Journal of Petroleum Science and Engineering 86-87 (2012) 225–236

Contents lists available at SciVerse ScienceDirect

Journal of Petroleum Science and Engineering

journal homepage: www.elsevier.com/locate/petrol

New upgridding method to capture the dynamic performance of the fine scale
heterogeneous reservoir
Mohammad Sharifi ⁎, Mohan Kelkar
McDougall School of Petroleum Engineering, The University of Tulsa, 800 S Tucker Drive, Tulsa, OK 74104, United States

a r t i c l e i n f o a b s t r a c t

Article history: Many tight gas reservoirs are characterized by large gross interval with significant presence of shale and
Received 29 June 2011 interspersed sand. The geo-cellular models of these reservoirs, which honor the resolution of log data,
Accepted 14 March 2012 contain large number of layers. Without significant upgridding of vertical layers, flow simulation of such
Available online 27 March 2012
reservoirs is very difficult. Ideally, the upgridding process should be conducted such that the dynamic
performance of geo-cellular model is preserved in the upgridded model. One way to ensure efficient upgridding
Keywords:
upgridding
is to preserve fine scale pressure profile in the coarse scale model.
upscaling In this study we observed that, in heterogeneous reservoirs, the vertical communication across layers has a big
vertical communication impact on the flow behavior in a fine scale model. In the presence of vertical communication, fine scale
pressure equilibrium heterogeneities are homogenized. That is, layers with different permeabilities have similar pressure profiles
tight gas reservoir during the depletion phase. Hence by combining these layers into a single layer would still predict a similar
dynamic performance. In contrast, when crossflow across layers does not exist, combining those layers would
result in much different dynamic behavior than fine scale behavior. It was observed that upgridding
performance is not only influenced by the extent of sand connectivity across the layers (which tells us about
vertical communication) but also how areally dispersed the connections are present. In the case of existence
of dispersed shale within reservoir it was found that the vertical flow is more helped by smaller sized vertical
connections between layers, which are areally dispersed, than large sized connections concentrated within a
certain region. Analytical solution is provided to quantitatively prove the effect of size as well as dispersed
nature of connection across the layers. Using the underpinnings of analytical solution, a new methodology is
developed which combines layers based on sand connectivity (vertical permeability) across adjacent layers as
well as their respective petrophysical properties (horizontal permeability).
Our approach is analytical. However, it was shown that the analytical upgridding error we estimate closely
mirrors the simulation error which represents the difference in the dynamic performance of fine versus coarse
layers. We validate our method by applying it to both synthetic as well as field cases. Through these examples,
we demonstrate that newly proposed method is superior to proportional as well as variance and flux based
upgridding methods.
© 2012 Elsevier B.V. All rights reserved.

1. Introduction A typical reservoir simulator can handle 10 6 to 10 7 grid blocks.

The number of grid blocks is a function of the production mechanism,
Conducting computationally efficient and an accurate simulation heterogeneity of reservoir and even post-processing of the simulation
of a reservoir model is the first priority of any simulation engineer. data for history matching purposes. It is increasingly important that
Over the last decade, the development of detailed three dimensional we not only predict the history, but also, uncertainties in the future
geologic descriptions has become common-place within asset and performance. This requires generating multiple reservoir models for
project teams worldwide. As the computational speed has improved, history matching purposes. Creation of multiple models requires
both geo-cellular and simulation models have become bigger. arriving at a compromise between a very detailed, history matched,
However computational requirements for geo-cellular models are single model versus multiple coarser models which are reasonably
always smaller than that of the corresponding flow simulation history matched and can predict the future performance uncertainties
models. Hence, there is a need to upscale the geo-cellular models. (Ates et al., 2005; Taware et al., 2011). That is, as long as coarse models
are able to properly resolve production data, it is better to use coarser
models which can capture the future uncertainties.
Therefore, reducing the number of grid blocks in a geo-cellular
⁎ Corresponding author. model is very important. The key to upscaling the geological model is
E-mail address: mohammad-sharifi@utulsa.edu (M. Sharifi). to retain sufficient amount of heterogeneities in the model so that

0920-4105/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.petrol.2012.03.016
226 M. Sharifi, M. Kelkar / Journal of Petroleum Science and Engineering 86-87 (2012) 225–236

