IPTC 10065

Markov Chain Monte Carlo Methods for Reserves Estimation

A.W. Wadsley, SPE, Curtin U. of Technology

Copyright 2005, International Petroleum Technology Conference

1. Reservoir Volumetrics, Fluid Contacts
This paper was prepared for presentation at the International Petroleum Technology
Conference held in Doha, Qatar, 2123 November 2005. 2. Material Balance, PVT, Aquifer
This paper was selected for presentation by an IPTC Programme Committee following review
of information contained in an proposal submitted by the author(s). Contents of the paper, as 3. Well Decline, Sweep Efficiency, Analogues
presented, have not been reviewed by the International Petroleum Technology Conference
and are subject to correction by the author(s). The material, as presented, does not necessarily Any estimate of reserves should be consistent with all three of
reflect any position of the International Petroleum Technology Conference, its officers, or
members. Papers presented at IPTC are subject to publication review by Sponsor Society these areas:
Committees of IPTC. Electronic reproduction, distribution, or storage of any part of this paper
for commercial purposes without the written consent of the International Petroleum Technology
Conference is prohibited. Permission to reproduce in print is restricted to an abstract of not
more than 300 words; illustrations may not be copied. The abstract must contain conspicuous
acknowledgment of where and by whom the paper was presented. Write Librarian, IPTC, P.O.
Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.

Abstract
Monte Carlo methods are used to integrate the data pertinent
to reserves estimation including material balance, production
decline, reservoir volumetrics, and petrophysics. Many of
these techniques produce independent estimates of reserves
and hydrocarbons initially in place (HCIIP): for example,
material balance and volumetric methods independently
estimate HCIIP. Similarly, independent estimates for recovery This paper shows how the Markov chain Monte Carlo
factors are obtained from production decline, analogue (MCMC) method can combine such information, and prior
reservoir studies and simulation. Traditional Monte Carlo parameter distributions, to produce consistent estimates of
methods are unable to combine such independent estimates in HCIIP and reserves. The MCMC approach is similar to the
a natural way. Markov chain Monte Carlo methods, on the acceptance-rejection method of Stoltz et al2 who used filtered
other hand, enable all such data to be integrated leading to Monte Carlo simulation3.
robust, unbiased and accurate estimates of HCIIP and reserves. Pressure History Match
The algorithms for achieving this are presented and illustrated Consider the observed pressures and material balance match
using field examples. shown in Figure 1 for a type I fractured reservoir with storage
Introduction only in the fractures.
Monte Carlo simulation for the estimation of hydrocarbon
reserves and fluids in place is a well established technique in
the oil and gas industry1. Traditionally, Monte Carlo
estimation of hydrocarbons initially in place (HCIIP) and
reserves uses samples from prior distributions of reservoir
parameters, such as gross rock volume, porosity, with these
samples being used to calculate the distribution of HCIIP or
reserves directly. Generally, apart from the range constraints
implicit in the prior distributions, no other constraints are
applied to either the calculated output distributions or input
variables.
It is likely, however, that additional quantitative
information is available which imposes constraints or Figure 1 Material Balance Match to Observed Reservoir
dependencies between variables or constraints on the Pressure
likelihood of the calculated HCIIP or reserves. For example,
reserves estimates should be consistent with material balance,
This match was achieved using a non-linear optimisation
volumetrics, decline curves and analogue or expert opinion. In
algorithm to minimise the error function:
general, there are three sets of over-lapping constraints which
impact reserves and HCIIP:
2 IPTC 10065

pkobs pkcal ( )
2
(y)
F ( ) = r ( x, y ) = min 1, . (4)
(x)
(1)
p
where represents a set of parameters; namely, aquifer
This algorithm is applied to the pressure match by selecting an
permeability, porosity, thickness, fracture compressibility, initial set parameters = {k , h, , cr , Dx, Dy, OIIP} and
perturbing these, + , as:
aquifer dimensions and oil-initially-in-place OIIP. Although +
the match is goodit captures the initial decline trend and the
final reservoir pressurebecause of the scatter in the pressure
k + k + N ( 0, k )
observations (obtained from many wells in the field and
corrected to datum) it is difficult to quantify the goodness-of- h + h + N ( 0, h )
fit from visual observation alone. Further, this match
calculates a single deterministic OIIP, which cannot be used + + N ( 0, )
directly in probabilistic analysis of reserves. (5)
Dx + Dx + N ( 0, Dx )
The Markov chain Monte Carlo method provides a
natural way of calculating the distribution of OIIP (and Dy + Dy + N ( 0, Dy )
associated aquifer parameters) as shown in the next section.
OIIP + OIIP + N ( 0, OIIP )
Markov Chain Monte Carlo
Markov chain Monte Carlo generates the target probability This defines a perturbation in the error function from which
distribution by running a cleverly constructed Markov chain the ratio r given in equation (4) is defined by
for a long time so that the limiting or stationary distribution of
this chain is the target distribution4. In the case of the error
function given in equation (1), we can stipulate that the error r=
( )
p +
(6)
should be normally distributed with zero mean and standard p ( )
deviation, p . This is achieved by using the Metropolis-
where
Hastings algorithm which simulates a Markov chain in the
configuration space x = ( , F ) composed of the parameters 1 pobs
k
pcal ( )
k
2

