Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Machine learning-based fracture-hit

detection algorithm using LFDAS signal


Ge Jin1, Kevin Mendoza2, Baishali Roy1, and Darryl G. Buswell1
https://doi.org/10.1190/tle38070520.1.
Downloaded 06/22/20 to 130.238.7.40. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Abstract Perforation cluster spacing is a key control parameter of plug-


Low-frequency distributed acoustic sensing (LFDAS) signal and-perf completion. Theoretically, cluster spacing should be
has been used to detect fracture hits at offset monitor wells during determined by reservoir permeability and stress anisotropy (dif-
hydraulic fracturing operations. Typically, fracture hits are manually ference between maximum and minimum horizontal stress). In
identified, which can be subjective and inefficient. We implemented low-permeability reservoirs, large cluster spacing designs may
machine learning-based models using supervised learning tech- drain the reservoir inefficiently due to sparsely distributed hydraulic
niques in order to identify fracture zones, which demonstrate a fractures. On the other hand, if the cluster spacing is too small,
high probability of fracture hits automatically. Several features are stress shadow interference between the nearby fractures may cause
designed and calculated from LFDAS data to highlight fracture-hit inefficient fracture growth, which leads to unnecessary completion
characterizations. A simple neural network model is trained to fit cost (e.g., Sebastian et al., 2015).
the manually picked fracture hits. The fracture-hit probability, as A recent study by Jin and Roy (2017) demonstrates that the
predicted by the model, agrees well with the manual picks in low-frequency (<0.05 Hz) component of DAS signal recorded at
training, validation, and test data sets. The algorithm was used in a an offset monitor well provides critical information to constrain
case study of an unconventional reservoir. The results indicate that hydraulic fracture geometry. This is one of the few technologies
smaller cluster spacing design creates denser fractures. so far that can locate fracture hits at an offset monitor well with
a spatial resolution less than 5 m. The data can be used to estimate
Introduction hydraulic fracture density, which is critical to determine cluster
Hydraulic fracturing completion design has a big impact on spacing for plug-and-perf designs.
the cost and production of unconventional wells. Plug-and-perf- The fracture hits in the low-frequency DAS (LFDAS) signal
type completion has become popular in recent years due to its are interpreted as fiber locations that experience strong extension
high efficiency in creating more evenly distributed hydraulic (fracture opening) during the pumping period and strong but
fractures in the reservoir. When designing a plug-and-perf comple- gradually decreasing compression after the pump stops. In contrast,
tion, many parameters must be determined, including cluster the fiber sections that are adjacent to fracture hits should experience
spacing, number of clusters per stage, number of perforations per compression (stress shadow) during the pumping period and
cluster, injection volume, proppant concentration, fluid type, usage extension (relaxation) after pumping has stopped. The identification
of diverter, etc. All of these parameters have an impact on induced of fracture hits in Jin and Roy (2017) relies on human interpretation
hydraulic fracture geometry, which is directly linked to production and manual picking, which is subjective and time consuming. To
performance. On the other hand, hydraulic fracture geometry is generate an objective estimation of fracture-hit locations automati-
also highly affected by reservoir rock property and local tectonic cally, we adopted the artificial neural network (ANN) algorithm
stress. The completion parameters must be optimized regionally to solve this problem.
to achieve the best economic results. The ANN concept was developed originally by McCulloch and
Fiber-optic sensing (FOS) technology has been used recently Pitts (1943). It became popular in recent years due to rapidly increasing
for the completion and production surveillance of horizontal uncon- computing power. The ANN algorithm loosely mimics the structure
ventional wells. It provides critical spatial information that can of a biological brain by forming a connected network of artificial
assist completion optimization. FOS technology utilizes laser energy neurons. The individual artificial neuron is a combination of input
scattered by impurities within fiber to measure temperature (dis- weights and an excitation function. Using the back-propagation
tributed temperature sensing) and strain rate (distributed acoustic algorithm (Werbos, 1974), ANN can adjust the weights in each
sensing [DAS]) along the deployed fiber with high spatial resolution neuron, which allows the model to be trained to make predictions
(1–10 m). Many FOS applications are developed for unconventional based on the input data. To get a better prediction performance, we
wells. The data can be used to evaluate perforation injection alloca- extract several features from LFDAS signal as the ANN input. After
tion for number of clusters per stage optimization (e.g., Boone et al., training using the manually labeled data, the ANN outputs
2015; Haustveit et al., 2017), stage isolation for plug design improve- the probability of fracture hits occuring at each fiber location
ment (e.g., Ugueto et al., 2016), crosswell communication for well during a certain stage. The probability can then be used to
spacing and completion design optimization (e.g., Holley et al., analyze the effect of cluster spacing on fracture density.
2015; Jin and Roy, 2017), production logging to estimate completion We use an unconventional case study to demonstrate the
efficiency and reservoir quality (e.g., In ’t Panhuis et al., 2014; Swan workflow implementation. In this data set, there are two designs
et al., 2018), etc. of cluster spacing; one is half of the other (double cluster number