the dynamic performance of the fine scale model is captured. Deter- similar fluxes are combined to reach an appropriate number of coarse
mining the optimum number of grid blocks in the coarse scale model grid blocks (Durlofsky et al., 1997). In general, the procedure tries to
to capture the dynamic performance is an active area of research. maintain the resolution in an area which has high fluxes and lump the
In the literature, the process of reducing the number of reservoir layers with low fluxes. Flow based methods may provide a better
layers is called upgridding and determining the properties of the upgridding procedure for a given well configuration, but the methods
generated grid is called upscaling. Different upscaling and upgridding require using fine scale simulations which may be difficult when the
procedures are appropriate in different situations. The ideal procedure fine scale model contains millions of cells. Further, the method may
to use on a particular problem depends on the simulation question be optimal for a given well configuration, but would be sub-optimal
being addressed, the production mechanism, and the level of detail under other well configuration scenarios.
that can be accommodated in the coarse model. For a problem
involving primary production with only oil or gas being produced, 2. Problem statement
the coarse model should correctly capture the effects of near well
heterogeneity as well as the general large scale flow response of the To summarize the current state of the art in upgridding, the static
reservoir. For scenarios involving displacement of oil by water or gas, methods are computationally efficient but may not reproduce
it may be important to accurately capture the effects of main flow dynamic characteristics of the fine scale model well; whereas, the dy-
paths between injection and production wells. This may require the namic methods are computationally demanding and may not provide
use of specialized gridding procedures (Durlofsky, 2005). Many optimal upgridding without precisely knowing the well configuration
upgridding and upscaling techniques are available in the literature as well as the flow processes under actual field conditions.
(Fincham and Christensen, 2004; Stern, 2005). These methods vary Our goal in this research is to address the deficiencies in the
from simple statistical methods to using advanced flow based current static and dynamic models. We want to develop a method
simulation techniques (Durlofsky, 1996; Durlofsky et al., 1994; King, which accounts for the dynamic characteristics of the fine scale
2006). These methods can be divided in two main categories. Those model, yet is efficient and robust enough that it can be used for various
methods that are based on static analysis of the reservoir properties well configurations as well as different flow processes. In this paper, we
are called static based methods and those that involve using flow only concentrate on primary depletion process.
simulation are called dynamic methods. In this research, our goal is to develop an optimum algorithm for
upgridding to reduce the number of layers of fine scale geological
1.1. Static based methods models for building a simulation model. Instead of focusing on static
properties, our focus is on the dynamic characteristic of the fine scale
Proportional and variance based methods are the two common model (i.e. production profile) so that the upgridded model will
static based methods used in the industry. The simplest approach is capture the dynamic behavior of the fine scale model. The method
a proportional method. In this method, the numbers of grid blocks has to be efficient so that it can be applied quickly for a large number
in vertical direction are combined in a fixed proportion without of grid blocks. In addition, the method has to be dependent on the
considering their underlying properties. As an example, for reducing flow process. To summarize, we want to develop an upgridding process
100 vertical layers to 20 vertical layers, we will simply combine five which has the following characteristics:
fine scale layers into one upscaled layer. Although this method is
currently used in the industry, the dynamic performance of the (1) It can capture most important dynamic characteristics of the
upscaled model may not reproduce the fine scale performance fine scale model;
(Hosseini and Kelkar, 2010). (2) It is robust and efficient so that it can be applied quickly and for
For overcoming this deficiency, some researchers used the variance various well configurations;
based method. Testerman (1962) was the first one who implemented (3) It would capture primary depletion process in actual field
such a method. Li et al. (1995), Li and Beckner (2000, 2001), Stern conditions.
and Dawson (1999)and more recently King et al. (2006) used variance
based method for upgridding purposes. In all of these methods, layers 3. Proposed method/hypothesis
with similar static characteristics are combined into a coarse layer.
The logic of static based method is to preserve the variance or hetero- The main hypothesis in the current methodology is that by
geneity of the fine scale model by isolating layers which are dissimilar. preserving the pressure distribution in the fine scale model, dynamic
The key assumption behind these methods is that key static properties characteristics of the fine scale model can be preserved (i.e. fluxes).
influence the dynamic characteristics; therefore, preservation of The potential or the pressure distribution determines the flow
heterogeneities would result in better reproduction of dynamic behavior; therefore, by developing an upgridding scheme which
performance. However in many instances, this assumption does not captures the pressure distribution, it can also capture the dynamic
hold true. As an example, in the case of a three layer model with characteristics of the coarse scale model.
permeabilities of 1, 10 and 30 md, the variance based method would To determine the pressure profile in the fine scale model, we
combine 1 and 10 md layers first since the difference is smaller; intend to use an analytical method. By nature, any analytical method
however, simulation of a gas production indicates that combining 10 requires some model simplification. Therefore, second hypothesis in
and 30 md layers would preserve the fine scale behavior better our research is that the assumptions we make for deriving the
(Hosseini and Kelkar, 2010). analytical model are valid for actual field conditions.
An actual fine scale model consists of three dimensional distribu-
1.2. Dynamic based methods tions of grid blocks of various sizes. To determine the combination of
the layers, we will need to determine the errors between the layers.
Dynamic based methods rely on simulating the fine scale model The error is defined later. However, imagine a column of grid blocks.
under realistic conditions and then determining the upgridding of We take two adjacent grid blocks and calculate the error. The error is
this model so that dynamic performance is preserved. One of such the difference between the dynamic behavior of an upscaled grid
methods is a flux based method that was suggested by Durlofsky block (which combines the two fine scale grid blocks) versus dynamic
(1996). In that method first the single phase flow in fine scale behavior in the two fine scaled grid blocks. Since we are not conducting
model is run and then the fluxes are output from each layer. Once any simulation, we are representing the dynamic behavior by average
the basic structure of coarse grid is specified, fine grid blocks with pressure within fine scale versus coarse scale grid blocks. Once we
 M. Sharifi, M. Kelkar / Journal of Petroleum Science and Engineering 86-87 (2012) 225–236 227

Fig. 1. Simple 3 layer model.

calculate the error for every pair of fine scale grid blocks within two All the integrals are calculated between time zero and ∞ that is our
adjacent layers, we add the error for all the fine scale grid blocks within final simulation time. The reason we used infinity is because we
two layers to calculate error between two layers. wanted our analytical expression to be utilized in upgridding process.
By considering only a limiting case, our solution becomes independent
3.1. Approach of time and hence easy to use. The subscripts 1, 2 and 3 represent the
first, second and third layers respectively and subscripts 12 and 23
Reservoir simulation solves a system of nonlinear partial differential represent the combined layers 1 and 2, and 2 and 3 respectively. For
equations for pressure and saturation distributions in the reservoir. The pseudo-steady state flow and single phase liquid, Eqs. (1) and (2) can
driving force in any reservoir is pressure; this can be supplied by be solved analytically. We used the upper limit as infinity so that we
external forces such as injection wells or by internal energy of the can get an expression which is independent of time. The detailed
reservoir fluids (e.g., expansion of fluids) itself during the primary derivation is shown in Appendix A. The final expression for the error
depletion. In our derivation, the focus is on a simple, single phase can be written in a simple form: i.e.,:
flow. The single phase can be either an oil or gas.



