Spatial variability of tight oil well productivity and the ... Justin Montgomery.pdf · forecasts and poor design choices for wells MIT Earth Resources Laboratory 2017 Annual Founding

Spatial variability of tight oil well productivity and the impact of technologyJustin Montgomery

PhD CandidateDepartment of Civil and Environmental EngineeringIn collaboration with Francis O’Sullivan and John Williams

MIT Earth Resources Laboratory2017 Annual Founding Members MeetingMay 31, 2017

Williston Basin of North Dakota was at the forefront of tight oil extraction but now faces economic uncertainty

MIT Earth Resources Laboratory2017 Annual Founding Members Meeting

Slide 2

Rising rig and well productivity suggest greater resilience than expected


Slide 3

2008 2010 2012 2014 2016

020

040

060

080

0

New−w

ell o

il pr

od. p

er ri

g (b

bl/d

)

050

100

150

200

250

Rig

cou

nt (#

of a

ctive

rigs

)

New−well oil prod. per rigRig count

2008 2010 2012 2014 2016

020

4060

8010

012

0

Mea

n ne

w−w

ell f

irst y

ear p

rod.

(Mbb

l)

010

020

030

040

050

060

0

New

wel

ls (w

ells

/qua

rter)

Mean new−well prod.New wells per quarter

Improvement of well productivity has been driven in part by changes in well and stimulation design


Slide 4

- Trends toward longer wells and larger stimulations (hydraulic fracturing) Increase in proppant (sand) per well over time

- Motivation for identifying impact:1. Forecast well productivity based on

anticipated changes2. Optimize wells

Another important dynamic is where wells are being drilled –“sweet-spotting” or “high-grading”


Slide 5

Well productivity heat map

Source: Schmidt, 2011

- Activity continuing to cluster in high productivity areas

- Motivation for identifying location influence:1. Need to control for this to accurately

understand impact of design changes2. Assess well portfolios and resource

economics based on location in field

How much of the improvement in well productivity is due to technology (design changes) vs location (sweet spotting)?


Slide 6

v Horizontal length of completed well

Technology Location v Target formation (Bakken/Three Forks)v Amount of water

injected to create fissures

v Amount of proppant(sand) injected to “prop” fissures open

v Latitude/longitude of wells

- Big public datasets available (Frac Focus, North Dakota Mineral Resources)- Can we use econometrics/machine learning to understand and make predictions?

Current regression models to understand the influence of technology on productivity


Slide 7

- Nonspatial linear regression (NS)- Fixed Effects (FE), such as county-level used by EIA

- Issues:- Not spatially granular enough- Residuals are spatially autocorrelated

à Omitted variable bias

Regression-kriging provides an appropriate tool for distinguishing between impact of location and technology


Slide 8

Estimate trend with linear regression

Estimate spatial correlations by kriging residuals

Remove interpolated spatial component

No

Estimates for model parameters

Converged?

Yes

Technologytrend

Spatial component

Start

RK improves accuracy (in 10-fold cross validation) compared to currently used regression models


Slide 9

0 100 200 300 400

010

020

030

040

0

Predicted first year prod. (Mbbl)

Actu

al fi

rst y

ear p

rod.

(Mbb

l)

Nonspatial regression

MASE = 0.938

0 100 200 300 400

010

020

030

040

0


Actu

al fi

rst y

ear p

rod.

(Mbb

l)

MASE = 0.873

0 100 200 300 400

010

020

030

040

0


Actu

al fi

rst y

ear p

rod.

(Mbb

l)

MASE = 0.62

Fixed effects-county Regression-kriging

Existing regression models overestimate the role of technology relative to location


Slide 10

Technology

Location

Overestimating the impact of technology leads to overoptimistic forecasts and poor design choices for wells


Slide 11

Forecasts for 2018 designsKey findings

1. Regression-kriging improves prediction accuracy

2. Shifts in well design and drilling location have contributed equally in recent years

3. County-level fixed effects inadequate to detect sweet-spotting à EIA forecast is likely overoptimistic

4. Current models encourage over-stimulation of wells

Future work


Slide 12

- Apply to other unconventional fields

- Predict decline rates

- Use to develop improved field-scale economic models

Thank you! Questions?


Slide 13

- Thank you to MIT Energy Initiative for supporting this research

- Full paper is:Montgomery, J. B., & O’Sullivan, F. M. (2017). Spatial variability of tight oil well productivity and the impact of technology. Applied Energy, 195, 344-355.