ConocoPhillips, Houston, Texas, USA. E-mail: ge.jin@conocophillips.com; baishali.roy@conocophillips.com; darryl.g.buswell@conocophillips.com.


1

University of Utah, Department of Geology and Geophysics, Salt Lake City, Utah, USA. E-mail: kevin.mendoza@utah.edu.
2

520 THE LEADING EDGE July 2019 Special Section: Machine learning applications
with same wellbore length). By applying the machine learning the pumping starts and two hours after the pumping stops for each
workflow, we quantitatively estimate the impact of cluster spacing stage. There are 76 stages from five injection-monitor well pairs
on hydraulic fracture density. being recorded. The fracture hits in the data are manually picked
to serve as a labeled data set. Because the spatial resolution of the
Data and method gauge data is five channels (meters), we label a five-channel window
The DAS data were recorded at a pad with four instrumented to each manually picked fracture hit as positive samples.
wells in an unconventional reservoir. A gun barrel view of the The 2D (channel and time) LFDAS data are transformed into
Downloaded 06/22/20 to 130.238.7.40. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

four wells is shown in Figure 1. The fracking sequence of the pad eight 1D features in the channel axis. The features are designed
is W1, W2, W3, and W4. W1 and W4 are completed with a to characterize the fracture-hit signal from different perspectives.
cluster spacing of L. W2 and W3 are completed with a cluster A summary of feature symbols and descriptions can be found in
spacing of 2L. Because the fracture-hit signal can only be observed Table 1. εd is the summation of strain rate during the pumping
in a monitor well that has not been fracked (Jin and Roy, 2017), period. For this feature, channels with fracture hits should show
only five injection-monitor well pairs can be used for this study strong positive values, and channels in the stress shadow should
(orange arrows in Figure 1). The fiber cables are strapped outside show negative values. In contrast, εp is the strain rate summation
casing and cemented in place to ensure mechanical coupling with two hours after pumping stops, where fracture-hit channels should
the formation. The DAS data are recorded using 5 m gauge length show strong negative values and the surroundings should show
with 1 m channel spacing output, and the raw data are sampled positive. Feature A calculates the root mean square (rms) of
at 10 kHz. The data are first downsampled to 1 s after a low-pass individual channel amplitude during the pumping period to
antialiasing filter and then 2D median filtered with a five-channel capture large amplitude variation at the fracture hits. Feature D
and 10 s window. Data are further low-pass filtered with a corner sums the absolute value of the difference between a channel with
frequency of 0.05 Hz to get LFDAS data (Jin and Roy, 2017). the channels that are three channels apart (1 m channel spacing
LFDAS data are then cut into sections by completion stages. and 5 m gauge length). Because the geometry of fracture hits is
The time window for data splitting is selected to be one hour before usually smaller than DAS spatial resolution, the signal polarity
at fracture-hit channels should be different than channels half a
gauge length away. To better separate individual fractures and
enhance weaker fracture-hit signals, we applied a high-pass filter
in channel with a corner wavelength of 30 channels (100 ft) on

Table 1. Description of features.

Feature symbol Description


εd Accumulate strain during pumping
εp Accumulate strain post pumping
A rms of LFDAS data during pumping
D Difference between current channel and
nearby channels
Figure 1. Gun barrel view of the well pad. Black double arrows indicate horizontal FH High-pass filter with a corner wavelength
and vertical well spacing, and orange arrows point from the injection wells to the
monitor wells.
of 30 channels (30 m)

Figure 2. LFDAS signal and features. The color map shows extension (red) and contraction (blue) of the fiber. Horizontal dashed lines indicate the locations of manual fracture-
hit picks. Thick vertical dashed lines are the beginning and ending of the pumping period. εd , FH (εd ), A, and FH (A) are calculated using data between the two dashed lines
(0.5–3.2 hours). εp and FH (εp ) are calculated from the data after the second dashed line (3.2 hours). D and FH (D) are calculated using the entire data frame.