φ φ

EIk;kþ1 ¼

k − kþ1

: ð3Þ
3.2. No crossflow scenario K k K kþ1

In this section, we propose a method based on a single phase flow That is, the best combination of layers is the one which minimizes
with two additional assumptions. We assume no crossflow between error in Eq. (3). The details are provided in Section 3.6. This result is
the layers (Kz = 0) and a pseudo steady state condition in the reservoir. in agreement with the result that was observed by Kim et al. (2010).
We start our model based on a simple three layer reservoir with Please note that in real 3-D reservoir models, there are no homoge-
different petrophysical properties in each layer and a well at the neous layers. Therefore, the error will have to be calculated grid block
center of the model. This is shown in Fig. 1. by grid block in vertical direction and the errors for all the grid blocks
We define the error as the difference between the average within two adjacent layers have to be combined to calculate the total
pressures in fine layers during depletion compared to average error between the two layers.
pressure in coarse layer during the depletion. That is, if we combine
layers 1 and 2 the error term (Error12) will be the difference between 3.3. Analytical solution for partial communication scenario
the average pressures of layers 1 and 2 in fine scale model with the
average pressure that will be generated in a coarse model. It is As mentioned before, Eq. (3) is based on no crossflow. In reality,
worth mentioning that after combining any two layers, the horizontal reservoir rarely contains layers with no crossflow. Many tight gas
and vertical permeabilities of a coarse layer are calculated by reservoirs contain combination of shale and sand. Sand grid blocks
arithmetic and harmonic averages of corresponding fine layer model may communicate vertically, whereas, shale will impose a barrier.
permeabilities respectively. After combining layers 1 and 2 or 2 and Depending on the percentage of shale and sand, partial blocking in
3 we can define these errors mathematically as: vertical direction is likely. Higher the shale percentage, less is the




vertical communication.

∞

∞
The two layer or three layer problem with partial vertical commu-
error1; 2 ¼
∫0 ðp1 −p12 Þdt
þ
∫0 ðp2 −p12 Þdt
ð1Þ
nication has no analytical solution without simplifying assumptions.
In this section, we extend our expression to account for the cross
flow effect. Here the effect of a barrier between two porous media






∞

∞
in the presence of two small openings is examined. In Fig. 2 well_A
error2; 3 ¼
∫0 ðp2 −p23 Þdt
þ
∫0 ðp3 −p23 Þdt
: ð2Þ
is producing at a constant flow rate q = 1000 STBD and is located at

Fig. 2. Effect of connection distribution on flow across the barrier.

228 M. Sharifi, M. Kelkar / Journal of Petroleum Science and Engineering 86-87 (2012) 225–236

the distance of 400 ft from the barrier (D0 = 400 ft). The distances
between openings 1 and 2 from the well_A are D1 and D2 respectively.
The distance between points 1 and 2 is called DL. For solving this
system analytically we simulate the connections as injection wells
(Zhao and Thompson, 2002). Both flux and pressure at the connections
are changing with time. Then pressure equation for the system that
consists of five unknowns is solved analytically (pressures and flow Fig. 4. 3 layer model containing shale barrier and two distant connections.
rates at points 1 and 2 and bottom hole pressure at well _A). The
detailed calculation procedure is given in Appendix B. We do not use behavior of combined layers (upgridded) would be similar to the fine
the equation derived in Appendix B directly in our solution procedure. scale model when significant vertical communication exists. This also
However, it is shown to illustrate the importance of opening in a sheds light on the correct upgridding process. If significant vertical
barrier on the flow behavior of a well. This provides us with an analog communication exists, heterogeneous layers can be combined under
when we are considering vertical communication through discontinu- the depletion process; whereas, two relatively similar layers may not
ous shale barriers. be combined if vertical communication does not exist across those
A production result (from other side of barrier) from analytical layers.
solution is shown in Fig. 2. It can be seen that only by having two Another important consideration in incorporating the vertical
small connections along the barrier, at late time, about 40% of the pro- communication between the layers is how well distributed the
duction of the well_A is coming from the other side of the barrier vertical communication is across the layer. That is, if the two adjacent
through points 1 and 2. Also, when the distance between the points 1 layers are communicating only in one part of the area between the two
and 2 is bigger (DL= 300 ft) in Fig. 2b, the amount of supply from layers, the vertical communication is not as effective. On the other
other side of the barrier is greater than the case where the connection hand, if there are many discontinuous regions where vertical commu-
points are closer to each other (DL= 10 ft) in Fig. 2a. nication exists, it would be lot more effective in “homogenizing” the
two layers. This is what we observed in our analytical solution as
3.4. Effect of crossflow in 3 layer model well as simple three layer model. We want to validate this observation
by considering more complex situations. For simplicity, we built a three
In the Section 3.3 we showed that existence of opening within the layer model with the top layer with 200 md and bottom layer with
barrier can cause more homogenization. In this section we are going 10 md. The middle layer is a shale barrier except that we created few
to see effect of the existence of connections between two layers that openings in that layer by assigning non-zero permeability and very
are separated by an impermeable layer in the middle. small porosity in that barrier. Fig. 6 shows the permeable paths in the
In Fig. 3 we have a three layer model including a well that operates shale barrier. As can be seen, various configurations of openings in
under constant bottom hole pressure. This is similar to Fig. 1 except shale barrier are considered. In the first figure (top, left) no opening
the middle layer is a shale barrier with two small openings which exists; in the second figure (top, middle) only one grid block within
connect the two layers. So, the middle layer is not contributing to the shale barrier has nonzero vertical permeability, in the third (top,
flow except providing communication across top and bottom layers. right), a total of 36 grid blocks have nonzero vertical permeability. In
The well is fully perforated. The top layer has a higher permeability the bottom left 16 grid blocks have nonzero vertical permeability for
than the bottom layer. The middle layer is shale except two connec- communication and are evenly distributed. In the bottom middle and
tions through which gas can flow from the bottom layer into top bottom right there are 36 grid blocks with nonzero vertical permeabil-
layer. Intuitively, the top layer would be depleted faster and due to ity for vertical communication, except that the locations of those open-
pressure difference between the top and the bottom layers, the gas ings are at different places. Fig. 7 shows the effect of the variability in
will flow from the bottom layer into top layer. configuration on the error (difference between three layer fine scale
Fig. 3 shows the case where two connections are close to each and one layer coarse scale production). It is clear from the graph that
other; whereas in Fig. 4 the distance between these connections is both the extent and distribution of connections are important.
larger. Finite difference simulation results for both cases are shown For quantifying the effect of these two factors (shale extent and
in Fig. 5. We compare our results with coarse scale model when the distribution) to the upscaling procedure, we used the RL concept
two productive layers are combined into a single layer. It is clear that was originally
qffiffiffiffi
ffi proposed by Zapata and Lake (1989). They defined
that the error from combining layers in the presence of connection RL ¼ HL Kz
and proposed that for RL greater than 10, the layers will
Kh
points is much smaller compared to the no vertical communication
reach the vertical equilibrium. L in the RL equation represents the
scenario. More than that, it can be seen that the result from upscaled
model is closer to fine scale model when the distance between two
connection points is larger. These results are in agreement to the
analytical solution that was obtained in Appendix B.

