obitko.com/tutorials/genetic-algorithms/ga-basic-description.php
Oil Price. (2013). Retrieved August 17, 2013, from http://www.oil-price.net/
Sambridge, M. (1999). Geophysical inversion with a neighbourhood algorithm II . Appraising
the ensemble. Geophysic Journal International.
Sarma, P., Durlofsky, L. J., Aziz, K., & Chen, W. H. (2007). A New Approach to Automatic
History Matching Using Kernel PCA. Houston: Society of Petroleum Engineers.
Schulze-riegert, R., & Ghedan, S. (2007). Modern Techniques for History Matching. In 9th
International Forum on Reservoir Simulation. Abu Dhabi.
Soleng, H. H. (1999). Oil Reservoir Production Forecasting with Uncertainty Estimation Using
Genetic Algorithms. In IEEE Proceeding of 1999 (pp. pps. 12171223, vol. 2, 1999).
Tatiana Tambouratzis, V. K. (2013). The Scale-Up Performance of Genetic Algorithms Applied
to Group Decision Making Problems. In M. Tomassini, A. Antonioni, F. Daolio, & P.
Buesser (Eds.), Adaptive and Natural Computing 11th International Conference
Proceedings, ICANNGA 2013 (pp. 161168). Lausanne.
U.S. G ologi l S v y. (n.d.). Th Wo ld W t . R t i v d A g t 16, 2013, f om
http://ga.water.usgs.gov/edu/earthwherewater.html
Valestrand, R., Sagen, J., Naevdal, G., Huseby, O., & Forghany, S. (2008, April). The Effect of
Including Tracer Data in the EnKF Approach. Proceedings of SPE/DOE Symposium on
Improved Oil Recovery. Tulsa: Society of Petroleum Engineers. doi:10.2118/113440-MS

57

Vazquez, O., MacMillan, A., Arnold, D., Young, C., Demyanov, V., Fisher, A., & Christie, M.
(2013, July). Produced Water Chemistry History Matching in the Janice Field. Proceedings
of 75th EAGE Conference & Exhibition incorporating SPE EUROPEC 2013. London:
Society of Petroleum Engineers. doi:10.2118/164903-MS
Vazquez, O., McCartney, R., & Mackay, E. (2013). Produced Water Chemistry History
Matching Using a 1D Reactive Injector Producer Reservoir Model. The Woodlands: Society
of Petroleum Engineers. doi:10.2118/164113-MS
Veldhuizen, D. A. Van, & Lamont, G. B. (1997). Evolutionary Computation and Convergence to
a Pareto Front. California: Morgan Kaufmann.

58

9 Appendix
9.1 Detailed Comparison of Front Advance by Region
Table 9-1 Comparison of Front Advance by Region
Comparison of Front Advance by Region
Case 2
Case 1
Dominant Front by

ITN
Oil &
Water ITN Oil & Water SWTP Oil&BHP
REGION
BHP
Misfit
BHP
Misfit Misfit
(N-X) refer
Misfit
Misfit
to Fig. 5.1.1
E-2
495
0.452 2.199 262 0.430 2.106 0.316 Case 1
157
0.462 2.077
C-3
B-2
B-3
A-3