Special Section: Machine learning applications July 2019 THE LEADING EDGE 521
these features to create four new features: FH(εd), FH(εp), FH(A), all the stages in W1-W2 and W3-W4 well pairs as a test data set
and FH(D), respectively. with the hope of getting a more objective fracture density com-
The features are then normalized based on individual injection- parison between the two well pairs with same in-zone well spacing
monitor well pairs, as different well spacing may generate different but different cluster spacing designs. Within the three remaining
strain strength. Normalization is performed by standardization, well pairs (W2-W4, W1-W4, W1-W3), 45 stages are randomly
which makes each feature zero mean and unit variance. The selected as training data and 10 stages as validation data. For each
Downloaded 06/22/20 to 130.238.7.40. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

normalization process can be presented as: stage in the training and validation data set, we only select channels
within the range of 10 channels below the toe-most fracture-hit
f −f channel to avoid interference from fractures generated by previous
f′ = , stages, and 100 channels above the heel-most one to exclude
σ
low-amplitude negative samples. The channel selection created
where
f − f is the mean of feature f, and σ is the standard deviation 6426 samples in the training set within which 794 samples are
f′ =
of theσfeature. An example of the normalized features with com- channels labeled as fracture hits. The positive samples are dupli-
parison to the original LFDAS data can be seen in Figure 2. cated four times to increase their percentage in the training data
To properly train the machine learning model, the data need set to achieve better training performance.
to be split into three data sets: training, validation, and test. The Because the designed features can characterize fracture-hit
training data set is used to train the machine learning model, the channels from the rest, we found that a simple three-layer neural
validation data set is used to optimize hyperparameters of the network model performs well in this case. The neural network
model, and the test data set is used to evaluate model performance. model has eight nodes in the input layer, 10 nodes in the hidden
In this study, we did not follow the conventional 60%/20%/20% layer, and one node in the output layer. The nodes in the model
split of training, validation, and test data sets. Instead, we hold are fully connected. We use the rectified linear unit function as
the activation function. The architecture of the neural network
is shown in Figure 3. The model takes in the values of the eight
features of a certain channel and outputs the fracture-hit probabil-
ity at the channel during a completion stage. We use F1 score to
evaluate model prediction. The final model scores 85% on the
training data, 86% on the validation data, and 79% on the test
data based on a 0.5 threshold.

Result and discussion


Figure 4 shows an example of model prediction for the same
injection stage at two different monitor wells. One of the LFDAS
data is from the training data set, and the other is from the test
data set. The data plotted in the left panel are the same as shown
in Figure 2. No obvious performance difference can be observed
between these two well pairs. Virtual examination shows a similar
prediction performance in all data sets.
To more objectively investigate the effect of cluster spacing
on fracture density, we reserved the well pair W1-W2 and W3-W4
Figure 3. The architecture of a selected neural network model. as the test data set, which was not involved in the model training

Figure 4. Model prediction on the same stimulation stage at two different monitor wells. The data of monitor well 1 are in the training data set, and that of monitor well 2
are in the test data set.

522 THE LEADING EDGE July 2019 Special Section: Machine learning applications
spacing. First, fracture-hit probability
prediction from different stages are
summed together, and a low-pass filter
with a corner wavelength of five chan-
nels is applied to smooth the curves.
Then, local maximums of the curves
Downloaded 06/22/20 to 130.238.7.40. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

within the section marked in Figure 5


and with values larger than 0.5 are
identified as fracture-hit locations. The
CDF of the spacing between fracture
hits is calculated and plotted in
Figure 6. The fact that the blue curve
is always above the orange curve indi-
cates that the tighter cluster spacing
design (W1-W2) creates a denser
Figure 5. Probability of hydraulic fracture hits at monitor wells for W1-W2 and W3-W4 well pairs. (a) W1 was fracture network than the wider one
completed with cluster spacing L. (b) W3 was completed with cluster spacing 2L. Different colors indicate fracture (W3-W4). In a low-permeability shale
hits from different stages. Dashed box shows the depth range used for spectrum analysis. reservoir, this observation may support
tighter cluster spacing designs.

Conclusion
We adopted a machine learning algorithm to automate
fracture-hit identification on LFDAS data. Eight features are
developed to extract fracture-hit characters from the data and
serve as the model input. A simple neural network model was
trained to provide fracture-hit probability estimation from the
input. We applied the workflow in an unconventional case study.
The model prediction results and fracture-hit CDF analysis
indicate that smaller cluster spacing completion design creates
denser hydraulic fractures in the reservoir.

