Sei sulla pagina 1di 17

URTeC: 383

Reservoir Fluid Data Acquisition Using Advanced Mud Logging Gas


in Shale Reservoirs
Tao Yang, Ibnu Hafidz Arief*, Martin Niemann, Marianne Houbiers; Equinor ASA

Copyright 2019, Unconventional Resources Technology Conference (URTeC) DOI 10.15530/urtec-2019-383

This paper was prepared for presentation at the Unconventional Resources Technology Conference held in Denver, Colorado, USA,
22-24 July 2019.

The URTeC Technical Program Committee accepted this presentation on the basis of information contained in an abstract
submitted by the author(s). The contents of this paper have not been reviewed by URTeC and URTeC does not warrant the
accuracy, reliability, or timeliness of any information herein. All information is the responsibility of, and, is subject to corrections by
the author(s). Any person or entity that relies on any information obtained from this paper does so at their own risk. The information
herein does not necessarily reflect any position of URTeC. Any reproduction, distribution, or storage of any part of this paper by
anyone other than the author without the written consent of URTeC is prohibited.

1. Abstract

Oil production from shale reservoirs has increased dramatically in the recent years. To identify drilling
targets and optimize well completions, it is important to get early access to reservoir fluid properties.
However, due to the low permeability of shale reservoirs, fluid samples often become available only after
most important development decisions have been made. Therefore, it has been an abiding challenge in the
industry how to acquire fluid properties data earlier in shale reservoirs.

Mud logging gas data acquired while drilling provide the earliest hydrocarbon response from the reservoir.
In an earlier study, we have demonstrated that advanced mud gas data have large potential to predict
reservoir fluid properties. In general, fluid properties are strongly correlated with thermal maturity of the
source rock. In shale reservoirs, reservoir fluids are still in the source rock, as low permeability limits
migration and convection of the reservoir fluids. As a result, the reservoir fluid systems in shale reservoirs
are relatively undisturbed and have a high degree of consistency. This provides the possibility to correlate
advanced mud logging gas data and reservoir fluid properties.

Based on a reservoir fluid database with more than 60 samples from different shale reservoirs, we developed
a machine learning algorithm to predict fluid properties from advanced mud logging gas data. The accuracy
of the new method is significantly improved compared with the previous model which used an explicit
correlation based on wetness. In addition, the new approach is more general and does not depend on a
specific shale reservoir. We applied the new model to 11 wells with advanced mud logging gas data. The
predicted gas oil ratios are close to the measurement from early production data when advanced mud
logging gas data are of good quality.

This publication demonstrates that advanced mud logging gas data can be used to acquire reservoir fluid
properties in shale reservoirs. Such approach provides a novel and cost-efficient solution for the sampling
challenges in early phase. In addition, the method provides continuous fluid data along entire well, as
opposed to a single fluid sample taken at a specific location. Hence the results provide insight in the fluid
URTeC 383 2

distribution in shale reservoirs. The method can be widely used for sweet spot identification and optimizing
fracking strategy in shale reservoirs.

2. Introduction

Due to the extremely low permeability, the commonly used wireline fluid sampling technology to acquire
PVT data cannot be applied in shale reservoirs. For shale producers, a reservoir fluid sampling program is
usually performed after production starts, with the disadvantage that the PVT data cannot be used to
optimize decisions that must be made before production starts. For the oil industry it has been a longstanding
problem how to get earlier access to fluid data in shale reservoirs.

Pressure coring (Bjorum et al., 2013) is a method to take in-situ representative fluid samples in shale
reservoirs during drilling. After the pressurized core sample is taken to laboratory, the released gas and
extracted oil from the core will be recombined to reproduce the reservoir fluid sample. However, the method
is only reported in academic research, not commercially available due to the high cost, and impractical
because of lagging time in the laboratory.

