Professional Documents
Culture Documents
00084372
00084372
00084372
兺 sជ
assume that the permeability is spatially correlated with a Gaussian 1
sជkt ≈ ŝជkf = f
k,i , . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (7)
correlation model—for example, that the permeability in gridblock N i=1
(i1,j1) is correlated with the permeability in gridblock (i2,j2) with
correlation coefficient where N is the sample size of the ensemble. With this approxima-
tion of the true state, an approximation of a left factor of the error
exp − 冋 冉 冊 冉 冊册
i1 − i2
l
2
−
j1 − j2
l
2
. . . . . . . . . . . . . . . . . . . . . . . . . . (3)
covariance matrix of the model is
1
关L兴kf = 关共sជk,1
f
− ŝជkf兲, · · · , 共sជk,N
f
− ŝជkf兲兴 . . . . . . . . . . . . . . (8)
Using the correlations specified in Eq. 3, a correlation matrix [C] 公N − 1
is computed. From this correlation matrix, the covariance matrices
used while generating the initial permeability field and the model a matrix with N columns. The approximation of the model error
noise are computed as 2[C], where is the standard deviation in covariance matrix then becomes
the permeability of each gridblock. The standard deviation used
关P兴kf = 关L兴kf 共关L兴kf兲T. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (9)
while generating the initial permeability field (Eq. 1) is usually
much higher than the one used while adding the model noise in Eq. The expression of the Kalman gain matrix is7
2. The correlation length, l, is also treated as a normally distributed
stochastic variable. This means that (the noise added to) each 关K兴k = 关P兴kf关H兴T 共关H兴关P兴kf关H兴T + 关R兴k兲−1. . . . . . . . . . . . . . . . . . . (10)
ensemble member is generated by first drawing l from a normal The analyzed state of each member of the ensemble is computed as
distribution, then generating [C] and finally drawing a permeabil-
ity field from a multinormal distribution with zero mean and co- sជk,i
a
= sជk,i
f
+ 关K兴k共dជ k,i − 关H兴sជk,i
f
兲. . . . . . . . . . . . . . . . . . . . . . . . . . . . (11)
variance matrix 2[C]. To generate the stepwise model noise, this The analyzed error covariance matrix, [P]ak of the model can be
process is repeated, starting by resampling a random correlation computed along the same lines as [P]fk. Because the updating of the
length l. ensemble is linear, the new estimate of the true state, based on the
To generate the noise added to time-dependent variables of the ensemble after the analysis step, is
model, we draw from a normal distribution with zero mean and
covariance matrix [Q]d. In the examples, we use [Q]d⳱[I], where sជkt ≈ ŝជka = ŝជkf + 关K兴k 共dជ k − 关H兴ŝជkf兲 . . . . . . . . . . . . . . . . . . . . . . . . . . (12)
≈2×10−16. In generation of the initial ensemble there is no noise
in the state variables, as all the ensemble members are started using and the model error covariance matrix after the analysis step is
the same equilibrium condition of the reservoir.
关P兴ka = 共关I兴 − 关K兴k关H兴兲关P兴kf. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (13)
The initial model error is due to uncertainty in the permeability
estimate. In addition, a stepwise model error is added. This is done The underlying assumptions behind the filter are that there is
both to account for the fact that we are not able to represent the zero covariance between the model error and the measurement
Fig. 1—Overview of reservoir model showing wells and ini- from the original field model. The porosity is shown in Fig. 2. The
tial saturations. measurements are generated by running the simulator with the
“true” permeability, and adding noise to the values obtained.
error and that both the model error and measurement error are Injection wells are controlled by historical injection rate. Pro-
uncorrelated in time. duction wells are controlled by historical reservoir volume rate.
In evaluating the quality of the solution we use the root mean The natural logarithm of the permeability is used in the computa-
square (RMS) error, defined as tions with the Kalman filter. All the data of the restart files from
冑
the reservoir simulator were stored for each ensemble member, and
M during the Kalman filter update, those values corresponding to the
兺 共X − X 兲 , . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (14)
1 t 2
j j state variables were manipulated according to Eq. 12.
M j=1 The initial ensemble is generated using a mean correlation
where M is the number of state variables, Xj is the estimate of state length of 16 gridblocks with a standard deviation of six gridblocks.
variable j, and Xtj is the true value of state variable j. The initial ln-permeability fields for the Kalman filter are gener-
ated using a standard deviation of 0.5. For the model noise (Eq. 2),
Example: Simplified Field Model. A simplified 2D field model the standard deviation is set to 5⭈10−4. The initial mean of
has been constructed by taking out a horizontal layer from a model the permeability of the ensemble is set to 1000 md for all
of a North Sea field. The dimensions of the simplified model are the gridblocks.
39×55, but with some inactive blocks. There are 1,931 active The measurement uncertainties are given in Table 1. We apply
gridblocks. The simulation spans a time range of 3,955 days. The the filter to take into account the information gained from four
reservoir model is shown in Fig. 1. measurements for each of the production wells. The measured
Measurements are assimilated at least once a month, but also quantities are pressure, oil production rate, water cut, and gas/
when new wells are starting production, or when wells are shut in. oil ratio.