Acknowledgment
Figure 6. CDFs of fracture-hit spacing for W1-W2 (cluster spacing L) and W3-W4 The authors would like to thank ConocoPhillips ESOI and
(cluster spacing 2L). business unit management for permission to publish this paper.

process. The two injection-monitor well pairs have the same in- Data and materials availability
zone distances (660 ft, as shown in Figure 1), but cluster spacing Data associated with this research are confidential and cannot
of W1 is only half of the cluster spacing of W3. The model predic- be released.
tion results, colored by stages, are shown in Figure 5. It is obvious
that W1 creates many more fracture hits than W3. However, Corresponding author: ge.jin@conocophillips.com
there are several design changes in terms of number of clusters
per stage along W3 that may impact fracture density. To better References
compare the two, we select the sections where W1 and W3 have Boone, K., R. Crickmore, Z. Werdeg, C. Laing, and M. Molenaar,
the same number of clusters per stage but different cluster spacings. 2015, Monitoring hydraulic fracturing operations using fiber-optic
The sections are marked by the dashed boxes in Figure 5. Within distributed acoustic sensing: Proceedings of the Unconventional
the sections, there are 419 channels at each monitor well. A total Resources Technology Conference, https://doi.org/10.15530/
of 276 channels of W2 are identified as fracture hits from W1 by urtec-2015-2158449.
Haustveit, K., K. Dahlgren, H. Greenwood, T. Peryam, B. Kennedy,
the model, while only 216 channels of W4 are identified as fracture
and M. Dawson, 2017, New age fracture mapping diagnostic
hits from W3. By doubling the perforation cluster density, fracture
tools — A STACK case study: Presented at Hydraulic Fracturing
density increases by 28% at the in-zone monitor wells 660 ft away. Technology Conference and Exhibition, SPE, https://doi.
On the other hand, fracking order of the pad may impact org/10.2118/184862-MS.
fracking performance. W3 was completed after W1 and W2. Holley, E. H., and N. Kalia, 2015, Fiber-optic monitoring: Stimulation
The geometry of the fracture network W3 created may be affected results from unconventional reservoirs: Proceedings of the
by the fracture system created by W1 and W2. To answer the Unconventional Resources Technology Conference, https://doi.
key question — whether the L ft cluster spacing design creates org/10.15530/urtec-2015-2151906.
a denser fracture network than 2L ft — we calculated the In ’t Panhuis, P., H. den Boer, J. Van Der Horst, R. Paleja, D. Randell,
cumulative distribution function (CDF) for the fracture-hit D. Joinson, P. B. McIvor, K. Green, and R. Bartlett, 2014, Flow

Special Section: Machine learning applications July 2019 THE LEADING EDGE 523
monitoring and production profiling using DAS: Presented
at Annual Technical Conference and Exhibition, SPE, https://
doi.org/10.2118/170917-MS.
Jin, G., and B. Roy, 2017, Hydraulic-fracture geometry charac-
terization using low-frequency DAS signal: The Leading Edge,
36, no. 12, 975–980, https://doi.org/10.1190/tle36120975.1.
McCulloch, W. S., and W. Pitts, 1943, A logical calculus of
Downloaded 06/22/20 to 130.238.7.40. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

the ideas immanent in nervous activity: The Bulletin of


Mathematical Biophysics, 5, no. 4, 115–133, https://doi.
org/10.1007/BF02478259.
Sebastian, H., I. Gil, M. Graff, J. Wampler, G. Merletti, T. Sun,
S. Cadwallader et al., 2015, 3-D hydraulic fracturing and
reservoir flow modeling — Key to the successful implementa-
tion of a geo-engineered completion optimization program in
the Eagle Ford Shale: Proceedings of the Unconventional
Resources Technology Conference, https://doi.org/10.15530/
urtec-2015-2149711.
Swan, H. W., G. Jin, K. R. Krueger, and B. Roy, 2018, Production
logs from distributed acoustic sensors: U.S. Patent 10095828B2.
Ugueto C., A. Gustavo, P. T. Huckabee, M. M. Molenaar, B.
Wyker, and K. Somanchi, 2016, Perforation cluster efficiency
of cemented plug and perf limited entry completions: Insights
from fiber optics diagnostics: Presented at Hydraulic
Fracturing Technology Conference and Exhibition, SPE,
https://doi.org/10.2118/179124-MS.
Werbos, P., 1974, Beyond regression: New tools for prediction
and analysis in the behavioral sciences: PhD thesis, Harvard
University.

524 THE LEADING EDGE July 2019 Special Section: Machine learning applications

You might also like