Page 1
© I
FP
Controlled CO2 | Diversified fuels | Fuel-efficient vehicles | Clean refining | Extended reserves
Reservoir Modeling with EnKF: A sensitivity study
L. Heidari (IFP), V. Gervais (IFP), M. Le Ravalec (IFP), H. Wackernagel (ENSMP)Ecole Des Mines de Paris (ENSMP)Institut Français du Pétrole (IFP)
4th Ensemble Kalman Filter Workshop, 22-24 June 2009 - Bergen , Norway
Page 2
© I
FP
2
Overview Background
Reservoir Study and History Matching Kalman Filters Ensemble Kalman Filter – EnKF
Application 2D synthetic case study Sensitivity Tests
Conclusions & perspectives
Page 3
© I
FP
3
Overview Background
Reservoir Study and History Matching Kalman Filters Ensemble Kalman Filter – EnKF
Application 2D synthetic case study Sensitivity Tests
Conclusions & perspectives
Page 4
© I
FP
4
Reservoir Study
?=U
ps c
a lin
g
(Ph
i , K
)
Geologicalmodel
(facies, porosity,permeability)
Flu
id F
low
S
imu
lati
on
Reservoirmodel
100
150
200
250
0 500 1000 1500 2000 2500 3000
Time (day)
BH
FP (b
ar)
0
100
200
300
400
0 500 1000 1500 2000 2500 3000
Time (day)
GO
R
0.0
0.1
0.2
0.3
0.4
0 500 1000 1500 2000 2500 3000
Time (day)
WC
UT
Observed Production
data
100
150
200
250
0 500 1000 1500 2000 2500 3000
Time (day)
BH
FP (b
ar)
0.0
0.1
0.2
0.3
0.4
0 500 1000 1500 2000 2500 3000
Time (day)
WC
UT
0
100
200
300
400
0 500 1000 1500 2000 2500 3000
Time (day)
GO
R
Simulated production data
Aim Constrain reservoir models to production history by
adjusting parameters: porosity, permeability, etc.
Page 5
© I
FP
5
Sequential Methods
Initial state
Timet0 t1
Measurements
Forecasted state (t1)
Timet0 t1
Dynamical model
Corrected state (t1)
Assimilation method
Forecasted state (t2)
Dynamical model t2
t2
Measurements
Correctedstate (t2)
Assimilation method
…
dfa21 AA
Kalman Filter
Page 6
© I
FP
6
Overview Background
Reservoir Study and History Matching Kalman Filters Ensemble Kalman Filter – EnKF
Application 2D synthetic case study Sensitivity Tests
Conclusions & perspectives
Page 7
© I
FP
Assumptions
Linear dynamic model
Gaussian probability distribution function
First Order Markov processes
Independent and uncorrelated model and measurement errors
Kalman Filter estimate is based on
Forecast Estimate
Measurements
7
Kalman filter
dfa21 AA
)t()t( 1 ka
kf F
td H
Dynamic Model Operator
Measurement Operator
True State
Measurement error
Page 8
© I
FP
Kalman filter estimate, a , minimizes
Kalman filter estimate, a :
8
Kalman filter analysis
fTfTffa HdCHHCHC 1
Kalman gain Simulated data
)()()()()(J 11 HdCHdC TffTf
( )( ) : Forecast error covariance matrixf t f t f TC
Distance between and f
matrix covarianceerror tsMeasuremen : TεεC
Distance between and d
dfa21 AA
Page 9
© I
FP
System Dynamics Linear
Non linear
Slightly Non Linear Models Extended Kalman Filter Taylor expansion of nonlinear model operator F
Highly Non Linear Models Ensemble Kalman Filter Ensemble representation of error covariance matrix
9
Kalman filter forecast
)()( ),()( 11T
ka
kf
ka
kf FtFCtCtFt
Model operator ?)( ,)()( 1 k
fk
ak
f tCtFt
Page 10
© I
FP
10
Overview Background
Reservoir Study and History Matching Kalman Filters Ensemble Kalman Filter – EnKF
Application 2D synthetic case study Sensitivity Tests
Conclusions & perspectives
Page 11
© I
FP
Ensemble representation of error covariances for Model State Measurements
Forecast The time evolution of uncertainty is obtained by integrating
each model state forward in time
Analysis Kalman linear analysis scheme applied to each model state
with the sample covariance
11
Ensemble Kalman filters (EnKF)
TeCC
Best-guess
Tffff
e
ff CC
fTfTfaj
aj HdCHHCHC
1
Page 12
© I
FP
12
Advantages of EnKF for History Matching
Only one fluid-flow simulation over the whole production history per model of the ensemble
Real-time assimilation of data, for any data frequency
Uncertainty quantification