3.5. Heterogeneous reservoir upgridding (partial crossflow)

The analytical solution as well as the results from a simplified

three layer model illustrates the importance of vertical communica-
tion. In effect, vertical communication “homogenizes” the static hetero-
geneity. Because the pressure equilibrates in vertical direction, the flow

Fig. 5. Simulation result for upscaled model with fine scale model in the presence and
Fig. 3. 3 layer model containing shale barrier and two close connections. absence of vertical communication.
 M. Sharifi, M. Kelkar / Journal of Petroleum Science and Engineering 86-87 (2012) 225–236 229

Fig. 6. Schematic diagram of different connection configuration (kz ≠ 0).

characteristic length in horizontal direction. For layered model, it is communication. On the other hand, if the vertical communication is con-
the total length in horizontal direction. H is the thickness of the two centrated in one part of the layer, it does not provide an effective
layers combined, Kz is the harmonic average of the layer vertical communication. To determine the effective vertical communication, we
permeabilities and Kh is the arithmetic average of the layer horizontal superpose a coarse grid on the fine grid and calculate the new fraction.
permeabilities. High value of RL signifies higher amount of vertical For having a more realistic fraction number that represents both the
communication and vice versa. extent and the distribution of connected sands within the layers, the
Note that L is not a grid block dimension. Therefore, for determining procedure was applied for at least four different superimposed coarse
the vertical communication, we need to consider overall dimension of configurations, i.e., different sizes of superimposed grids. In each step,
the layer. If ‘x’ and ‘y’ dimensions of the two layers are different, then the fraction is calculated and finally the minimum of those four
we can consider square root of area as a representative length. We calculated fractions was considered as an effective fraction. Formula for
used the same concept with some modification. For applying this calculating effective fraction for layer ‘k’ is shown in Eq. (4).
formula to each of the two vertically adjacent grid blocks, “RL” was
calculated locally. In the formula, we used harmonic average for ∑Nt
i¼1 mi wi
vertical permeability and arithmetic average for horizontal permeabil- f ef f ;k ¼ ð4Þ
∑Nti¼1 wi
ity of top and bottom grid blocks. The height also is the summation of
height of both layers. Instead of ‘L’, we used the effective length
concept. Effective length is calculated globally as a square root of the In this equation ‘Nt’ is the total number of grid blocks in each layer
effective area. Effective area is the summation of areas of all of the and the weight (wi) assigned to a grid in a coarse model is inversely
grid blocks in that layer multiplied by an effective fraction. In the proportional to the number of fine scale connections in the coarse
limit, if all the grid blocks in vertical direction are communicating, grid.
the fraction is one. If there is no communication, the fraction is zero. The method is illustrated using a simple example. Assume our fine
The procedure for calculating the effective fraction is discussed in the scale model has Nt grids whereas our coarse superposed model is 2 by
next step. The reason for using effective fraction is to ensure that we 2. If within one of coarse grids that consists of Nt/4 fine scale grids, ‘F’
consider the distribution of vertical crossflow areas in the layers as connected grids exist, then wi for all of the fine grids within the coarse
well as their extent. block will be 4F/Nt. In the case of no connected fine grids in the coarse
For calculating the effective fraction, we assume that if the vertical grid, we assign wi = 1 for all of fine grids inside that coarse grid. In
connections between two layers are well distributed, they provide better addition to the weight factor, for accounting for the extent of connec-
tion we assign mi = 0 for non-connected grids and mi = 1 for connected
grids. For example, let us consider two different connection
distributions in Fig. 8a and c. In both cases the number of connected
grid blocks is the same (real fraction is 0.11). Here we superimpose
once two coarse grid blocks (Fig. 4a and c) and once 4 coarse grid
blocks (Fig. 4b and d) on the top of the fine model. From Eqs. (5-a)
to (5-d), for different configurations we get different fraction and
based on our methodology we choose the minimum fraction as the
effective fraction of the system. Let us explain the procedure for
calculating of effective fraction of Fig. 8b. In this case, the fine scale
model was divided to four coarse grids. In the top-left superimposed
Fig. 7. Difference in oil production between coarse scale (different connection configura- grid, four small openings exist out of nine fine grids. So for all the 9
tion) and the fine scale models. grids wi = 4/9. In other three coarse grids due to nonexistence of
230 M. Sharifi, M. Kelkar / Journal of Petroleum Science and Engineering 86-87 (2012) 225–236

Fig. 8. Different effective fractions for different connection configuration.