US tight oil production growth has demonstrated the potential of shale and other unconventional formations – Combined output from three of the main US plays alone is now equivalent to the total output of China or Canada

14Source: F. O’Sullivan, United States Energy Information Administration, HPDI Production Database

Illustration of crude oil production growth from some select major U.S. unconventional oil plays since 2005MMbbls of oil per day

0

1

2

3

4

5Eagle Ford

Bakken

PermianCombined, the Bakken, Permian and Eagle Ford have added more than 4 MMbbls per day of production to US output over the past 5 years

Tight oil plays now support more than 50%

of US Crude

15

Surface trend analysis (productivity fit to polynomial of coordinates)

Fixed effects – county or township level

Some other approaches that have been used to control for location

Results of regression kriging – Productivity forecast with typical well designs for 2018


Slide 16

−104.0 −103.5 −103.0 −102.5

47.0

47.5

48.0

48.5

49.0

Longitude

Latitude

[40,70)[70,100)[100,130)[130,160)[160,190)[190,220)[220,250]

Predicted firstyear prod. (MBbl)

17

Each model provides a good fit to the mean productivity over time

6080

100

120

140

Mea

n ne

w−we

ll firs

t yea

r pro

d. (M

bbl)

2012 2013 2014 2015 2016

NonspatialFESTASEMRK

ActualActual−IQR

18

Training only with data from early wells shows that mean production can be reliably forecasted based on changes in location and technology

7580

8590

9510

010

511

0

Mea

n ne

w−w

ell f

irst y

ear p

rod.

(Mbb

l)

2012 2013 2014 2015 2016

NonspatialFESTASEMRKActual

Training Hindcastvalidation

19

8090

100

110

120

130

Mea

n ne

w−w

ell f

irst y

ear p

rod.

(Mbb

l)

2012 2014 2016 2018

NonspatialFESTASEMRKActual

Using EIA projected designs for 2018

These models are useful for forecasting production and economics of future wells – Important differences between RK and existing approaches such as FE become clear

20

Differences in impact attributed to different parameters

NS FE STA SEM RK

Shar

e of

pro

duct

ivity

incr

ease

(%)

020

4060

8010

0

Location (acreage quality)

Proppant

Water

Lateral Length

NS FE STA SEM RK

Shar

e of

pro

duct

ivity

incr

ease

(%)

020

4060

8010

0


Proppant

Water

Lateral Length

NS FE STA SEM RK

Shar

e of

pro

duct

ivity

incr

ease

(%)

020

4060

8010

0Location (acreage quality)

Proppant

Water

Lateral Length

NS FE STA SEM RK

Shar

e of

pro

duct

ivity

incr

ease

(%)

020

4060

8010

0


Proppant

Water

Lateral Length

NS FE STA SEM RK

Shar

e of

pro

duct

ivity

incr

ease

(%)

020

4060

8010

0


Proppant

Water

Lateral Length

21

– Spatial trends and patterns result from physical processes over long lengths of time– Occur at various scales (e.g. macro: formation thickness, grain size/porosity, thermal maturity;; micro: natural fractures)

– Geological controls may be poorly understood or hard to quantify

Location is important because key geological controls on production vary spatially across basin

22

Depth (ft)

Location is important because key geological controls on production vary spatially across basin

Amount of proppant has been increasing over time and is correlated with productivity


Slide 23

24

Water trends

25

Lateral length trends

26

Definition of models:

Multiple linear regression model:

Ordinary least squares:

Multiple linear regression model with variance-‐covariance matrix:

Generalized least squares:

27

One approach to estimating the effect of technology on productivity is linear regression with ordinary least squares – Omitted-‐variable bias is a problem though

28

Technology

ProductivityGeological controls (omitted variable)

More realistically:

Bias of Estimate:

Bias is introduced if:

29

Evaluating the models

Moran’s I-‐1 10

Highly dispersed

Highly clusteredRandom

Moran’s I to measure spatial autocorrelation:

Back transformation:

Model accuracy:

10-‐fold cross validation:

Comparison of models:

Spatial weights matrix W: Coefficient estimates:

0 10 20 30 40 50

0.00

0.05

0.10

0.15

0.20

k−neighbor neighbor

Mea

n sp

atia

l wei

ght

SEMRK

Comparison of models’ estimates of technology and location driven improvement in productivity

Mea

n ne

w−w

ell p

rodu

ctiv

ity

(inde

xed

to Q

1−20

12)

11.

11.

21.

3

2012 2013 2014 2015 2016

FEFE−technology constant

Mea

n ne

w−w

ell p

rodu

ctiv

ity

(inde

xed

to Q

1−20

12)

11.

11.

21.

3

2012 2013 2014 2015 2016

STASTA−technology constant

Mea

n ne

w−w

ell p

rodu

ctiv

ity

(inde

xed

to Q

1−20

12)

11.

11.

21.

3

2012 2013 2014 2015 2016

SEMSEM−technology constant

Mea

n ne

w−w

ell p

rodu

ctiv

ity

(inde

xed

to Q

1−20

12)

11.

11.

21.

3

2012 2013 2014 2015 2016

RKRK−technology constant

Documents

Spatial variability of tight oil well productivity and the ... Justin Montgomery.pdf · forecasts and poor design choices for wells MIT Earth Resources Laboratory 2017 Annual Founding