and Well Performance Prediction in Water-

clangedbivalveΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 3 χρόνια και 7 μήνες)

73 εμφανίσεις

Using
Neural Networks

for Candidate Selection
and Well Performance Prediction in Water
-
Shutoff Treatments Using Polymer Gels

A. Saeedi
*
, K.V. Camarda,

J.T. Liang

Chemical & Petroleum
Engineering Department

SPE 101028

* Now with Chevron Corp.

Water
-
shutoff Treatments



World wide,
> 75 billion barrels/yr

of water produced
from oil and gas fields.


Averaging > 3 barrels of water for each barrel of oil
produced.


Estimated disposal costs
~ 40 billion US dollars/yr
.


Financial & Environmental Incentives

Polymer Gels for

Water
-
shutoff Treatments


Used extensively in field applications to suppress
excess water production.


Candidate selection critical to the success of gel
treatments.


Anecdotal guidelines for candidate
-
well selection
resulting in inconsistent outcomes.

Objective of Study


Develop a methodology using
neural
networks

to identify candidate wells
based on
predicted outcomes

for
polymer gel treatments.


22 wells

treated with polymer gels in
Arbuckle formation in Kansas

were used to
develop the neural networks.

A.

Arbuckle Structure Map

B.

Arbuckle Cum. Oil Production by


County (MMBO)


Franseen, 2003

Franseen, 2003

Arbuckle Formation in Kansas


Main oil producer in Kansas (~2.2 billion barrels).


Fractured
-
controlled karstic reservoirs with porosity
and permeability influenced by basement structural
patterns and subaerial exposures.


Strong bottom aquifer with high WOR.


Open
-
hole completion at top 1/3 of pay zone to avoid
water coning.


Reservoir poorly characterized.

Gel Treatments in Arbuckle


Cr(III)
-
HPAM gels very successful in treating high
water
-
cut well.


>250 wells treated with Cr(III)
-
HPAM gels.


Candidate selection based mainly on vendor’s past
experience (reservoir poorly characterized).


Neural Networks for

Candidate Selection


Candidate selection based on predicted
treatment outcomes using pre
-
treatment data
is a
pattern
-
recognition

problem.


Neural network

is a powerful tool for solving
pattern recognition problems.


Artificial Neuron

After Mohaghegh (2000)

Three
-
layer Neural Network

Input Layer

Hidden Layer

Output Layer

Weight

Weight

Neural Network Development


Data divided into two sets.


Training set used to adjust connecting weights.


Verification set used to evaluate the trained
network.


Verification set not seen by network during
training.

Supervised vs. Unsupervised
Neural Networks

Unsupervised Neural Networks


Training set consists of input patterns only.


No feedback is provided.

Supervised Neural Networks


Training data consist of many pairs of input an
output patterns.


Feedbacks are provided during training.

Feedforward
-
Backpropagation Neural
Networks
(Supervised Training)

Weight

Weight

Feedforward

Backpropagation

Input #1

Input #2

Input #3

Input #4

Input #5

Input #6

Output #1

Output #2



An iterative process to minimize error.



Training Step

Neural Network

Feedforward

Backpropagation

Adjusted

Weights

Input

Output

Example Training Set

Well

No.

Input #1

(Depth to the top
of the Arbuckle)

Input #2

(Oil Prod.
Rate BT)

Input #3

(Water Prod.
Rate BT)

Output

(Cum. Oil Prod.
one month AT)

1

3,670

5

475

2,625

2

3,825

9

1,125

3,519

3

3,824

3

385

2,054



Verification Step

Input

Trained

Neural Network

Predicted

Output

Actual

Output

Accuracy

Evaluated

Example Verification Set

Well

No.

Predicted

Output*

Actual

Output

Error**
Fraction

4

1,219

1,557

0.22

5

1,135

962

0.18

* Cumulative oil production one month after treatment

**Error Fraction = |(Actual Value


Predicted Value)|/Actual Value



Predictive Step

Input

Trained

Neural Network

Outcome
Prediction

Example Outcome prediction

Input #1

(Depth to the
top of the
Arbuckle)

Input #2

(Oil Prod.
Rate BT)

Input #3

(Water Prod.
Rate BT)

3,467

7

756


Neural Network

Predicted

Outcome

(Cum. Oil Prod.
one month AT)


1,000

Neural Network
1*

Input Parameters

Output Parameter

Normalized Latitude and Longitude
of the Well

Cumulative Oil Production
One Month after Treatment

(BBLS)

Total Depth of the Well (ft)

Depth to the Top of the Arbuckle (ft)

Oil Production Rate before
Treatment (BPD)

Water Production Rate before
Treatment (BWPD)

*
18 Training Sample Wells + 4 Verification Sample wells

Neural Network
1

Neural Network
1

Well
No.

Actual
Value

Neural
Network’s
Prediction

EF
*

AEF
**

1

1862

1572

0.155

0.16

2

3519

2533

0.280

3

2054

1616

0.213

4

1557

1561

0.003

*
EF =
|(
Actual Value


Predicted Value)
|
/Actual Value

**
AEF: Average EF

Neural Network
2*

Input Parameters

Output Parameter

Total Depth of the Well (ft)

Cumulative Oil Production
Three Months after Treatment

(BBLS)

Depth to the top of the Arbuckle (ft)

Oil Production Rate before
Treatment (BOPD)

Water Production Rate before
Treatment (BWPD)

*
16 Training Sample Wells + 4 Verification Sample wells

Neural Network
2

Neural Network
2

Well
No.

Actual
Value

Neural
Network’s
Prediction

EF

AEF

1

2883

3021

0.048

0.10

2

4081

2995

0.266

3

2933

3134

0.069

4

3568

3585

0.005

Conclusions


With only pre
-
treatment data as input parameters,
the neural networks in this study can accurately
predict cumulative oil production one and three
months after gel treatment.


Neural networks

allow the candidate selection to be
based on the accurate predictions of
treatment
outcomes

using only
pre
-
treatment data
.


This method is far superior to the anecdotal
guidelines based solely on vendor experience.


Thank you!