connection wi = 1. Finally, as it was mentioned mi = 0 for all noncon- In this step we can calculate the RL by using the effective fraction
nected fine grids. The result is shown in Eq. (5-b). from previous step in Eq. (6).
The interesting point is that in the case of uniform distribution of
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
connected grids (Fig. 8c and d), the effective fraction is higher than Layer area ðAk Þ  f k Arithmatic average of KHk ; KH kþ1
RLk;kþ1 ¼
the case of concentrated distribution (Fig. 8a and b) and the fractions Hk þ Hkþ1 Harmonic average of KZ k ; KZ kþ1
are equal to the real fraction.
ð6Þ
(Fig. 8a)
Fig. 9 shows the relation between RL that was obtained from our
4 calculation and the dynamic error for all three layer model configura-
∑i¼1 184
fi ¼ ¼ 0:04 ð5  aÞ tions that were shown in Fig. 7. As it was mentioned before, error is
∑i¼1 18 þ ∑18
18 4
i¼1 1 the difference between cumulative production of coarse and fine
scale models:
(Fig. 8b) t¼ts
Dynamic Error ¼ ∑t¼0 ðQ –C ðt Þ−Q F ðt ÞÞ ð7Þ

where Q_C and Q_F are coarse and fine scale cumulative production,
∑4i¼1 49
fi ¼ ¼ 0:057 ð5  bÞ and ts is final simulation time. We would expect that better vertical
∑9i¼1 49 þ ∑27
i¼1 1 communication would result in a smaller error. That means, the
larger the value of RL (better vertical communication), the smaller
would be the simulation error. We use the effective fraction for all
(Fig. 8c)
the configurations in Fig. 6 using our analytical approach and plot
that value against the simulation error. Fig. 9 demonstrates that the
∑4i¼1 18
2 dynamic error in flow simulation is well presented by the static
fi ¼ ¼ 0:11 ð5  cÞ error based on our analytical approach. In other words we are able
∑36 2
i¼1 18

(Fig. 8d)

4
∑i¼1 19
fi ¼ ¼ 0:11: ð5  dÞ
∑36 1
i¼1 9

By this procedure, smaller fraction is calculated for the layers in

which connections are not evenly distributed. As an actual example,
the comparison between real and effective fractions of the real
reservoir is discussed in the result section (Fig. 13). What this implies
is that, more evenly distributed the communication between the two
layers, closer the pressure will be within the layers. Fig. 9. Effect of RL on production error on 6-different 3-layer model configuration.
 M. Sharifi, M. Kelkar / Journal of Petroleum Science and Engineering 86-87 (2012) 225–236 231

Start Upgridding
Number of initial Layers=Nz

Column by column error term for all the grid blocks

ϕ ϕ
E I i , j ,k ,k +1 = i , j , k − i , j , k + 1
K i, j,k K i, j ,k +1

Calculating the effective fraction fi and RL

for each two adjacent layers using Equations 4 and 6

EF i , j ,k ,k +1 = EI i , j ,k ,k +1 (0.9 exp(−0.5 RLk , k +1 ) + 0.1)

nx ny
EL k ,k +1 = ∑∑ EF i , j ,k ,k +1
i=1 j=1

Find Minimum
Layer Error
( EL )
m −1 , m

Combine that layer and update error

= EL + EL
U
EL m −1 , m m −1 , m m , m +1

= EL + E L m +1,m + 2
U
EL m , m +1 m , m +1

Nz=Nz-1

Yes
If Nz> Desired Coarse Layer Number

No

End

Fig. 10. Upgridding algorithm flowchart.

232 M. Sharifi, M. Kelkar / Journal of Petroleum Science and Engineering 86-87 (2012) 225–236

to correctly predict what type of dynamic error will result from (5) After we have calculated Nz-1 errors corresponding Nz vertical
upgridding fine scale layers. layers, we arrange them in ascending order. We select the first
Finally we need to define the final error term (E_F) based on both two layers combined as the one with the smallest error. The
preserving the pressure profile criteria (Eq. (3)) and crossflow effect only errors updated are the ones which are below and above
(Eq. (6)). By running different synthetic models we found that the the combined layers. Assume we combine layers m and
following formula can be considered as the representative of the m + 1. After combining m and m + 1 layers, only the error
final error term. terms adjusted are:
    u
EF i; j;k;kþ1 ¼ EIi; j;k;kþ1 0:9Exp −0:5RLk;kþ1 þ 0:1 ð8Þ EL m−1;m ¼ ELm−1;m þ ELm;mþ1 ð11Þ

u
In this way, when RL = 0 final error is equal to the initial static EL m;mþ1 ¼ ELm;mþ1 þ ELmþ1;mþ2 ð12Þ
error (Eq. (3)). The error gets smaller as RL increases; however,
never reaches below 10% of the static error. and the remaining error term will not change. This is done to
ensure that prior errors are accounted for while making the
3.6. Upgridding algorithm decision about which next two layers to combine.
(6) We continue the process regressively so that in each step on
In this section we summarize the workflow for the proposed more vertical layer is eliminated by combining with another
upgridding procedure. layer. This continues till we reduce our fine scale model
enough for running reservoir simulation.
(1) Basic error term is calculated between any two adjacent grid
blocks in vertical direction. This error represents the maximum The summary of upgridding procedure based on both initial error
error possible if no vertical communication exists. and crossflow effect is shown in Fig. 10.




φ


i; j;k φi; j;kþ1
4. Results
EIi; j;k;kþ1 ¼
−
ð9Þ

K i; j;k K i;j;kþ1

4.1. Synthetic examples
(2) Degree of vertical communication between all adjacent layers
Fig. 11 shows the results from different methods for upscaling
is calculated using the procedure that was discussed in
from 50 layers model to a 10 layer coarse model. Here we assume
Section 3.4. So for layer k and k + 1, RLk,k + 1 is calculated for
that Kz = 0. We compare the fine scale cumulative production (50
every pair of vertically adjacent layers.
layers) with 10 layer models generated using various upgridding
(3) We update the error term (EFi,j,k,k + 1) using Eq. (8) for every pair
methods. These methods include proportional layering, flux based
of vertically adjacent grid blocks.
upgridding and the proposed, new, method.
(4) Since our upgridding scheme is based on combining layers
It can be seen that the methodology based on Eq. (3) does better
(and not combining vertical grid blocks), we calculate the
job in terms of having a production profile closer to the fine scale
total error between layers k and k + 1 by using the following
model. For this case the result from both proportional and flux
equation. It is a summation of all the adjacent vertical block
based methods is almost the same and optimistic.
errors over grid blocks in x and y directions.
Eq. (3) is based on no crossflow assumption. In Fig. 12 it can be
nx ny seen that in the case of a perfect crossflow between the layers
ELi; j;k;kþ1 ¼ ∑i¼1 ∑j¼1 EF i; j;k;kþ1 ð10Þ
(Kz = Kh), all the methods will give almost the same result and it is

Fig. 11. Resulting of upgridding from 50 to 10 layers.