Standard mud gas logging has been a safety tool while drilling for a long time. Besides the measurement of
total gas, the mud gas logging tools also provide a chromatographic breakdown of the gas for the range C1-
C5. However, due to the physical aspects of the degassing process, compositional fractionation takes place
resulting in a gas composition that is artificially enriched in the lighter gaseous hydrocarbons (Niemann et
al., 2007). Two decades ago, advanced mud gas logging technology became available (Kandel et al., 2001;
Ellis et al., 1999, 2003; Berkman et al., 2002). Due to a controlled and thermodynamically stable degassing
process at elevated temperature, corrections for hydrocarbon recycling and extraction efficiency are
possible. The corrected gas composition from advanced mud gas (AMG) is comparable with the reservoir
fluid composition (normalized to C1-C5) obtained from PVT analysis. Fully corrected AMG data (currently
limited to C1-C5) have been widely used to provide a qualitive interpretation of reservoir fluids encountered
while drilling.

Although mud gas is the first detection of hydrocarbons in the reservoir while drilling, it is not commonly
used to acquire reservoir fluid data because the data only provides partial information of gas components
(C1-C5), but no information of oil components (e.g. C7+). Therefore, the quantitative prediction of reservoir
fluid properties from mud gas data has remained distant from a practical application. For the past two
decades, there are limited efforts reported in the literature to quantitatively predict reservoir fluid properties
(e.g. gas oil ratio) from mud gas data (Wright, 1996; Sima et al., 2014; and Melo et al., 2016). However,
these efforts did not show acceptable prediction accuracy for practical applications. Recently we published
a machine learning approach to predict GOR based on AMG data with focus on conventional reservoirs
(Yang et al., 2019). The fluid property prediction accuracy based on AMG data is significantly improved
compared with all previous work. The results confirm the method is ready for large industrial applications
in conventional reservoirs.

For shale reservoirs, the gas oil ratio (GOR) can be correlated with the wetness of reservoir fluid sample
compositions (C1-C5), as shown in our previous work (Yang et al., 2014; Yang, 2017-2018). This is also
supported by later studies (Akihisa et al., 2017; Whitson et al., 2018). We applied the wetness-based
correlation to AMG data from a pilot well and the predicted GORs were close to the early production GOR
from the well. The pilot test showed the potential that AMG data can be used to provide continuous and
real time GOR prediction. However, to prove the technology readiness for large applications in shale
reservoirs, it is necessary to verify the method on multiple applications with reliable prediction results.
URTeC 383 3

Another drawback of the correlation based on wetness is that it is only sufficiently accurate for a specific
shale play, and each shale play has its own empirical correlation. It will be highly desirable to develop an
algorithm which is more generally applicable than the shale-dependent correlation based on wetness.

As shown in PVT correlations for shale reservoirs (Yang et al., 2014; Yang, 2017-2018), fluid properties
are strongly related with thermal maturity due to the limited migration and convection of reservoir fluids.
The consistency of reservoir fluids in shale reservoirs provides a better possibility to correlate AMG data
and fluid properties than in conventional reservoirs. In this paper, we developed a machine learning
algorithm to predict GOR from AMG data from 11 wells in shale reservoirs. Although we only present
GOR prediction to demonstrate the methodology, other fluid properties (e.g. density, saturation pressure,
formation volume factor) can be derived in a similar workflow.

3. Methodology

The general workflow for predicting GOR from AMG data is similar as what we documented in the paper
for conventional reservoirs (Yang et al., 2019). The basic steps in the workflow are:
1) Prepare and quality control PVT data and AMG data.
2) Select and train the machine learning model to predict GOR using a large reservoir fluid database
with only PVT samples from shale reservoirs.
3) Investigate compositional similarity between AMG data and analog PVT samples.
4) Apply the machine learning model trained on PVT samples to AMG data for GOR prediction.
5) Compare the predicted GORs from AMG data with GOR measurements from early production to
reach conclusions.

The detailed documentation of these steps and in particular of the machine learning method will not be
repeated in this paper. However, application to shale reservoirs has special considerations which are
different from conventional reservoirs. These differences are addressed below.

