Assisted History Matching and Optimization

Remus Hanea
 Some historical data

• Ph.D. at TU Delft in Applied Mathematics - Inverse modeling for air quality (2006)

• TNO/TU Delft joint position (2006 - 2013) (researcher and Asssitent Professor) – Assisted
History Matching and Optimization

• Statoil, Bergen, Norway – RDI (Research, Development and Innovation)

 Reservoir Engineering and Simulations

• «Reservoir engineering is a branch of petroleum engineering that applies scientific

principles to the drainage problems arising during the development and production of oil
and gas reservoirs so as to obtain a high economic recovery»

• «Reservoir simulation is an area of Reservoir engineering in which computer models are

used to predict the flow of fluids (typically, oil, water, and gas) through porous media»

• Dynamics are described by the flow simulations

Petroleum vs. Atmospheric/Oceanography

Differences

Atmospheric/Oceanography Reservoir
• Unstable chaotic dynamics • Stable diffusive transport
• State estimation problem • Parameter estimation problem
• Recursive problem • Prediction is determined given
• 4D VAR the model
• ES

Similarities

• Dynamical prediction problem

• Uncertainty in the predictions
• Conditioning on lots of data
My first impression about reservoirs and trapped oil
 I was almost right
rocks

pores

6
 Important properties

• porosity – the fraction of the rock

that contains fluids

• permeability - the ability of fluids

to flow from one pore to another

high permeable
low permeable

7
 «Holy grail»

Probabilistic History Matching and uncertainty reduction:

- Provides multi-model history match for improved uncertainty description

– Predicts of the future behavior of the reservoir in existing and new wells with increased
confidence

– Maintains geological realism

– ?
Incorporates of all quantifiable information (seismic, gravity, tracer etc.)
 Model
• The discrete model
x k = f ( x k −1 , θ )
where
x = [pT sT ]T
• There is no model error

• Uncertainty in model parameters (constant) – Model calibration

• Combine the state variable and the uncertain parameters to

obtain an augmented state vector

9
 Geological uncertainties

• Structural uncertainties:
• surfaces (top/bottom)
• faults

• Facies uncertainties

• Petro-physical uncertainties
• Perm/poro/NTG
• Rel.perm
 Observations

• The relation between the measured data y and the state variable x is
described by a nonlinear function h

y k h ( x k ) + ηk
=

• Conventional observations :
• Bottom Hole Pressure (BHP)
• Production rates
• Unconventinal observations:
• Seismic, tracers, gravity, satellite
 Inverse problem – Data assimilation

• Variational methods ⇔ Minimization of the cost function ⇔ Gradient based method ⇔

Requires an adjoint
 Inverse problem – Data assimilation
The adjoint «way of life»
 Inverse problem – Data assimilation
Life of the adjoint-deprived masses
 Inverse problem – Data assimilation

• Variational methods ⇔ Minimization of the cost function ⇔ Gradient based method ⇔

Requires an adjoint

• Sequential methods ⇔ No adjoint required ⇔ Update every time when a measurement is

available ⇔ Minimizing the variance of the estimation every analysis step.
 Filter, prediction, smoother

•
 Filter, prediction, smoother

•
 Ensemble Kalman Filter
N
1
x(t k +1 ) =
N
∑ i (tk +1 )
ξ f

i =1

ξ ia (t k ) = ξ i f (t k +1 ) + K (t k +1 )[ y o (t k +1 ) − H (t k +1 )ξ i f (t k +1 ) + vi (t k +1 )]
P f (t k +1 ) ≈ Pe (t k +1 ) = E[( x(t k +1 ) − x f (t k +1 ))( x(t k +1 ) − x f (t k +1 ))T ]
f

P a (t k +1 ) ≈ Pe (t k +1 ) = E[( x(t k +1 ) − x a (t k +1 ))( x(t k +1 ) − x a (t k +1 ))T ]

a
 Ensemble smoother

• Similar to EnKF, but with a single update with all data available, i.e., no sequential data
assimilation.

• Faster and easier to implement than EnKF (no restarts).

• Almost cost free (without iterations)

• Easy to implement

• Indicated for the parameter estimation problem

Security
Classificati
19 Classification: Internal 2012-12-06 on: Internal
- Status:
 Filtering problem

•
 Joint sequential smoothing
•
Multiple Data Assimilation (MDA)
 Challenges
• The dimension of the state vector (dynamic and static parameters) is huge

• The forward simulation takes a lot more time that the update. (95% to 5%)

• The dimension of the ensemble goes to our infinity = 100 (magic number)

• Small number of conventional (well) measurements

• Not directly related to the parameters of interest (observability / identifiability)

• Large amount of unconventional measurements (seismic , gravity)

• Almost unknown (very uncertain) and very complex prior knowledge

 Challenges
1) Initial step
− Geological uncertainties (properties, facies and structural) huge impact
− Generation of the initial ensemble
2) Forecast step / model update
− Need to be customized for each particular reservoir simulator
3) Update step / measurement update
− Spurious correlations (unphysical updates or updates that are not making sense)
− Collapse of the ensemble (divergence of the filter)
Solutions:
1) Meaningful parameterizations of the geological parameters and a tool of generating the initial
ensemble
2) No solution for this (or stick with only one simulator)
3) Inflation of the variance and localization (covariance localization or local analysis)
4) Integrated workflow
 Inflation trick
•
 Inflation trick

•
 Localization trick

• Covariance localization (covariance filtering) Houtekamer and Mitchell (2001);

Whitaker and Hamill (2002)