 M. Sharifi, M. Kelkar / Journal of Petroleum Science and Engineering 86-87 (2012) 225–236 233

Fig. 12. Oil production comparison in the presence and absence of crossflow.

very close to the fine scale result. By comparing Fig. 12a and b it is grids according to their extent and distribution. This indicates an
clear that crossflow will help the reservoir appear as a homogeneous importance of distribution of sand connectivity across the layers. The
reservoir and the errors from all methods are very small. Using the result from different upgridding techniques can be seen in Fig. 14.
formula for calculating the effective fraction for perfect crossflow It can be seen that the result from the new method is better than
Eq. (4) always gives the same answer and the effective fraction is the result from other methods and it could capture the dynamic
equal to one. Because of that, the value of RL will be maximum response of fine scale model better than the other methods.
(Eq. (6)) and the calculated final error will be minimum (Eq. (8)).
In a real reservoir we have regions where crossflow exists, and 4.3. Real case example_2
regions where the crossflow is negligible. We test proposed upgridding
methodology on the two real cases in the next section. Second real case is a 33 × 33 × 575 model from one heterogeneous
shaly reservoir (reservoir_A). The initial reservoir pressure was
4.2. Real case example_1 4000 psi and the reservoir fluid was dead oil. All the wells are produc-
ing at 1200 psi bottom hole pressure control.
In this section new methodology is applied for upgridding Vible Due to high heterogeneous nature of reservoir the number of
reservoir. Vible is located in Green River Basin, Wyoming, USA. Because layers was reduced from 575 layers to 68 layers. The result of
of the depositional character of the field, the average rock permeabil- different upgridding scheme is shown in Fig. 15. It can be seen here
ities are around 10− 3 md. The model has fifteen producing wells. The that again new method performance in prediction of cumulative oil
reservoir contains single gas phase and includes 53 × 69 × 228 grid production is better than the proportional and flux based method. It
blocks in x,y and vertical directions respectively. In this case the is worth mentioning that proposed method is analytical and hence
reservoir was upgridded from 228 to 28 layers. does not require any flow simulation. This method is very efficient
Before showing the production profile result, the result of and takes just about 3 min for both real cases using a standard per-
calculating effective fraction and the comparison with real fraction is sonal computer.
shown in Fig. 13. It shows that the fraction number obtained from
our method is always less than or equal to real fraction of connected

Fig. 13. Comparison between real and effective fraction in Vible case. Fig. 14. Gas production of Vible reservoir.
234 M. Sharifi, M. Kelkar / Journal of Petroleum Science and Engineering 86-87 (2012) 225–236

p dpave −J
∫pinit   ¼ ∫t0 dt: ð6  aÞ
pave −pwf C t vp

By solving Eq. (6-a), we can express pressure changes in the each

layer. i.e., in the first layer
" #
  −J 1
p1 −pwf ¼ pinit −pwf exp t ð7  aÞ
ct V p1

" #
−J 1
p1 −pwf ∝ exp t : ð8  aÞ
ct V p1

Now we define the simulation error as the difference between the

average pressure in fine layers that are to be combined, and the
coarse grid combined layer average pressure between time zero and
end of simulation time ts.





Fig. 15. Oil cumulative production of reservoir_A.
ts

ts

error1; 2 ¼
∫0 ðp1 −p12 Þdt
þ
∫0 ðp2 −p12 Þdt
ð9  aÞ

5. Summary and conclusions
ts

ts

error2; 3 ¼
∫0 ðp2 −p23 Þdt
þ
∫0 ðp3 −p23 Þdt
ð10  aÞ
In this study relatively simple but computationally efficient (no
need for simulation) method was proposed for upgridding. We used Eq. (9-a) can be expanded in the following line:
analytical solution to determine the best layering solution so that

   

pressures in fine scale will be closer to the coarse scale model during
ts

error1; 2 ¼
∫0 p1 −pwf − p12 −pwf dt
ð11  aÞ
the depletion process. Further, we also considered the impact of
discontinuous sand shale distribution on vertical communication
   


ts

between the layers. We observed that the vertical communication þ
∫0 p2 −pwf − p12 −pwf dt
:
between two adjacent layers is a function of the total number of
connected grid blocks as well as how they are distributed. We devel- The reason that we were able to put absolute value outside of
oped a method to calculate effective communication between two integral in Eqs. (9) and (10) is that the sign of each term does not
adjacent layers. We tested our methodology by comparing the new change with time. The reason that sign would not change is because
method with proportional as well as flux based methods. For both the average pressure in fine scale layers would either be smaller or
the synthetic as well as field cases, we observed that our method is larger than average pressure in the combined layer throughout the
superior to other methods. depletion process. For example we want to prove that {exp[−At]-
exp[− Bt]} always is negative or positive where A = J1/(ctVp1) and
Appendix A B = J12/(ctVp12). Regarding the fact that A and B are constants in
time, above statement is always true. If A > B then the expression is
For simple three layer model with different permeability and always negative and is B > A then the expression is always positive.
porosity and similar height, we can develop a simple solution for To ensure that we can obtain a solution which is independent of
pressure difference between fine and coarse scale models in the follow- time, we change the upper limit of integral to infinity. This would
ing steps: allow us to get a solution independent of time (that can be easily
Flow rate and productivity index of a pseudo steady state system implemented in our upgridding scheme) Substituting,
can be expressed as Eqs. (1-a) to (3-a). (Dake, 1978)

