Vous êtes sur la page 1sur 37

Uncertainty,  sensitivity  and  rejection  in    

predictive  reservoir  modeling  


 
Jef  Caers  
Stanford  University,  USA  
Stanford  Center  for    
Reservoir  Forecasting  

‡ Quantitative  modeling  of  geological  heterogeneity  


‡ Modeling  uncertainty  in  the  context  of  decision  
making  
‡ Building  3D/4D  models  accounting  for  scale  and  
accuracy  of  geological,  geophysical  and  reservoir  
engineering  data  
SCRF  overview  
y Reservoir  geology  
y Multiple-­‐point  /  pattern-­‐based  geostatistics  
y Surface-­‐based  geostatistics  
y Structural  modeling  
y Basin  modeling  
 
y Reservoir  geophysics  
y Seismic  reservoir  characterization  
y Rock  physics  
y 4D  seismic  
 
y Reservoir  Engineering  
y Sensitivity  analysis  /  History  matching  
y Upscaling  
y Uncertainty,  decision  analysis  and  value  of  information  
y Proxy  models  /  model  complexity  
 
Uncertainty,  sensitivity  and  rejection  in    
predictive  reservoir  modeling  
 
Jef  Caers  
Stanford  University,  USA  
Current  practice  
͞ŚĂƐŝŶŐĚĂƚĂ͟  
2D  seismic   The  reservoir   The  reservoir  

3D  seismic  

3D  seismic+production  
Life-­‐time  

Life-­‐time  
4D  seismic+production  

Industry  practice  of   ͞ǁŝĚĞƌƉƌŝŽƌƐ͟  


͞ŚŝƐƚŽƌLJŵĂƚĐŚŝŶŐ͟   sensitivity/rejection  
Why?  

History  matched  permeability    


in  a  real  field  currently  in  production  
The  sensitivity  argument  
y Any  modeling  of  uncertainty  is  irrelevant/impossible  
without  a  decision  or  prediction  goal  
 
y Need  for  an  understanding  and  discovery  of  what  
impacts  flow  processes  and  decision  variables  
y More  than  just  a  computational  issue  !  
 
y Challenge:  
y Flow:  very  non-­‐linear  process  
y Most  important  and  impacting  variables  are  discrete  
 
 
The  rejection  argument  

y Karl  Popper  (1959):  physical  processes  are  laws  that  


are  only  abstract  in  nature  and  can  never  be  proven  
correct,  they  can  only  be  disproven/falsified  with  
facts  or  data  
 
y Popper-­‐Bayes   P(model|data)    P(data|model)  P(model)
 
Application  to  reservoir  case  study  

New  well  planned  


P3  
P1  
P4  

P2  

West-­‐Coast  Africa  (WCA)  slope-­‐valley  system  

Data  courtesy  of  Chevron  


Sensitivity  
y ͞ZŽĐŬ͟ƵŶĐĞƌƚĂŝŶƚLJ  
y Depositional  model  (Training  Image)  
y Spatial  uncertainty  (for  given  depositional  model)  
y Kv/Kh  ratio  
 
 
y ͞&ůƵŝĚ͟ƵŶĐĞƌƚĂŝŶƚLJ  
y Residual  oil  saturation  
y Maximum  water  relative  permeability  value  
y Water  Corey  exponent  
 
What  matters  for  prediction  ?  
Generalized  Sensitivity  Analysis  (GSA)  
underlying  principle  
Dim.  Reduction  
classification  

A  big  modeling  box  


  C1  
input   output  
Geology/geophysics   C2  
parameters   response  
Stochastic  
Flow   C3  

A  measure  of  sensitivity  is  the  difference    between  the  


frequency  distributions  of  input  parameters  per  each  class  
Distance-­‐based  GSA  

m  
Proxy  flow   Response  r  
Structure   model  complexity  
Rock   p  
Fluid   Time  
stochastic  

1 1

0.8
pi   0.8
pj  
0.6
CDF  

0.6
CDF  
cdf

cdf

0.4 0.4

0.2 0.2

0
0 0.2 0.4 0.6 0.8 1
0
2 2.5 3 3.5 4
Dim.  Reduction  
KvKh watExp
Kv/Kh   Corey  Water  Exp   classification  
Generalized  sensitivity  
any  parameter,  any  response  
y A  measure  of  sensitivity  is  the  L1  norm  difference  
between  a  class-­‐conditional  and  marginal  cdfs  
  pi  =  corey  exponent   pj  =  TI  ;  pi  =  corey  exponent    
TI|krwMax - class # 1
1 1