359

0.503

1.079

Front Lagging
262

0.525

0.632

Front Lagging

Most
Pareto
Optimal

Difference Diff. as %
Between
of Mean
Lowest
Misfit
Misfit
Water

Oil Mean=
4.34 Water
Mean=4.03
0.022
0.513%
0.075

1.739%

0.075

1.739%

0.213

4.911%

0.167

1.995%

286

0.428

1.058

458

0.290

0.838

0.351 Case1
0.497 Case 1

493

0.460

0.529

0.066

171

0.474

0.344

0.517

Case 1

0.115

2.856%

315

0.653

0.409

0.205

Case 1

0.045

1.105%

Case1

387

0.612

0.459

379

0.663

0.453

A-5

419

0.976

0.418

397

0.816

0.371

0.092

Case 1

0.048

1.182%

A-6 & A-7

100

1.194

0.358

473

1.229

0.187

0.182

Case 1

0.171

4.235%

A-4

59

9.1.1 Case 1: Excluding Sea Water Tracer Production Data

Case 1: SWTP Pareto Plot in Regions


3

2.5

Water

D
1.5

C
1

0.2903

0.5

0
0

X-

0.1875

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

OIL & BHP

7
Remainder

Pareto

To compare Pareto, by misfit, the trade-offs between objectives has been marked into ranges
by the regions. Pareto points falling into each of the frontal regions is be compared for the
two cases
Figure 9-1 Case 1: SWTP Pareto Plot in Regions

60

9.1.2 Case 2: Excluding Sea Water Tracer Production Data

Case 2: Pareto Plot in Regions


N

3
2.5

E
Water

D
1.5

C
1

B
0.5

A
0
0

X-

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8 OIL & BHP


Remainder

Pareto

Figure 9-2 Case 2: NSWTP Pareto Plot in Regions

9.2 3 Dimensional Pareto Plot for Case 1

Not
e
that the red balls represent the pareto models, the blue balls are the remainder of the ensemble.
Figure 9-3 3D Pareto Plot Case 1: Include Sea Water Tracer Production Data

61

9.3 Field Oil Production Total (FOPT) Intervals for Case1, C-Case and Case 2
Case -C FOPT-SWTP

C-Case FOPT-SWTP

4300000

4500000

4100000

4000000

3900000

3500000

3700000

3000000

3500000

2500000

3300000

2000000

3100000

1500000
1000000

2900000

500000

2700000
2500000
4000

4500

5000

5500

6000

6500

0
-2000
0
-500000

2000

4000

6000

TRUTH FOPT

FOPT (P10)

TRUTH FOPT

FOPT (P10)

FOPT (P50)

FOPT (P90)

FOPT (P50)

FOPT (P90)

Figure 9-4 Control Case -C FOPT-SWTP shows


same trend as Case 1: SWTP

8000

62

Fig. 9.3.3 Case1: SWTP FOPT

Fig. 9.3.4 Case 2:NSWTP FOPT


4500000

4500000

4000000

4000000

3500000

3500000

3000000

3000000

2500000

2500000

2000000

2000000

1500000

1500000

1000000

1000000

500000

500000

0
-2000
0
-500000

2000

4000

6000

8000

0
-2000
0
-500000

2000

4000

6000

TRUTH FOPT

FOPT (P10)

TRUTH FOPT

FOPT (P10)

FOPT (P50)

FOPT (P90)

FOPT (P50)

FOPT (P90)