3.1 Data preparation and quality control (QC)

In this subsection, we describe the data preparation for this study, including PVT data, AMG data and
production data.

PVT data

The reservoir fluid data database is an internal database with PVT data from more than 60 PVT samples.
Most of the PVT samples are from the Eagle Ford shale play, while other PVT samples are a different shale
asset. The PVT samples include a wide spread of fluid types, from lean gas, to rich gas, gas condensate,
volatile oil and black oil. All samples were taken at an early production stage. The separator gas and oil
samples were recombined to reproduce the reservoir fluid samples. Standard PVT measurements were
performed at commercial laboratories.

We used only single-flash GOR measurements since these were reported for all samples and have the same
separation reference conditions. Outliers were removed through visual inspection and a manual verification
process. Although the PVT data contain the full compositional (including C7+) analysis for each sample, we
URTeC 383 4

only used the C1-C5 composition and renormalized them. The resulting composition of PVT samples is in
the same format as from AMG composition.

Advanced mud gas data

AMG data from 11 wells are available for this study. AMG data are recorded while drilling by analyzing
the released gas composition of returned drilling mud and cuttings. The composition of AMG data is usually
from C1 to C5 (C6+ components can be measured using specific equipment, but these components are
normally only corrected for hydrocarbon recycling, not for extraction efficiency which we need here). Even
after the recycling and extraction efficiency corrections, it is necessary to QC the AMG data before any
applications. Understanding the quality of the AMG data is essential before we perform the GOR prediction
because the data quality has a significant impact on the prediction accuracy. We employed the same quality
control (QC) methods as were used in the study for conventional reservoirs (Yang et al., 2019). Among
others, a QC metric is used to flag low-quality or unreliable AMG data values. Examples of characteristics
in the AMG data that are flagged are:
1) Large fluctuations of a component within a small depth range.
2) The first observations after missing measurements.
3) C1 content below a given threshold.
4) C4 or C5 content below a given threshold.

To quantify the quality of the AMG data, the QC metric is defined in a range from 0 to 1. High-quality
AMG data would have QC metric value close to 1. If one or more of the above factors are found, then the
QC metric is penalized. Low-quality AMG data is indicated with QC metric close to 0. A single numeric
quality measure between 0 and 1 can be plotted together with a predicted GOR log to visualize the
confidence level associated with each prediction, based on AMG data quality.

Production data

Unlike the study for the conventional reservoirs, it is not possible to take the corresponding PVT sample
versus AMG data at a specific depth. None of the PVT samples in the PVT database is associated with a
well with AMG data. Therefore, to make a judgement about correctness of the predicted GORs from AMG
data, we use the early production data from the well to determine a representative, ‘true’ GOR value for
each of the 11 wells.

3.2 Predictive models for GOR

All predictive models discussed in this paper, are based on the reservoir fluid database consisting of the
PVT samples from shale reservoirs. Compared with AMG data and production GORs, the C1-C5
composition of PVT samples has better quality, and the GOR data of PVT samples are measured in the
laboratory conditions with good control. Using the high-quality PVT data in the modeling work leads to
more robust results compared to when using the noisier AMG data and production GOR data directly.
URTeC 383 5

Empirical correlation based on wetness

In our previous study (Yang et al., 2014; Yang, 2017-2018), we used wetness to derive the relation between
C1-C5 composition and GOR of the corresponding sample. The wetness is defined in Equation 1. Depending
on the applications, wetness can be defined for separator gas composition or for reservoir fluid composition.
For application to AMG data, we defined the wetness using reservoir fluid composition. As shown in Figure
1, the correlation is strong between GOR and wetness for Eagle Ford samples (in black dots). A best fit
model for the Eagle Ford samples is given in Equation 2.