  FǷ(pi ) FǷ(pi | p j , ck )
0.8 0.8

 
cdf  

0.6 0.6

 
cdf

cdf
0.4 0.4 FǷ(pi |ck )
  0.2 0.2

  FǷ(pi |ck )
0 0
0 0.2 0.4 0.6 0.8 1 TI1 TI3 TI8 TI9 TI10 TI13
x TI
 
y A  measure  of  interaction  sensitivity  is  the  L1  norm  
difference  between  a  conditional  class-­‐conditional  and  
conditional  cdfs  
Application    
to  WCA   watExp
Training  Image  
Max  Water  rel  perm  
krwMax
 
Water  Corey  exponent  TI
 
Kv/Kh  
KvKh
 
Res.  Oil  Sat.  
SOWCR
  Sensitive
NotSensitive

  0 0.5 1 1.5 2 2.5

y 28  wells,  Response  =  oil  production  in  20  producers  


y Two  classes,  kernel  k-­‐medoid  clustering  
y Proxy  simulator  :  streamlines,  gave  same  classification  
as  Eclipse,  150.000  cells  
y 110  reservoir  models,  5  parameters  (rock  +  fluid),  total  
CPU  =  700  min.  
 
Conditional  Interaction  
'r KvKh|SOWCR
krwMax|watExp
'Kv / Kh| Sowcr TI|SOWCR
SOWCR|TI
watExp|SOWCR
SOWCR|watExp

TI  watExp TI|krwMax
watExp|krwMax
Wat  exp  
krwMax SOWCR|KvKh
KvKh|TI
Krw  Max  TI watExp|TI
KvKh|krwMax
KvKh  
KvKh
krwMax|KvKh
TI|KvKh
SOW  CR  
SOWCR
Sensitive KvKh|watExp
NotSensitive
TI|watExp
0 0.5 1 1.5 2 2.5
'r krwMax|TI
krwMax|SOWCR
'Kv / Kh SOWCR|krwMax Sensitive
watExp|KvKh NotSensitive
‡ Interaction  is  asymmetric     0 0.2 0.4 0.6 0.8 1

‡ one-­‐way  sensitivity  often  not  fully  informative  


 
Rejection  
Geosciences  are  interpretative  sciences  
  Depositional  model  
  Type  of  fracture  hierarchies  
  Rock  Physics  model  
  Fault  Hierarchy  

P Model,Scenario|Data ¦  P(M| Scenario ,D)    P Scenario |D


k
k k

                               P Scenariok|D :  to  reject  scenarios  without  any  HM


                               P(M| Scenariok , D) :    HM  per  acceptable  scenario  k
Data:  geology  and  production  

geological  
scenario    
uncertainty:    
3  training  images  
TI1:  50%   TI2:  25%   TI3:  25%  

Production  
Data:  
Water  rate/well  
Two  modeling  questions  

               P Model,TI|Data ¦  P(M|TI ,D)    P TI |D


k k
k

P TIk|D :  reject  data-­‐inconsistent  training  images

P(M|TIk , D) :  sampling  with  the  remaining  ones  


Generate  initial  ensemble    
of  180  scoping  models  
TI1:  50%   TI2:  25%   TI3:  25%  
Production  data  &  180  Scoping  runs  
Well  1   Well  2  
Water  rate  

Well  3   Well  4  

Time/Days  
Trying  to  falsify  TIs  with  data  
represent  data  in  lower  dimensions  using    
multi-­‐dimensional  scaling  
Eigencomponent  2  

Production  data  
TI1  responses  
TI2  responses  
TI3  responses  

Eigencomponent  1  

MDS:  distance  =  difference  in  water  rate  response  for  all  wells  
9  dimensions  =  99%  of  variance  
f  (Data  |  TIk  )  
Kernel  density  estimation  in  9D  
for  TI1   for  TI2   for  TI3  

Water  rate  data  

f data|TIk P TIk
P TIk|Data #        
¦  f data|TI P TI
k k k

P TI1|Data 0.8%        P TI2|Data 38.5%        P TI2|Data 60.7%        


History  match  for  each  TI  
Regional  probability  perturbation  
Why  regional  PPM?  
     Geological  realism  
     Works  for  facies  models  
     Easy  optimization  with  region  parameters  