( " # " #)∞


cV −J 1 ct V p12 −J 12

 
t p1

kh error1; 2 ¼
exp dt − exp dt

q¼  re  3 pave −pwf ð1  aÞ
−J 1 ct V p1 −J 12 ct V p12

rw − 4 þ s
ln 0

" #
( " # " #)∞

cV −J 2 ct V p12 −J 12

kh
t p2

J¼  re  3 ð2  aÞ þ
exp dt − exp dt


−J 2 ct V p2 −J 12 ct V p12

ln rw −4 þ s 0







 
ct V p1 ct V p12

ct V p2 ct V p12

φ1 φ1 þ φ2

q ¼ J pave −pwf ð3  aÞ ¼

−
þ



J −
∝
−


K


J 1 J 12 2 J12 1K þK

1 2

We also know that we can express wellbore pressure as a function
φ φ þ φ2



φ1 K 2 −φ2 K 1



φ2 K 1 −φ1 K 2


þ

2 − 1
¼

þ
:ð12  aÞ
of productivity index by following Eqs. (4-a) to (7-a): K2 K1 þ K2 ðK 1 þ K 2 ÞK 1 ðK 1 þ K 2 ÞK 2

1 ∂vp Two final terms
always
have opposite sign, so we can simplify
ct ¼ − ð4  aÞ
φ

vp ∂p them more, i.e.:
φK kk − K kþ1
kþ1

.
Using Eqs. 3-a and 4-a pressure changes in the system as a func- Generally speaking, the final result can be expressed in a very simple
tion of time: form:



dpave dvp  
φ φ

−ct vp ¼ ¼ q ¼ J pave −pwf ð5  aÞ errork;kþ1 ¼

k − kþ1

ð13  aÞ
dt dt K k K kþ1
 M. Sharifi, M. Kelkar / Journal of Petroleum Science and Engineering 86-87 (2012) 225–236 235

Appendix B These flow rates can be calculated as:

Here we determine the amount of contribution from the other n o

side of a barrier in total production of well_A in Fig. 2. For simplicity qc Δpcu;I ðd2 Þ−Uqv1 Δpcu;II ðdL Þ þ Δpcu;I ðdL Þ
qv2 ¼ n o ð6  bÞ
and without loss of generality, we assume that reservoir properties U Δpcu;II ðr w Þ þ Δpcu;I ðr w Þ
on both part of reservoir are the same.
In the absence of any connection point across the barrier it is very
easy to solve this system analytically by using image method (Dake, Δpcu;II ðdL ÞþΔpcu;I ðdL Þ
1978). However, having a connection through the barrier, makes the qc Δpcu;I ðd1 Þ−Δpcu;I ðd2 Þ
Δpcu;II ðrwÞþΔpcu;I ðrwÞ
problem complicated. For this purpose we simulate each connection qv1 ¼ ð7  bÞ
fΔpcu;II ðdL ÞþΔpcu;I ðdL Þg
2

point by an imaginary well. Then instead of solving a very complicated 1−

fΔpcu;II ðrw ÞþΔpcu;I ðrw Þg
2

geometry we consider our system as two regions (Zhao and Thompson,

2002). Region_I consists of well_A, its image and two injection well_1
and well_2. Region_II consists of two imaginary producing wells, Where:
well_1 and well_2. The amount of production from well_1 and 2 is
sffiffiffiffi!!
equal to their injection rate. 5:615 U
For better understanding let us explain the pressure drop at Δpcu;I ðrw; U Þ ¼ K rw ð8  bÞ
hφct 4πηU 0 η
well_A that is located in Region_I (Eq. (1-b)). Pressure change at I

well_A is affected by producing from well_A itself, its image well

and points 1 and 2. The reason is that we know that in absence of sffiffiffiffi!!
any connection (perfect barrier) pressure drop at well_A is summa- 5:615 U
Δpcu;II ðrw; U Þ ¼ K rw ð9  bÞ
tion of pressure drop caused by well_A that is working in infinite res- hφct 4πηU 0 η
II
ervoir, plus the pressure drop caused by its image well with the same
rate that is located at distance 2d1 from well_A (Dake, 1978). But we
know that because of existence of two connections (imaginary wells),
0:006328  k
the real pressure drop should be less than summation of well_A and η¼ : ð10  bÞ
μct φ
its image. The effect of these two imaginary wells is shown in third
and fourth terms in Eq. (1-b). Third and fourth terms are pressure
support from connection numbers one and two respectively. It is K0 is zero order modified Bessel function and U is Laplace variable.
worth mentioning that neither pressure nor flow rate is not constant Finally by using Stephest algorithm, we can solve Laplace Eqs. (6-
at the imaginary well_1 and well_2. Regarding the fact that both pres- b)-(10-b) in time domain for obtaining flow rate. The final result is
sure and flow rate are changing at the wells_1 and 2, we write these shown in Fig. 2c for different distances between points_1 and 2.
equations by applying the Duhamel principle (Thompson and
Reynolds, 1986). Here qc is flow rate at well_A that is constant. qv1 Nomenclature
and qv2 are variable rates at points 1 and 2. Δpcu, I and Δp ' cu, I are pres-
sure change and derivative of pressure change with time, assuming
that wells are in Region_I. Similar to this, we can write Symbol Definition
Eqs. (2)–(5-b). Eq. (2-b) which show the pressure drop at point_1 as-
Ct Total compressibility
suming it is an injection well in Region_I. Eq. (3-b) shows the pres- D0 Distance between well and barrier
sure drop at point_1 considering it as a production well in a second D1 Distance between well and point 1
region. Eq. (4-b) shows the pressure drop at point_2 assuming it is D2 Distance between well and point 2
an injection well in Region_I. Eq. (5-b) shows the pressure drop at DL Distance between point 1 and point 2
EF Grid block Final error
point_2 considering it as a production well in the second region.
EI Grid block Static error
EL Layer error
t ′
ΔpA ¼ qc Δpcu;I ðr w Þþqc Δpcu;I ð2d0 Þ−2∫0 qv1 Δp cu;I ðd1 ; t−τÞdτ−2 ð1  bÞ feff Effective fraction
H Height
t ′
×∫0 qv2 Δp cu;I ðd2 ; t−τÞdτ k permeability
kh Horizontal permeability
kz Vertical permeability
t ′ t ′
Δpo1 ¼ −2∫0 qv1 Δp cu;I ðr w ; t−τ Þdτ−2∫0 qv2 Δp cu;I ðdL ; t−τ Þdτ ð2  bÞ L Length
m Connection indicator
þ2qc Δpcu;I ðd1 ; t Þ Nt Total number of grids in each layer
Nx Number of grid in x direction
Ny Number of grid in y direction
t ′ t ′
Δpo1 ¼ 2∫0 qv1 Δp cu;II ðr w ; t−τ Þdτ þ 2∫0 qv2 Δp cu;II ðdL ; t−τÞdτ ð3  bÞ Nz Number of layer in z direction
Pwf Bottom hole pressure
Pinit Reservoir initial pressure
t ′ t ′
Δpo2 ¼ −2∫0 qv2 Δp cu;I ðr w ; t−τ Þdτ−2∫0 qv1 Δp cu;I ðdL ; t−τ Þdτ ð4  bÞ q flow rate
Q_C Coarse scale cumulative production
þ2qc Δpcu;I ðd2 ; t Þ Q_F Fine scale cumulative production
qc Constant flow rate
t ′ t ′ qv Variable flow rate
Δpo2 ¼ 2∫0 qv2 Δp cu;II ðr w ; t−τ Þdτ þ 2∫0 qv1 Δp cu;II ðdL ; t−τÞdτ ð5  bÞ rw Wellbore radius
ts Simulation time
U Laplace variable
We can simplify these five equations for calculating the five un- Vp Pore volume
knowns. After solving Eqs. (1-b)–(5-b) for the flow rate and applying w Weighting factor
the Laplace transform, we get the expression for flow through J Productivity index
points_1 and 2 in Fig. 2 (Eqs. (6-b) and (7-b)). Here we are interested η Diffusivity coefficient
φ Porosity
to see the amount of flow rate that is passing through these points.
236 M. Sharifi, M. Kelkar / Journal of Petroleum Science and Engineering 86-87 (2012) 225–236