Practical implementation
Can be plugged with any fluid-flow simulator Suitable for parallel computation Relatively low computational cost
Page 13
© I
FP
13
Overview Background
Reservoir Study and History Matching Kalman Filters Ensemble Kalman Filter – EnKF
Application 2D synthetic case study Sensitivity Tests
Conclusions & perspectives
Page 14
© I
FP
14
Application : 2D synthetic case
P P
P P
I
Reservoir Size : 500 × 500 × 1 m Grid : 50× 50 × 1
Petrophysical Properties porosity = cte , Kv/Kh = cte Reference ln(Kh): spherical variogram
Fluid Flow oil-water, IFP simulator: PumaFlow Water Injection
Well plan 5 wells: 4 producers and 1 injector
Production data Bottom hole Pressure
(BHP) Surface Oil Rate
(QOS) Water cut
(WCT)
Page 15
© I
FP
15
Reference data
P1
P2 P3
P4
Inj
Ref ln(Kh)
Page 16
© I
FP
16
Assimilation with EnKF
Ensemble of 50 models State-parameter vector
ln( )
k
k
Kh
P
WSat
QOS
BHP
WCUT
Pressure
Water saturation
Surface oil rate
Bottom hole pressure
Water cut
Log of permeability Static parameter
Dynamical state
Simulated measurements
One value pergrid block
One value per well
17 assimilations, Once every month
Page 17
© I
FP
17
Evolution of ln(Kh) mean and variance
1st assimilation 4th assimilation 10th assimilation 17th assimilationInitial Model
Me
an
Va
rian
ce
ln(Kh) reference model
Page 18
© I
FP
18
Simulated answer and prediction – Prod3
Initialensemble
Finalensemble
Ensemblemembers
Ensemblemean Reference
Prediction
Page 19
© I
FP
19
Root Mean Square Error - RMS
Total number of grids
Ensemble Mean value for each
grid
Reference value at each grid
Number of time steps
Number of data per time step
Reference model Measurements
Simulated data for ensemble member
Page 20
© I
FP
20
Overview Background
Reservoir Study and History Matching Kalman Filters Ensemble Kalman Filter – EnKF
Application 2D synthetic case study Sensitivity Tests
Conclusions & perspectives
Page 21
© I
FP
21
Effect of Initial Ensemble 10 Ensembles of size 50 Prediction of cumulative oil production after 5.5
years
True Oil Cumulative Oil Production
Page 22
© I
FP
22
Effect of Assimilation Time Interval
Long Assimilation Interval
Base Case
Short Assimilation Interval
2 Months
Time
Page 23
© I
FP
23
Effect of Ensemble Size - 1
Reference ln(Kh)
Init
ial
Fin
alF
inal
Init
ial
Mea
nV
aria
nce
Ensemble Size
50 100 200 500
Page 24
© I
FP
24
Effect of Ensemble Size - 2Ensemblemembers
Ensemblemean Reference
50 Members
100 Members
200 Members
500 Members
BHP QOS WCT
Prediction
Page 25
© I
FP
25
Measurement Uncertainty
High Case : 10%Base Case : 5%Low case : 1%Initial Model
Me
an
Va
rian
ce
Page 26
© I
FP
26
Overview Background
Reservoir Study and History Matching Kalman Filters
Ensemble Kalman Filter Theory Implementation
Application 2D synthetic case study Sensitivity Tests
Conclusions & perspectives
Page 27
© I
FP
27
Main challenges Finite number of models used to represent uncertainty
Quality of initial sampling Variability Preservation
Linear analysis for a non linear model Physical inconsistency between parameters and state after
analysis
Non Gaussian distributions Multimodal distributions are not well preserved (e.g.
channel and barrier permeability distribution, facies distributions)
Large amount of data 4D seismic
Page 28
© I
FP
28
Conclusions and perspectives EnKF application to a 2D model and Sensitivity studies
Several ensembles of the same size for a sound uncertainty quantification
Smaller assimilation time interval : Better Match Lower Variance
Larger ensemble size : Better match for production data Variance preservation Larger simulation and computational cost
Larger Measurement Uncertainty Variance Preservation Less efficient for matching production data
Perspectives Facies History Matching Real field case study
Page 29
© I
FP
Thank you for your attention !Any Questions ?