p ( ) = exp (7)
2 p
and the error F ( ) . Starting with any configuration
(0)
x
the Metropolis-Hastings algorithm proceeds by iterating the
1
following two steps5: Note that the normalising constant is not required in
(n) 2
1. Randomly perturb the current state x so as to (7) because it cancels in equation (6). If the perturbations
+
generate a new configuration x from a proposal given by (5) result in parameters going outside reasonable
function or distribution T x ( (n)
, x+ .) physical bounds (for example, k < 0 ) set
p ( ) = 0
+
so
that the update will not be accepted. In this example, the lower
2. Draw a uniform variate u U 0,1 and update to [ ] limit on OIIP was set to 25 MMbbl in order for OIIP to always
exceed cumulative oil production to date of 23.4 MMbbl.
the (n+1)th level as:

( n +1)
x + , if u r ( x( n ) , x + )
x = (2)
x ( n ) , otherwise
where

( y ) T ( y, x )
r ( x, y ) = min 1, (3)
( x ) T ( x, y )
and is the target distribution. The function T ( x, y ) is
called the proposal distribution for state y given state x . In
the case that T is symmetric (that is, T ( x, y ) = T ( y , x ) ,
Figure 2 Material Balance OIIP distribution using MCMC

equation (3) reduces to Figure 2 shows the calculated distribution of OIIP after
500,000 trials using the above MCMC technique. This
IPTC 10065
distribution is approximately lognormal with mean 43 MMbbl,
median (P50) 38 MMbbl and standard deviation 17 MMbbl.
The mode, or most likely value, of 35 MMbbl coincides with
the minimum residual error.
Figure 3 OIIP variation over 400,000 iterations of Markov chain
The 500,000 trials used in this example is typical of the
number of iterations required to generate a good
approximation to the underlying probability distribution using
MCMC. Figure 3 shows how OIIP varies over the simulation
from 1 to 400000 iterations. This behaviour, with oscillation
around the most likely value and then a significant move away
from the most likely value for a considerable number of
iterations is typical behaviour of a random walk. In this
example there are no constraints on the underlying parameters
apart from those required to maintain physical feasibility.
Thus the random walk can move without restriction into areas
of low probability though it ultimately moves back to the high
probability (most likely) parts of parameter space. A large
number of iterations is required in order to account for this
behaviour. Furthermore, the output is highly correlated, many
values not moving at all until the probability of change is
sufficiently high. This is seen in Figure 4 where a 5000 trial
sample of the output is shown. The character of the MCMC
trajectory is stepwise, and not random as usually seen in
Monte Carlo analysis where samples are directly taken from
the underlying parameter distributions, rather than incremental
perturbations as given here.
Figure 4 OIIP variation over a subset of 5000 iterations
The ability of the random walk to cover all of the probability
space is termed mixing. There is a trade-off between
3
efficient mixingthat is, the random walk does not spend too
much time in low probability regions of parameter space and
visits all parts of parameter spaceand increased likelihood of
rejection of the update. If the parameters are allowed to jump
randomly over their range, it is possible and indeed, highly
likely, that the resulting outcome will have very low
probability (that is, the pressure match is very poor) and be
rejected. Figure 5 shows a portion of the Markov chain for
same pressure match when the standard deviations given in
equation (5), which define the magnitude of the perturbations
of the reservoir parameters, have been increased by a factor of
5. It can be seen that the incidence of rejection has increased
significantly with the chain spending large periods of time
stuck on the same value.
Figure 5 OIIP variation for large parameter step size
Combining OIIP Estimates
The previous section calculated a probability distribution for
OIIP derived from material balance and observed reservoir
pressures. By changing the stipulated standard deviation, p ,
the OIIP range can be narrowed around the most likely value
(or value of minimum residual error). In practice, we also
generate an OIIP distribution calculated volumetrically from,