Altogether, there are 171 points in time when measurements are In Fig. 3, the development in the estimated permeability is
assimilated. The measurements are assimilated from 14 producing shown together with the “true” permeability and the reference
wells. Two of them start from the first day. The reservoir is pro- solution. The reference solution is the mean of the initial ensemble,
duced by gas injection, and the producers are located along the and because the initial ensemble is based on a model with constant
oil/water contact. There are four gas injectors. There are no mea- mean permeability of 1000 md, the reference model will have a
surements from the injectors. permeability of approximately 1000 md in all cells. One can ob-
In addition to the 14 producers and 4 injectors, there are some serve that after 1,611 days, one obtains a good match of the per-
dummy wells included in the simulation to adjust for flow through meability field especially close to the producing wells. The loca-
the boundary. The rates of the dummy wells are kept constant tions of the producers are shown with circles. For the permeabili-
while generating the data and running the ensemble Kalman filter. ties along the boundaries, we observe a slight drifting toward the
The quantities that are tuned with the Kalman filter are the end of the simulation. Further development of the filter is needed
horizontal permeability, which is assumed to be the same in the x- to avoid this effect.
and y-direction, together with the dynamic variables, which are the The estimated uncertainty in the ln-permeability is presented in
pressure, saturation of water and gas, the solution gas/oil ratio, and Fig. 4. The uncertainty is lowest close to the producers, where the
the vapor oil/gas ratio. This means that we have both vaporized oil measurements are obtained. The highest uncertainty is close to the
and dissolved gas. boundary of the reservoir. In general, the uncertainty decreases as
The “true” model is obtained from a stochastic realization of more measurements are assimilated. Close to the boundaries, the
the field. Relative permeabilities and PVT properties were taken model error added in Eq. 2 dominates the reduction in uncertainty
obtained from measurements at later times.
In Figs. 5 through 8, we present the outcome of the filter for
some of the measured variables, together with the reference solu-
tion. We have selected to present results from measurements that
have among them the largest relative deviations between the truth
and the solution obtained from the Kalman filter. Notice that the
filter is able to follow the measurements closely.
In Figs. 9 through 12, we show forecasts based on the en-
semble mean after 4 days, 819 days, and 3,050 days for the same
measurements as presented in Fig. 5. As expected, the quality of
the forecasts is improved as more data are assimilated. It is clearly
seen that the forecasts improve when more data are included, and recover the main trends of the permeability field rapidly because of
that the forecasts based on data up to 3,050 days are almost perfect. the continuous downhole pressure measurements. However, al-
In Fig. 13, we show how the RMS error defined in Eq. 14 though new measurements are added continuously, the error in the
evolves. The RMS error for the permeability decreases very rap- estimated permeability increases slightly late in time. The reason
idly and then stabilizes. Still, the predictions clearly improve with for this instability is not fully understood.
time as additional measurements become available. Note also that The initial states of the dynamic variables are determined from
with the exception for RS and RV, the RMS values for the state an equilibrium state. This fixes the initial state of the reservoir.
variables stabilize relatively fast at quite low levels. Therefore, the difference between the estimate obtained from the
filter and the true value of the dynamic variables has to increase in
Discussion the beginning. After a while, the differences stabilize.
The ensemble Kalman filter technology seems promising for tun- It is crucial for a successful application of the methodology that
ing permeability and dynamic variables, such as the pressure and the ensemble of models correctly reflects the uncertainty at all
saturations, in reservoir simulator models. The filter was able to times. Therefore, the generation of the initial ensemble and how to
Fig. 4—Uncertainty in estimated ln-permeabilities. The values are the standard deviation in ln-permeability for each block. Note that
the plots in the bottom row use the same color scale as those in the upper right plot.
add model noise are important questions that also have to be fur- Conclusions
ther addressed.
Technically, there is no limit with respect to how many model An ensemble Kalman filter technology is developed for continuous
parameters can be updated, but the effect of adding a large number updating of reservoir simulation models. The example is shown in
of parameters on filter stability, as well as other concerns, has not 2D, but the methodology can easily be extended to 3D and applied
yet been considered. to any existing reservoir simulator.
Fig. 6—Measured values, filter solution, true solution, and the reference solution for the produced oil rate in some wells. The legend
is the same as in Fig. 5. The true solution is mostly hidden behind the filter solution. Lower limit for the x-axis corresponds to well
startup time. Upper limit corresponds to end of history.
The filter is used to tune the permeability and dynamic vari- filter is able to track the measurements and tune the permeability
ables such as pressure and saturations. The continuous updating field, and as more measurements are assimilated the forecasts
ensures that the predictions always start from a solution that are improved.
matches the observed production data. Although the presented results are promising, this technology is
The methodology is applied to a “semisynthetic” field model novel to reservoir simulation, and further development of the filter
with 14 producers and four injectors. The ensemble Kalman is needed for this application.
