2008 IFAC World Congress: Oil and gas production optimization - lost potential due to uncertainty, Steinar M .Elgsæter

1

Oil and gas production optimization; lost potential due to uncertainty

Steinar M. Elgsæter, Norwegian University of Science and Technology (NTNU)

Olav Slupphaug, ABB

Tor Arne Johansen, NTNU

2

Production optimization• Day-to-day optimization of production

Multiphase/Gas-lift/Production chokes/Well tests

u,d

y

3

Current practice and challenges

• Production very difficult to model (multiphase pipe flow)

• Current practice: optimization with rigourous ”engineering” models, with varying success

• Fields operated such that there is little variation/excitation in data against which to fit models*

• ”feed forward” production optimization

* Elgsaeter, S.M., Slupphaug, O. And Johansen, T.A. (2007), Challenges in Parameter Estimation of Models for Offshore Oil and Gas Production Optimization, International Petroleum Technology Conference, Dubai, 4.Dec 2007, IPTC11728

4

Excitation can ”close the loop”, but has costs and risk

* Elgsaeter,S.M., Slupphaug, O. and Johansen, T.A.(2008), Production Optimization; System Identification and Uncertainty Estimation, Intelligent Energy ’08, Amsterdam, Feb 2008, SPE112186

5

Deadlock

Practicioners do not see the value of excitation and hence will not allow us to implement it

We need to implement excitation to doucment its value

6

”Data-driven” models• ”Stationary” (no dynamic terms)• Motivated by the concepts of system identification

– Simple/pragmatic/inferred from data

• Locally valid • Model the oil, gas and water produced from each well, fitted to

measured total rates of oil, gas and water• Two versions for comparison:

”well test””production choke kernel” (modeled disturbance )

”gas-lift kernels”(decision variables)

Fitted paramters

7

Real-world data set

• Offshore North Sea field

• 20 gas-lifted platform wells

• Data:– 5 months

– ”normal operations”

– Apparenlty some excitation but not excitation is not aboundant

8

Quantifying model uncertainty• Bootstrapping

– ”Multiple model”

– Computationally intensive

– Re-solves parameter estimation problem many times with resampled data-sets

– Multi-variable uncertainty estimates

• Result: – ”uncertainty fans” which

express information content or lack thereof in data

– Several houndred similar but different paramters which are all plausible

9

What is the cost of not being able to differentiate among the models identified?

10

Production optimization

• Profits(M) and constraints(c) are related to modeled rates of oil, gas and water– Profit: total produced oil– Constraint 1: max total produced gas (<= current rate)– Constraint 2: max total produced water (<= current rate)– Constraint 3: (Uprc is design variable)– Constraint 4: (flow assurance on some wells)

Model

Constraints

Profit

11

Realized and lost potential

• Approach:estimate distributions for Po and Lu using Monte-Carlo simulation

• If there is little excitation and model uncertainty as a result, we cannot expect to realize all the potential of production optimization

12

Monte-Carlo simulation of loss due to uncertainty

• Solve production optimization problem for all models identified using bootstrapping

• For each plausible model identified using bootstrapping (model t)

– For each plausible model identified using bootstrapping (model f)

• Simulate how much of the potential production optimization is able to realize if we implement decision variable found by optimization with (model f) when in reality production is described by (model t)

13

• Results model-dependent• 1.order model results in more

conservative losses (and with less variation)

• Lu and Po depend on choice of Uprc,

the ratio between the two less so• Analysis indicates that

– a potential in excess of 2% exists– as much as half of this potential

may be unrealized unless excitation is introduced

Blue:1.order(linear) modelGreen:2.order modelTop: upper 95. percentileCenter: meanBottom lower 95.percentile

ResultsPo(%)

Lu (%)

Lu/Po

14

Conclusions

• A way to estimate the value of closed-loop modeling/optimization that is consistent with data – attempt to ”let the data speak”

• A tool for communicating the value of closing-the-loop to non-control engineers. – The link between how a field is operated and the success of subsequent

production optimization is highlighted

– This analysis can hopefully convince practicioners to operate their fields in a manner which introduces more excitation

15

Thank you for your attention