typically, gross rock volume, porosity, net-gross, elevation of
the oil-water contact (OWC) and connate water saturation. For
example, for the fractured reservoir of the previous section, a
histogram of OIIP was calculated from input distributions of
fracture porosity, elevation of OWC and effective fractured
area in the reservoir.
Figure 6 Independent Volumetric and Material Balance
Estimates of OIIP
4 IPTC 10065
This calculated distribution is shown in Figure 6 together with
the distribution calculated from material balance. There is a
significant difference in distributions with the mean of the
volumetric estimate being 85 MMbbl (compared to 43
MMbbl) with standard deviation 54 MMbbl (compared to 17
MMbbl). The most likely volumetric OIIP is 49 MMbbl,
compared to the 35 MMbbl calculated from material balance.
Simple Merge Algorithm
MCMC provides a natural way to merge these different
distributions to reflect uncertainty in both the material balance
and volumetric parametrisations. Given two distributions
f and g , these can be combined using the following MCMC
( 0) ( 0)
algorithm. Let x f and y g be initial values
sampled from the given distributions. The update from the nth
to the (n+1)th configuration uses the steps:
+ (n) + (n)
1. Draw x f x and y g y
2. Draw a uniform variate u U 0,1 and update to [ ]
th
the (n+1) level as:
+ + ( x+ y + )
x ,y , if u
x( ( x( n ) y ( n ) )
n +1)
, y(
n +1)
= (8)
( n) ( n)
x ,y , otherwise
where is the target distribution which is usually stipulated
to have zero mean. Generally set N ( 0, ) , a normal
distribution with mean zero and standard deviation .
Figure 7 Volumetric and Material Balance OIIP Estimates after
Simple Merge Algorithm
Figure 7 shows the result of applying the above MCMC
algorithm to the material balance and volumetric OIIP
distributions where the target distribution for the difference
between the OIIP estimates is normal with zero mean and
standard deviation 5 MMbbl. The distribution for the
differences is shown in Figure 8 which, in this example, has a
calculated mean 0.26 MMbbl and standard deviation 4.9
MMbbl, close to the stipulated target distribution.
The main problem with the direct implementation of
this algorithm, in this case, is that the material balance
estimate is strongly dominating the volumetric estimate, and
the material balance estimate was obtained only through a
general random walk on the reservoir parameters, without
using knowledge of the likely distribution of the these
parameters. In the fractured reservoir under consideration, the
only parameter common to both the volumetric and material
balance estimates of OIIP is effective formation porosity (if it
is assumed that fracture porosity in the reservoir is
representative of fracture porosity in the aquifer). However,
formation thickness, formation permeability, formation
compressibility can all be assigned distributions based upon
geology and core analysis. These prior distributions can be
used in both the material balance and volumetric estimates of
OIIP which can be estimated simultaneously and forced to be
consistent using a analogue of equation (8).
Figure 8 Distribution of OIIP Differences after Simple Merge
Algorithm
Random Walk over Prior Distributions
In order that the Markov chain mixes well, we should choose
small perturbations of the reservoir parameters, and then
construct a random walk over the specified prior parameter
distributions. Again, this can be simply achieved using the
MCMC formalism. Let f be the distribution from which we
wish to sample the parameter x f . Rather than sampling
directly from f we wish to form the update x+ x + x ,
where the x is a random perturbation, usually sampled from
a normal distribution with zero mean, x N ( 0, ) and
( n)
specified standard deviation, . Given the nth iterate x
update using the steps:
1. Select x = x
+ (n)
+ x N ( 0, ) .
2. [ ]
Draw a uniform variate u U 0,1 and update to
th
the (n+1) level as:
+ f ( x+ )
x , if u
f ( x( n ) )
x(
n +1)
= (9)
(n)
x , otherwise
A test example of this algorithm is shown in Figure 9 where
the target distribution was triangular with minimum 0, mode
30 and maximum 100. As can be seen, even though the
IPTC 10065 5

random walk uses a normally distributed step-size, the target

distribution is simulated correctly.