Figure 9-6 Case 1: SWTP FOPT full plot

Figure 9-5 Case 1: SWTP FOPT full plot

8000

63

64

9.4 Comparison of History Match for Best Five Models of Case 1 and Case 2
Well Water Production Rate

WWPR Pro-1

cubic metre

40
35

Case 1

30

Case 1

25

Case 1

20

Case 1

15

Case 1

10

Case 1

Case2

0
-500

500

1000

1500

2000

2500

3000

3500

4000

4500

Time

Case2
Case2

Figure 9-7 WWPR Pro-1-Best Five Models of Each Case

WWPR Pro-4
250

Case 1
Case 1

cubic metre

200

Case 1
Case 1

150

Case 1
Case 1

100

Case2
Case2

50

Case2
Case2

0
-500

500

1000

1500

2000

2500

Time

9-8 WWPR Pro-4 Best Five Models of Each Case

3000

3500

4000

4500

Case2
Truth

65

WWPR Pro-5
160

Case 1

cubic metre

140

Case 1

120

Case 1

100

Case 1
Case 1

80

Case 1

60

Case2

40

Case2
Case2

20

Case2
0
-500

500

1000

1500

2000

2500

3000

3500

4000

4500

Case2
Truth

Time

9-9 WWPR Pro-5-Best Five Models of Each Case

WWPR Pro-11
40

Case 1

cubic metre

35

Case 1

30

Case 1

25

Case 1
Case 1

20

Case 1

15

Case2

10

Case2
Case2

Case2
0
-500

500

1000

1500

2000

2500

Time

9-10 WWPR Pro-11-Best Five Models of Each Case

3000

3500

4000

4500

Case2
Truth

66

WWPR Pro-12
250

Case 1
Case 1

cubic metre

200

Case 1
Case 1

150

Case 1
Case 1

100

Case2
Case2

50

Case2
Case2

0
-500

500

1000

1500

2000

2500

3000

3500

4000

4500

Case2
Truth

Time

9-11 WWPR Pro-12-Best Five Models of Each Case

WWPR Pro-15
1.2

Case 1
Case 1

cubic metre

Case 1
0.8

Case 1
Case 1

0.6

Case 1
Case2

0.4

Case2
0.2

Case2
Case2

0
-500

500

1000

1500

2000

2500

Time

9-12 WWPR Pro-15-Best Five Models of Each Case

3000

3500

4000

4500

Case2
Truth

67

9.5 Well Bottom Hole Pressure

WBHP Pro-1
260

Case 1

250

Case 1

Bars

240

Case 1

230

Case 1

220

Case 1

210

Case 1

200

Case2
Case2

190

Case2

180

Case2
170
-500

500

1000

1500

2000

2500

3000

3500

4000

4500

Case2
Truth

Time

9-13 WBHP Pro-1-Best Five Models of Each Case

WBHP Pro-4
260

Case 1

250

Case 1

Bars

240

Case 1

230

Case 1

220

Case 1

210

Case 1

200

Case2
Case2

190

Case2

180

Case2
170
-500

500

1000

1500

2000

2500

Time

9-14WBHP Pro-4-Best Five Models of Each Case

3000

3500

4000

4500

Case2
Truth

68

WBHP Pro-5
260

Case 1

250

Case 1

Bars

240

Case 1

230

Case 1

220

Case 1

210

Case 1

200

Case2
Case2

190

Case2

180

Case2
170
-500

500

1000

1500

2000

2500

3000

3500

4000

4500

Case2
Truth

Time

9-15 WBHP Pro-5-Best Five Models of Each Case

WBHP Pro-11
260

Case 1

250

Case 1

Bars

240

Case 1

230

Case 1

220

Case 1

210

Case 1

200

Case2
Case2

190

Case2

180

Case2
170
-500

500

1000

1500

2000

2500

Time

9-16 WBHP Pro-11-Best Five Models of Each Case

3000

3500

4000

4500

Case2
Truth

69

WBHP Pro-12
260

Case 1

250

Case 1

Bars

240

Case 1

230

Case 1

220

Case 1

210

Case 1

200

Case2
Case2

190

Case2

180

Case2
170
-500

500

1000

1500

2000

2500

3000

3500

4000

4500

Case2
Truth

Time

9-17 WBHP Pro-12-Best Five Models of Each Case

WBHP Pro-15
260

Case 1

250

Case 1

Bars

240

Case 1

230

Case 1

220

Case 1

210

Case 1

200

Case2
Case2

190

Case2

180

Case2
170
-500

500

1000

1500

2000

2500

Time

9-18 WBHP Pro-15-Best Five Models of Each Case

3000

3500

4000

4500

Case2
Truth

70

9.6 Comparison of History Match for Most Pareto Models for Case 1(Iteration 493
Run2) Case 2(Iteration 262 Run3)-Well Bottom Hole Pressure