• Local analysis (Evensen, 2003; Anderson, 2003; Ott et al., 2004; Hunt et al.,
2007)
 Localization trick

• Covariance localisation: Hamill and Whitaker (2001); Houtekamer and Mitchell (2001)

• Local analysis: Evensen (2003); Anderson (2003); Ott et al. (2004)

• Smooth tapering in parallel ESRFs: Hunt et al. (2007)

• Adaptive methods: Anderson (2007); Bishop and Hodyss (2007, 2009)

• Dean Oliver 2012, bootstarping of the Kalman gains

• Al Reynolds 2010 , adaptive choice of the lenght

 Localization trick

• Localization in the EnKF is an ad-hoc modification of the analysis scheme

• One must use localization if the ensemble size is smaller than the model subspace
dimension

• Localization makes the analysis schemes suboptimal and therefore inconsistent

(sometimes - strongly) in regard to estimation of posterior covariance

• Localization makes it possible to recover the modal structure from observations

even with a rank-deficient ensemble
Norne field
 Norne field

Security
Classificati
31 Classification: Internal 2012-12-06 on: Internal
- Status:
 Norne field
Uncertain parameters:

• porosity, permeability, NTG, MULTZ (grid-based properties)

• MULTFLT - fault transmissibility multipliers (53 parameters)
• Rel. perm – End-point water and gas relative permeability of four zones (8 parameters)
• OWC Initial oil-water contact depth in 4 regions

In total we have 150000 uncertain parameters

Measurements used in the AHM process:

• Water injection rate
• Gas injection rate
• Oil production rate
• Water production rate
• Gas production rate
• RFT pressure
In total we have 2000 measurements
Oil production rates matching

Field permeabilty updates

Fault transmisibility multiplier

 Multiple realizations / The ensemble

• How to handle multiple realisations in daily

operations
• Model QC
• Management decisions (predictions)
• Well planning

• Where to start and finish with parameterisation &

uncertainty
 Reservoir management (cont’d)

Optimize drilling priority using Fast Model Update (FMU) and Ensemble Reservoir Tool
(ERT)

• Not only on the base case and also tacking the dynamics into account – link with a NPV
maximization (optimization tool)

• The key functions are the ability to open a well at a varying date in the schedule file.

• Ability to model constraints between the dates (e.g. minimum of 30 days between well1
and well2 opening due to drilling time).
 Reservoir management

Well placement under geological uncertainty

(replicate geo-steering) Wells “falling out” of the analysis (placed in
non-net)

• The aim is to optimize the planned well path

under this geological uncertainty

• This may have a significant impact of NPV

estimations

• We will not actually do any geo-steering, but will

try to adjust a planned well path in the
simulation model in a way consistent with geo-
steering
 Robust optimization

Robust optimization
• • Optimization (multiple realizations)
(base case)
− ICD, ICV (choke control)
− Well placement
− Number of wells
− Production strategies
− Drilling priority Well placement and priority
ICD

• Robust optimization (multiple realizations)

− Optimization under geological uncertainty

• IOR and accelerated recovery due to improved and optimized production strategy
 Robust optimization
• Three control strategies applied to a set of realizations.

Reactive Control: Nominal optimization: Robust Optimization:

• Based on a single realization. • Optimization on one realization and • Maximizing expected
• What we get from the today applying controls on all other realizations. NPV over all
workflows. • Repeat optimization procedure for all realizations.
realizations and reapply.

38
Run movie
www.statoil.com
 Filter, prediction, smoother

•
 Smoothing algorithms

• Different types of smoothing problems

• Since some information must be propagated backward in time, no obvoius algortihm

prevails

• Hence, a large number of different algorithms, compared with the filtering and prediction
problems
 Sequential smoother

• It runs along with a filter

• The really new concept in the smoother, with respect to the Kalman filter, is the use
(explicit or implicit) of the following cross-covariance matrices:

When the filter process the observations at time k to update the state at time k, the
smoother uses the same observations to update the states at times i < k. So, the above
define are the cross-covariances between the smoother forecast and the analasis at time i
with the filter analysis at time k.
 Sequential smoother

The sequential smoother can be derived by extending the filter state vector with past states,
e.g.:
 Sequential smoother

The extended covariance matrix has the form

45 8th EnKF workshop 2013-05-27

 Sequential smoother

The extended model is obtained:

46 8th EnKF workshop 2013-05-27

 Sequential smoother

The propagated state is:

47 8th EnKF workshop 2013-05-27

 Ensemble smoother

• The cross-coavriance matrix

• Ensemble-based approaximation

48 8th EnKF workshop 2013-05-27

 Ensemble smoother
• The EnKF analysis

The filter corrections are linear combinations of ensemble perturbations

• The smoother analysis

The smoother corrections are linear combinations of ensemble perturbations

49 8th EnKF workshop 2013-05-27

 Sequential smoother

• Almost cost free (without iterations)

• Easy to implement

• Easy to switch from fixed-interval to fixed-lag to fixed-point types

• Indicated for the parameter estimation problem

• Does not give a better estimation than the EnKF of the current time

• The choice of using the smoother is tailored to the specific field of application (based on practical
considerations)

50 8th EnKF workshop 2013-05-27

Robust reservoir management

• Three control strategies applied to a set NominalOptimization:

Reactive
Robust optimization:
Control:
of 100 realizations. • Optimization
Applied onrealizations.
to 100
Maximizing over one realization
all individual and
realizations.
• applying
No controls on all other realizations.
optimization
• All curves represent PDFs from 100
• Repeat optimization procedure for all
realizations
realizations and reapply.