27
Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

Embed Size (px)

Citation preview

Page 1: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

Classification: Internal Status: Draft

Using the EnKF for combined state and parameter estimation

Geir Evensen

Page 2: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

2

Outline

• Reservoir modelling and simulation

• History matching problem and uncertainty prediction

• Ensemble Kalman filter (EnKF)

• Field case example

Page 3: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

3

Reservoir Geophysics and Fast Model Updating

• Business challenge

– To reduce uncertainty in reserves and production targets

• Project goal

– Provide continuously updated and integrated models with reduced and quantified uncertainty

• Activities

– Seismic acquisition and imaging

– 4D quantitative analysis

– Integrated use of 4D seismic data

– Well based reservoir monitoring

– Model uncertainty and updating

– Integrated IOR work processes

Page 4: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

4

The geological model

Log(K)

Phie

Geological 3D model

Structural framework(Seismic data)

Depositional model Rock properties distribution

Lithology: facies, porosity and permeability

Depth of fluid contacts and fluid properties

Page 5: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

5

Production data

Time (days)

Oil

flo

w r

ate

(m

3/d

ay)

History matching reservoir models

• Traditional parameter estimation

• Find parameter-set that gives best match to data

– Production and seismic data

• Definition of quadratic cost function

– Perfect model assumption

• Minimization of cost function

– Adjoints, gradients, genetic algorithms, ensemble methods

• Traditional workflow updates only simulation model

Simulation model

Page 6: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

6

History matching and uncertainty prediction

History Prediction

Initial uncertainty Predicted uncertainty

Reduced initial uncertainty

Reduced predicted uncertainty

Page 7: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

7

Assisted history matching

• Parameterization

• Definition of cost function

• Minimization/sampling

• High-dimensional problem

• Highly nonlinear problem

• Model errors ignored

• Multiple local minima

• Hard to solve

Page 8: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

8

General formulation

Find posterior pdf of state and parameters given measurements and model with prior error statistics

Combined parameter and state estimation problem

Bayesian formulation

Page 9: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

9

Bayesian formulation

Bayes’ theorem

Gaussian priors Markov model

Independent data

Quadratic cost-function Sequential processing of measurements

Sequence of inverse problems

p(x|d)~p(x)p(d|x)

Minimization/Sampling

”Ignore model errors”

Solve only for parameters? Ensemble methods

Page 10: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

10

History matching and uncertainty prediction

EnKF procedure

Todays posterior is tomorows prior

p(x|d1) ~ p(x) p(d1|x)

p(x|d1,d2) ~ p(x|d1) p(d2|x)

Page 11: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

11

Ensemble Kalman Filter

• Sequential Monte Carlo method

• Representation of error statistics by an ensemble of model states

– Mean and covariance

• Evolution of error statistics by ensemble integrations

– Stochastic model equation

• Assimilation of measurements using a variance minimizing update

– Sequential updating of both model state and static parameters

– Model state and parameters converge towards true values

– Information accumulates and uncertainty is reduced at each update

Page 12: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

12

EnKF can update geo-realizations

Geo-model

Geo-realizations

Simulation realizationsE

nKF

Log data

RFT/PLT data

Production rates

4D seismics

Page 13: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

13

Oseberg Sør reservoir model

• Dimensions: • Field 3 km x 7 km, 300m thick• Cells size 100 x 100m, z variable• 60 ‘000 active cells

• Complex reservoir • Heterogeneous flow properties• Many faults, poorly known properties • Fluid contacts poorly known

• Parameters to estimate • Porosity and permeability fields• Depth of fluid contacts• Fault properties• Relperm parameterization

• Condition initial ensemble on production data • 4 producers, 1 water injector • 6 years of production history

Permeability field

Page 14: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

14

Initial ensemble uncertainty span

Oil production rate

Water cut

Measurements

Initial ensemble

Page 15: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

15

OPRWCT

Measurements

Initial ensemble

EnKF updated ensemble

Posterior prediction and uncertainty span

Oil production rate

Water cut

Page 16: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

16

Oil Water relative permeability

0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1

0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1

Water Saturation

Krw Initial mean

Krow Initial mean

Krw Updated

Krow Updated

Page 17: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

17

Porosity layer 19 (UT), prior and posterior

Initial EnKF updated

Page 18: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

18

Porosity standard deviation layer 19, prior and posterior

Initial EnKF updated

Page 19: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

19

Improved estimate of initial WOC depth

2907± 5m

2890 ± 2m

Page 20: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

20

Fault transmissibility estimation

Page 21: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

21

• Grane reservoir

– Grid consists of 90x168x20 grid cells

– Homogenous/high permeability

– Unclear vertical communication

– Poorly known initial contacts

• Parameters to estimate

– PORO and PERM

– MULTZ

– WOC & GOC

– RELPERM

• Conditioning on production

– 3 years production history, 19 wells

– OPR, WCT, GOR

Real time prediction of oil production using the EnKF

Page 22: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

22

Page 23: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

23

Page 24: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

24

Page 25: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

25

Conclusions

EnKF can efficiently history match complex reservoir models

General tool for parameter and/or state estimation.

Practically no limitation on parameter space.

Problem with local minima avoided.

Workflow and EnKF method allow for:

Consistency in model chain.

Estimates with quantified uncertainty.

Real time and sequential updating of models.

Updated ensemble provides future prediction with uncertainty estimates

Page 26: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

26

Issues and future challenges

• EnKF with general facies models

– Involves non-Gaussian variables

• Pluri-Gaussian representation

• Kernel methods

• EnKF for estimating structural

parameters like faults and surfaces

– Changes model grid

• Conditioning geo-models

– Consistent links between geo- and

simulation model

• Operational workflow / best practice

– Generally applicable

Page 27: Classification: Internal Status: Draft Using the EnKF for combined state and parameter estimation Geir Evensen

27

Operational ocean prediction system

TOPAZ system:

27 000 000 unknowns

148 000 weekly observations

100 ensemble members

Local analysis