Streamline  geometry    
at  final  time  step  
Example  of  region  geometry  
History  match  results  for  all  TIs  
CPU:  Average  of  24  flow  simulations/model  
 
A  few  history  matches  
From  TI2   From  TI3  

Notice  the  absence  of  any  region  artifacts  


Rejection  sampler  on  TI  and  facies  

1. Draw  randomly  a  TI  from  the  prior  


 
2. Generate  a  single  geo-­‐model  m  with  that  TI  
 
3. Run  the  flow  model  simulator  to  obtain  a  response  d=g(m)  
 
4. Accept  the  model  using  the  following  probability  
 
§ RMSE(dobs , g(m)) ·
p exp ¨  ¸
© 2V 2
¹
Rejection  sampler  results  
Prediction  in  newly  planned  well  for  
next  1  year    
1000
Rejection
PPM
800
Water Rate, stb/day

600
P10  

400

200
P50  
0 P90  
2000 2100 2200 2300
Time, days
Comparison  
Runs/  
P(TI1|D)   P(TI2|D)   P(TI3|D)  
model  

Method   1%   38%   61%   24  

Rejection  
3%   33%   64%   250  
Sampler  

Further  speed-­‐up  by  realizing  that  


‡ HM  per  scenario  had  little  impact  on  reduction  of  
uncertainty  
‡ use  of  proxy  flow  models  because  only  relative  likelihood  
accuracy  is  needed    
Some  observations  
y There  is  no  need  for  a  history  match  to  get  a  good  
prediction  (this  is  case  dependent)  
 
y No  need  to  run  full-­‐physics  on  all  models  
y Sensitivity:  proxies  may  provide  accurate  classification  
y Rejection:  only  relative  likelihood  is  needed  
 
y Increased  importance  on  providing  geological  
uncertainty  through  multiple  scenarios    
The  importance  of    
quantitative  geological  modeling  
variogram  

MPS  

Boolean  

Surface  based  

Process  based  
The  missing  link  
y Geological  interpretation:  attempting  to  
understand  the  genesis  and  process  of  past  
deposition  
  ?  
y Geostatistics:  attempting  to  model  the  
geometries  currently  present  with  a  practical  
application  in  mind  

Two  challenges  
1. What  methodology  bridges  this  gap?  
2. If  so,  how  to  bridge  this  gap?  
Limitation  of  covariances  
data   model  

1   2   3  

Variograms  EW   Variograms  NS  


1.2   1.2  
1  
2  
0.8   0.8  
3  
0.4   0.4  

0   10   20   30   40   0   10   20   30   40  
Training  images  
From  Boolean   From  high  resolution  seismic  

From  process-­‐based  models  


data   Conditioning  
Training  image  

Geostatistical  model  
High  performance  
Training  image  

Geostatistical  model  

4.5  million  cells,  7  seconds   1  million  cells,  1  second  


Honarkhah,  M.  and  Caers,  J.  (2012)  Math.  Geosc.,  44:651ʹ672.  
Direct  pattern-­‐based  simulation  of  non-­‐stationary  geostatistical  models  
 
Pejman  Tahmasebi  et  al.  (2012)  Comp.  Geosc.,  16:779ʹ797.  
Multiple-­‐point  geostatistical  modeling  based  on  the  cross-­‐correlation  functions  
 
Fenwick,  D.,  Scheidt,  C.,  and  Caers,  J.  (2012)  submitted  
A  distance-­‐based  generalized  sensitivity  analysis  for  reservoir  modeling  
 
Park,  H.,  Scheidt,  C.  Fenwick,  D.  Boucher,  A  and  Caers,  J.  (2012)  submitted  
History  matching  and  uncertainty  quantification  of  facies  models  with  multiple  
geological  interpretation  
 
Scheidt,  C.,  Renard,  P  and  Caers,  J.  (2012)  submitted  
Uncertainty  Quantification  in  Inverse  Problems:  Model-­‐based  versus    
Prediction-­‐Focused    
 
Aydin,  O.  and  Caers,  J.  (2012)  submitted  
Image  transforms  for  determining  fit-­‐for-­‐purpose  complexity  of    
geostatistical  models  in  flow  modeling  
 
 
PDFs  available  

Vous aimerez peut-être aussi