Fig. 8—Measured values, filter solution, true solution, and the reference solution for the gas/oil ratio in some wells. The legend is
the same as in Fig. 5. The true solution is mostly hidden behind the filter solution. Lower limit for the x-axis corresponds to well
startup time. Upper limit corresponds to end of history.
Fig. 10—Forecasted values for some of the oil production rates. The forecasts are based on the mean of the initial ensemble (0
days), the forecast after 4 days, after 819 days, after 3,050 days, and the true solution (T). Legends are as in Fig. 9. Lower limit for
the x-axis corresponds to well startup time. Upper limit corresponds to end of history.
Subscripts Acknowledgments
d ⳱ dynamic variables This work has been supported financially by ENI-Agip SpA,
i ⳱ index (ensemble members) Norsk Hydro ASA, Statoil, and the Research Council of Norway.
j ⳱ index (state variables) References
k ⳱ index (assimilating time) 1. Evensen, G.: “Sequential data assimilation with a nonlinear quasi-
geostrophic model using Monte Carlo methods to forecast error statis-
Superscripts tics,” J. Geophys. Res. (1994) 99, 143.
2. Haugen, V.E. and Evensen, G.: “Assimilation of SLA and SST data
a ⳱ analyzed (a posteriori) into OGCM for the Indian Ocean,” Ocean Dynamics (2002) 52, 133.
f ⳱ forecasted (a priori) 3. Margulis, S.A. et al.: “Land data assimilation and estimation of soil
m ⳱ model noise moisture using measurements from the Southern Great Plains 1997
t ⳱ true Field Experiment,” Water Resources Research (2002) 38, No. 12,
T ⳱ matrix transpose 1299, DOI: 10.1029/2001WR001114.
Fig. 12—Forecasted values for some of the gas/oil ratios. The forecasts are based on the mean of the initial ensemble (0 days), the
forecast after 4 days, after 819 days, after 3,050 days, and the true solution (T). Legends are as in Fig. 9. Lower limit for the x-axis
corresponds to well startup time. Upper limit corresponds to end of history.
4. Lorentzen, R.J. et al.: “Underbalanced and Low-Head Drilling Opera- 11. Casella, G. and Berger, R.L.: Statistical Inference, Duxbury Press,
tions: Real-Time Interpretation of Measured Data and Operational Sup- Belmont, California (1990).
port,” paper SPE 71384 presented at the 2001 SPE Annual Technical 12. Burgers, G., van Leeuwen, P.J., and Evensen, G.: “On the analysis
Conference and Exhibition, New Orleans, 30 September–3 October scheme in the ensemble Kalman filter,” Monthly Weather Review
2001. (1998) 126, 1719.
5. Nævdal, G., Mannseth, T., and Vefring, E.H.: “Near-Well Reservoir
Monitoring Through Ensemble Kalman Filter,” paper SPE 75235 pre-
sented at the 2002 SPE/DOE Improved Oil Recovery Symposium,
Tulsa, 13–17 April. Geir Nævdal is a chief scientist at RF-Rogaland Research in
Bergen, Norway. e-mail: geir.naevdal@rf.no. He holds a PhD
6. Nævdal, G., Mannseth, T., and Vefring, E.H.: “Instrumented wells and
degree in mathematics from the U. of Trondheim. Liv Merethe
near-well reservoir monitoring through ensemble Kalman filter,” Proc.,
Johnsen is a reservoir engineer at Norsk Hydro, and has 4 years
2002 European Conference on the Mathematics of Oil Recovery, of experience there as a reservoir engineer and research sci-
Freiberg, Germany, 3–6 September. entist. Before joining Norsk Hydro, she worked for 4 years at the
7. Maybeck, P.S.: Stochastic Models, Estimation, and Control, Vol. 1., Nansen Environmental and Remote Sensing Centre as a re-
Academic Press, New York City (1979). search scientist. She holds an MSc degree in applied math-
8. Wan, E.A. and Nelson, A.T.: “Dual extended Kalman filter methods,” ematics from the U. of Bergen. Sigurd Ivar Aanonsen is a prin-
Kalman Filtering and Neural Networks, Ch. 5, John Wiley & Sons Inc., cipal researcher and professor at the Centre for Integrated
New York City (2001) 123–173. Petroleum Research, U. of Bergen. He has 18 years of experi-
ence with Norsk Hydro as a reservoir engineer and research
9. Lorentzen, R.J., Nævdal, G., and Lage, A.C.V.M.: “Tuning of param-
scientist. Aanonsen holds a PhD degree in applied mathemat-
eters in a two-phase flow model using an ensemble Kalman filter,” ics from the U. of Bergen. Erlend H. Vefring is a research director
International Journal of Multiphase Flow (2003) 29, 1283. at RF-Rogaland Research and a professor at the U. of Bergen.
10. Houtekamer, P.L. and Mitchell, H.L.: “Data assimilation using an ensem- He holds a PhD degree in applied mathematics from the U.
ble Kalman filter technique,” Monthly Weather Review (1998) 126, 796. of Bergen.