Figure 10 Integrated Merge of OIIP Estimates

Figure 9 MCMC simulation of Triangular Distribution using
Random Walk
Fully Integrated and Constrained Merging
This procedure can be implemented within the MCMC
algorithm for combining two distributions or minimising the
pressure error. Further constraints can be introduced in a
natural way. For this reservoir, fracture storativity was
determined for each well from interference test and earth-tide
analysis6. Storativity is the product of effective porosity and
total compressibility and provides a constraint on both
material balance estimates of OIIP (porosity, compressibility)
and volumetric estimates (porosity).
The full integration of the material balance and volumetric
estimates of OIIP is formulated by:
1. Select reservoir parameters from prior distributions
that have been derived from log, core and other
analyses. These distributions are used to sample
porosity, compressibility, storativity, aquifer
permeability, aquifer thickness, aquifer dimensions,
fracture extent and oil-water contact using the
MCMC update equation (9);
2. Generate the material balance estimate of OIIP using
the MCMC update given by equations (2) and (6);
3. Ensure consistency between correlated parameters
such as storativty, s = c , and the material balance
and volumetric estimates of OIIP using the MCMC
update equation (8).
The updates given in steps 2 and 3 are chained together by
defining a combined proposal function
r = rp rs rOIIP
where rp is given by equation (6), rs is given by
The distribution for a combined estimate of OIIP, using the
algorithm given above, is shown in Figure 10. It is seen that
the resulting distribution combines aspects of both the material
balance and volumetric distributions: the mean OIIP is 53
MMbbl, median 50 MMbbl. The most likely OIIP is 43
MMbbl, compared to 49 MMbbl and 35 MMbbl calculated
from volumetric and material balance estimate, respectively.
Discussion
The examples given above used MCMC techniques to
generate OIIP estimates which combined independent material
balance and volumetric estimates of OIIP into a single,
consistent distribution. The methodology is easily extended to
reserves estimation where, for example, recovery efficiency
may have been estimated independently from well decline and
sweep efficiency in the reservoir. The steps are essentially
those given previously:
1. Sample reservoir parameters from prior
distributions that have been derived from
laboratory, field observation, expert opinion,
analoges, or other sources. Use the MCMC
approach to perform a random walk over these
parameter distributions using equation (9);
2. Incorporate constraints or ensure consistency
between parameters or calculated values using the
MCMC proposal equation (8);
3. Merge independent estimates of reserves, or other
calculated values, using the proposal equation (8);
4. Update using the combined proposal function
given by a generalized version of equation (10).
The technique has been applied successfully by the author for
reserves reporting and reserves dispute resolution. Comparison
(10) with actual field performance has shown excellent agreement
between the values calculated from the MCMC methodology
and observed outcomes. Implementation of this methodology,
which rigorously creates a zone of consistency among

rs =
(
s s + + c + ) rOIIP
reseroivr parameters, provides a robust and consistent method

( )
(n) ( n) ( n)
and is given by for the estimation of hydrocarbon reserves and fluids in place.
s s c

rOIIP =
(
OIIP OIIPp+ OIIPv+ )
( )
.
OIIP OIIPp( ) OIIPv( )
n n
6 IPTC 10065

Definition of Symbols
F = error between calculated and observed reservoir pressures New Zealand, SPE 39714, 1998 SPE Asia Pacific
Conference on Integrated Modelling for Asset
= set of reservoir parameters Management, Kuala Lumpur, Malaysia.
3
pk
obs
= observed reservoir pressure at time k C. Kuhne, G. Wiggs, D. Beeson J. Madelone, M. Gardner.
Using Monte Carlo Simulation for Probabilistic Design,
pk
cal
( ) = calculated reservoir pressure at time k given parameters 4
Proceedings of the 2005 Crystal Ball User Conference.
Gilks, W.R., Richardson, S. and Spielgelhalter D.J. (1996)
p = standard deviation for pressure Introduction to Markov chain Monte Carlo. In Markov
x = current state of reservoir (as set of parameters) Chain Monte Carlo in Practice (Gilks et al.), pp. 1-19.
Chapman and Hall, London.
(n)
x = state of reservoir at time n 5
Liu, J.S. (2001) Monte Carlo Strategies in Scientific
Computing. Chapter 5. Springer-Verlag, New York.
u = uniform radnom number between 0 and 1 6
Z.Yang, G.Weir, A.W.Wadsley: Fracture Porosity
+
x = proposed new configuration for reservoir state Identification of Naturally Fractured Gas Reservoir from
the Analysis of Earth Tidal Effects. To Appear, 2005.
+
T = proposal distribution from which x is sampled
r = proposal function
= target probability distribution
y = state of reservoir
+
k , k = reservoir/aquifer permeability
+
h, h = reservoir/aquifer thickness
+
s , s = reservoir/aquifer storativity
+
c, c = reservoir/aquifer compressibility
, = porosity
+

+
Dx, Dx = x-dimension of aquifer
+
Dy , Dy = y-dimension of aquifer
+
OIIP, OIIP = oil originally in place
p = error function for distribution for pressure
s = error function for distribution for storativity
OIIP = error function for distribution for OIIP
N ( , ) = normal probability distribution with mean and variance
2

U [ 0,1] = uniform probability distribution between 0 and 1

f , g = probability distributions or density functions
r = proposal function
rp = proposal function for pressure error
rs = proposal function for storativity error
rOIIP = proposal function for OIIP error

References
1
McCray, A.W. (1975) Petroleum Evaluations and Economic
Decisions. Prentice-Hall, New Jersey.
2
L.R.Stoltz, M.S.Jones and A.W.Wadsley: Reserves
Identification in the Fractured Limestone, Waihapa Field,