WBHP PRO-1
300
250
200
150
100
50
0
-500

500

1000

1500

2000

2500

3000

3500

4000

4500

-50
Truth

Case 1-IT493

Case 2-IT262

9-19 WBHP PRO-1 Case1 vs. Case2 Most Pareto Optimal Models

WBHP PRO-5
300
250
200
150
100
50
0
-500

500

1000

1500

2000

2500

3000

-50
Truth

Case 1-IT493

Case 2-IT262

9-20 WBHP PRO-5 Case1 vs. Case2 Most Pareto Optimal Models

3500

4000

4500

71

WBHP PRO-4
300
250
200
150
100
50
0
-500

500

1000

1500

2000

2500

3000

3500

4000

4500

3500

4000

4500

-50
Truth

Case 1-IT493

Case 2-IT262

9-21 WBHP PRO-4 Case1 vs. Case2 Most Pareto Optimal Models

WBHP PRO-11
300
250
200
150
100
50
0
-500

500

1000

1500

2000

2500

3000

-50
Truth

Case 1-IT493

Case 2-IT262

9-22 WBHP PRO-11 Case1 vs. Case2 Most Pareto Optimal Models

72

WBHP PRO-12
300
250
200
150
100
50
0
-500

500

1000

1500

2000

2500

3000

3500

4000

4500

4000

4500

-50
Truth

Case 1-IT493

Case 2-IT262

9-23 WBHP PRO-12 Case1 vs. Case2 Most Pareto Optimal Models

WBHP PRO-15
300
250
200
150
100
50
0
-500

500

1000

1500

2000

2500

3000

3500

-50
Truth

Case 1-IT493

Case 2-IT262

9-24 WBHP PRO-15 Case1 vs. Case2 Most Pareto Optimal Models

73

9.7 Comparison of History Match for Most Pareto Models for Case 1(Iteration 493
Run2) Case 2(Iteration 262 Run3) - Well Water Production Rate

WWPR PRO-1
0.012
0.01
0.008
0.006
0.004
0.002
0
-500
0
-0.002

500

1000
Truth

1500

2000

2500

Case 1-IT493

3000

3500

4000

4500

Case 2-IT262

9-25 WWPR PRO-1 Case1 vs. Case2 Most Pareto Optimal Models

WWPR PRO-4
250

200

150

100

50

0
-500

500

1000

1500

2000

2500

3000

-50
Truth

Case 1-IT493

Case 2-IT262

9-26 WWPR PRO-5 Case1 vs. Case2 Most Pareto Optimal Models

3500

4000

4500

74

WWPR PRO-5
60
50
40
30
20
10
0
-500

500

1000

1500

2000

2500

3000

3500

4000

3500

4000

4500

-10
Truth

Case 1-IT493

Case 2-IT262

9-27WWPR PRO-5 Case1 vs. Case2 Most Pareto Optimal Models

WWPR PRO-11
20

15

10

0
-500

500

1000

1500

2000

2500

3000

-5
Truth

Case 1-IT493

Case 2-IT262

9-28 WWPR PRO-11 Case1 vs. Case2 Most Pareto Optimal Models

4500

75

WWPR PRO-12
160
140
120
100
80
60
40
20
0
-500

500

1000

1500

2000

2500

3000

3500

4000

4500

-20
Truth

Case 1-IT493

Case 2-IT262

9-29 WWPR PRO-12 Case1 vs. Case2 Most Pareto Optimal Models

Fig. 9.7.6 WWPR PRO-15


0.009
0.008
0.007
0.006
0.005
0.004
0.003
0.002
0.001
0
-500
0
-0.001

500

1000
Truth

1500

2000
Case 1-IT493

2500

3000

3500

Case 2-IT262

9-30 WWPR PRO-15 Case1 vs. Case2 Most Pareto Optimal Models

4000

4500

76

9.8 Field Oil Production Total

Field Oil Production Total


2620000

2600000

2580000

2560000

2540000

2520000

2500000
4000

4050
Case 1- ITN493 FOPT

9-31 Field Oil Production Total

4100

4150
Case 2 ITN 262 FOPT

4200
Truth FOPT

4250

Potrebbero piacerti anche