References King, M.J., Burn, K.S., Wang, P., Muralidharan, V., Alvardo, F., Ma, X., Datta-Gupta, A.,
2006. Optimal coarsening of 3D reservoir models for flow simulation. SPE Reserv.
Ates, H., Bahar, A., El-Abd, S., Charfeddine, M., Kelkar, M., Datta-Gupta, A., 2005. Eval. Eng. 317–334 (October 20–22).
Ranking and upscaling of geostatistical reservoir models using streamline simula- Li, D., Beckner, B., 2000. Optimal uplayering for scaleup of multimillion-cell geologic
tion: a field case study. SPE Reserv. Eval. Eng. 8 (1), 22–32. models. SPE Annual Technical Conference and Exhibition (October 1–4).
Dake, L.P., 1978. Fundamental of reservoir engineering. Elsevier, Amsterdam. Li, D., Beckner, B., 2001. A new efficient averaging technique for scaleup of
Durlofsky, L.J., 1996. Scale up of heterogeneous three dimensional reservoir descriptions. multimillion-cell geologic models. SPE Reserv. Eval. Eng. 297–307.
SPE J. 313–326. Li, D., Cullick, A.S., Lake, L.W., 1995. Global scale-up of reservoir model permeability
Durlofsky, L.J., 2005. Upscaling and gridding of fine scale geological models for flow with local grid refinement. J. Pet. Sci. Technol. 149–157.
simulation. Presented at the 8th International Forum on Reservoir Simulation Stern, D., 2005. Practical aspects of scaleup of simulation models. J. Pet. Technol. 57 (9),
(Stresa,Italy). 74–81.
Durlofsky, L.J., Behrens, R.A., Jones, R.C., Bernath, A., 1994. Application of a new scale up Stern, D., Dawson, A.G., 1999. A technique for generating reservoir simulation grids to
methodology to the simulation of displacement processes in heterogeneous preserve geologic heterogeneity. SPE paper 51942, presented at SPE Reservoir
reservoirs. SPE paper 28704, presented at the SPE International Petroleum Confer- Simulation Symposium held in Houston, Texas (Feb 14–17).
ence and Exhibition of Mexico, Veracruz (Oct 10–13). Taware, S., Friedel, T., Datta-Gupta, A., 2011. A practical approach for assisted history
Durlofsky, L.J., Jones, R.C., Milliken, W.J., 1997. A nonuniform coarsening approach for matching using grid coarsening and streamline-based inversion: experiences in a
the scale up of displacement processes in heterogeneous porous media. Adv. giant carbonate reservoir. SPE paper 141606, Presented at The Woodlands, Texas
Water Resour. 335–347. (February).
Fincham, A., Christensen, F.A., 2004. Up-gridding from geological model to simulation Testerman, J.D., 1962. A statistical reservoir-zonation technique. J. Pet. Technol.
model: review, applications and limitations. SPE paper 90921,presented at SPE 889–893.
Annual Technical Conference and Exhibition (September 26–29). Thompson, L.G., Reynolds, A.C., 1986. Analysis of variable-rate well-test pressure data
Hosseini, S.A., Kelkar, M., 2010. Analytical upgridding method to preserve dynamic using Duhamel's principle. SPE formation evaluation, pp. 453–469.
flow behavior. SPE Reserv. Eval. Eng. 473–484 (June). Zapata, V.J., Lake, L.W., 1989. A theoretical analysis of viscous crossflow. SPE Annual
Kim, J.U., Datta-Gupta, A., Jimenez, E.A., Hohl, D., 2010. A dual scale approach to Technical Conference and Exhibition, San Antonio, Texas (October 4–7).
production data integration into high resolution geologic models. J. Pet. Sci. Eng. Zhao, G., Thompson, L.G., 2002. Semi analytical modeling of complex-geometry
147–159. reservoirs. SPE Reserv. Eval. Eng. 437–446.
King, M.J., 2006. Upgridding and upscaling: current trends and future directions. SPE
112810-DL, presented during Distinguished Lecture.