2008 Steinar M. Elgsæter phd defence: Modeling and optimizing the offshore production of oil and...

Post on 22-Jun-2015

783 views 2 download

Tags:

description

Steinar M. Elgsæter phd defence presentation. A data-driven approach to production modeling and model updating is suggested. Simple model structures are inferred from observations of measured production, motivated by the concepts of system identification and a desire for models which are sufficiently accurate while being easily solved numerically. A production model needs to be updated periodically to reflect changes in reservoirs and wells. Updating models against observations of recently measured data describing normal operations is suggested. As data describing normal operations may have low information content, parameters estimated may be subject to significant uncertainty. Bootstrapping is considered as a means of estimating uncertainty in fitted parameters. Methods for the explicit treatment of such uncertainty based on Monte-Carlo analysis are investigated. Methods for estimating lost potential due to uncertainty, result analysis, excitation planning and choosing active decision variables are suggested. For full pdf download, see also sites.google.com/site/steinarelgsaeter/

Transcript of 2008 Steinar M. Elgsæter phd defence: Modeling and optimizing the offshore production of oil and...

1

Modeling and optimizing the offshore production of oil and gas under uncertainty

Steinar M. Elgsæter - October 14, 2008

2

Thesis introduction

• supervised by Professor Tor Arne Johansen (NTNU) and Dr.Ing Olav Slupphaug (ABB),

• funded by ABB, Norsk Hydro (later StatoilHydro) and the Norwegian Research Council,

• work conducted in the period 2005-2008,• three conference papers presented,• two journal papers submitted,• one patent application submitted.

3

”slow” dynamics on the timescales of months and years

”fast” dynamics on the timescales of hours and days

4

production

disturbance

decision variables

measured output:profits and capacities

production optimization timescale: hours and days

5

Model-based production optimization

Production

Disturbances Decision Variables(valves)

Measured output (Profits and capacity utilization)

Production constraints(capacities) and object function(profit measure)

Production optimization

Production Model

Model parameters:

Watercut,GOR,well potential etc.

current practice: an ”engineering” approach to modeling•detailed physical models•emprical relations for closure•commerical simulators

6

Challenges of current practice

1. challenging production modeling– complexity of systems considered

– multiphase flow

– measurement difficulties (such as multiphase flow meters)

– disturbances (reservoir depletion)

2. model updating (high update frequency, laborious)

3. numerical and optimization issuses (numerical stability,identifiability,convexity,run-time)

7

Part I: A data-driven approach to production modeling and model updating

8

production data contains information that can be exploited in optimization

9

A data-driven approach to production modeling and model updating

Production

disturbances decision variables(valves)

measured output (Profits and capacity utilization)

Parameter andstate

estimation

fitted parameters and states

Production model

-

Difference (residual)

model parameters

Production constraints(capacities) and object function(profit measure)

Production optimization

Production Model

A ”closed loop”

modeled output

10

Challenge

• data describing normal operations are usually not sufficiently informative, models fitted to data are subject to parameter uncertainty

11

Part II: Methods for uncertainty analysis and uncertainty handling

12

Quantifying uncertainty

• bootstrapping– multiple-model

– computational

– based on data-set resampling

• models– locally valid

– simple ”performance curves”

– motivated by concepts of system identification

13

realized potential

Uncertaintydue to low information content in data

max

current

?

12

3

Experiments

Optimization

Eliminating uncertainty is not a practical option

Cost

14

An approach for structured uncertainty handlingmy thesis proposes a five-element strategy for

optimization with uncertain models

1. result analysis

2. excitation planning

3. active decision variables

4. operational strategy

5. iterative implementation and model updating

15

1.Result analysisrealizedpotential

uncertaintydue to low information content in data

max

current1

Different simulated plausible outcomes

16

1

2. Excitation planning

realizedpotential

uncertaintydue to low information content in data

current2Experiment

Cost

Simulated plausible outcomesof optimization without exictation

Simulated outcome of excitation

Simulated plausible outcomesof optimization with exictation

17

3. Active decision variables

realized potential

uncertaintydue to low information content in data

current1

Simulated change in all decision variablesSimulated change in active decision variables

18

4. Operational strategy

When models are uncertain, a target setpoint can be infeasble when implemented

An opertational strategy is an iterative implementation of setpoint change while monitoring profits and constraints

19

4. Operational strategy...

Production

Decision Variables

Measured output

Parameter andstate

estimation

Fitted parameters and states

Production optimization

Operational strategy

Target

20

realized potential

uncertaintydue to low information content in data

max

current1

2

3

5.Iterative implementation and model updating

4

optimize

update model and re-optimizeupdate model

and re-optimize

21

Perf o rm ex c ita tion p lanning

Perf o rm produc tion optimiz a tion

O ptiona lly : s e lec t ac tiv e dec is ion

v ariables

Implement s e tpo int c hange s ugges ted

by produc tion optimiz a tion ac c ording to

opera tiona l s tra tegy

Is the c os t/benef it tradeo f f o f any

planned ex c ita tion f av orable?

Implement p lanned

ex c ita tion

Y es

Update mode l: Es timate parameters

and parameter unc erta inty

Is res ult ana ly s is f av orable?

No

Y es

W ait until new data bec omes av ia lable

No

Perf o rm res ult ana ly s is

Combined the elements provide a framework for optimizing oil and gas production with uncertain models

22

Results

• Methods applied to two sets of real-world production data from North Sea oil fields

• Simulations indicate:– promising active decision variable candidates found

– in simulations 30-80% of potential profits were realized using uncertain models in combination with the suggested framework

23

Results: Active decision variables(1)

24

Results: Active decision variables(2)

25

Discussion and Conclusions

26

I. Data-driven modeling and model updating

• adresses weaknesses of current practice:– models easy to design– models updated with less effort

• this may increase frequency at which production optmization can run

– models are less prone to issues of convexity, numerical stability, identifiability and computational effort.

– models especially well suited for iterative optimization (each iteration reveals information)

• challenge– requires measurement maintenance and may be prone to issues of

low information content in data

27

II. Framework for optimizing production with uncertain models • a method that can exploit current real-world data as a

starting point• iterative approach ideal for combination with low-

maintenace data-driven models• analog to the current approach

– but: decision support based on objective analysis at every step of decision-making process

• relationship between current manner of operation, uncertainty and production optimization is made explicit

28

Further work

29

A ”low-hanging fruit” for practicioners

• perform a ”proof of concept” experiment– implement setpoint change according to active decision variables

method

• an experiment that – will be profitable with high confidence

– validates the ”control” approach of this thesis

30

Thank you for your attention