Figure 1 Correlation between GOR and wetness for different shale plays

∑5𝑖𝑖=2 𝐶𝐶𝑖𝑖
𝑊𝑊ℎ = 100% × (Eq. 1)
∑5𝑖𝑖=1 𝐶𝐶𝑖𝑖
𝐺𝐺𝐺𝐺𝐺𝐺 = 5.615 × 10−0.0773.𝑊𝑊ℎ +5.0618 (Eq. 2)

The advantage of an empirical correlation is that it is easy to apply and does not require a sizeable database.
The disadvantages of the correlation based on wetness are the limited accuracy and the shale-dependence.
As shown in Figure 1, the correlation derived for the Eagle Ford shale play does not work well for a different
shale play (in red triangles). The shale-dependent correlation limits the general applicability of the
approach, especially for a new shale play with limited data available.

Machine learning approach

Our objectives of the study are 1) to increase the model prediction accuracy compared to the wetness-based
correlation and 2) to establish a generic model which is not shale dependent as the empirical correlation
based on wetness. When a sizable database is available, machine learning is a good alternative to establish
the relation between C1-C5 composition and GOR of PVT samples. The compositional input parameters
can vary significantly and therefore will generate a large amount of input options. These input parameters
include individual component concentrations (C1-C5), wetness, balance, character, and numerous
component ratios (e.g. C1/C2, C1/C3). Machine learning offers a robust and efficient way for screening
through a large variety of so-called modelling approaches.
URTeC 383 6

A modeling approach is the combination of a selected predictive model, the features (input variables) made
available to it, the feature selection or reduction method applied to the feature set, possible transformation
of the target variable (e.g. the log of GOR), and possible transformation of the input variables (e.g. log of
C1 concentration). The predictive models considered in this study are Gaussian Process, Universal Kriging,
KMeans, Random Forest, and linear regression regularized with Elastic Net. A modeling strategy is a
collection of such modeling approaches considered for selection.

For the model building and evaluation on PVT data, we conducted two types of analysis:
1) Generalization assessment workflow, in which we analyzed how well a model trained on PVT input
from certain wells, would perform on a different set of wells.
2) Model selection stability, in which we analyzed how robust the data-based selection of the best
modeling approach is.

The details of the generalization assessment workflow and model selection stability are not repeated in this
paper; more details can be found in the paper (Yang et al., 2019) or other standard machine learning
publications. To evaluate the model accuracy of a modelling approach, we use the Mean Absolute
Percentage Error (MAPE). The MAPE definition for GOR can be found from Equation 3.

𝑁𝑁𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
100% �𝐺𝐺𝐺𝐺𝑅𝑅𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚,𝑖𝑖 − 𝐺𝐺𝐺𝐺𝑅𝑅𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐,𝑖𝑖 � (Eq. 3)
𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 (%) = × �
𝑁𝑁𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 𝐺𝐺𝐺𝐺𝑅𝑅𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚,𝑖𝑖
𝑖𝑖=1

3.3 Compositional similarity between AMG data and analog PVT samples

Since the wetness correlation was derived from PVT data, and similarly, the machine learning algorithm
was trained and developed based on PVT data, we need to make sure that C1-C5 compositions between
AMG data and the PVT samples are similar. A comparison of C1-C5 compositions between AMG data and
PVT samples is an important step before we apply the predictive model based on PVT samples to AMG
data.

Because we cannot take wireline fluid samples in shale reservoirs, it is not possible to compare C1-C5
compositions from AMG data with corresponding PVT samples at the same depth. Therefore, we chose to
compare the C1-C5 composition from representative AMG data with analog PVT samples. Radar plots are
used to investigate the compositional similarity. The traditional radar plot can demonstrate well the
compositional similarity between the representative AMG data with the analog PVT samples.

3.4 Apply the machine learning model to AMG data for GOR prediction

Once the modeling strategy is accomplished on PVT data using the generalization assessment and model
selection stability workflows, the resulting predictive model needs to be evaluated with AMG data as input.
For this study, we selected the best modeling approaches using the simple feature set of the normalized C1-
C5 composition, instead of automatically selecting from all the modelling approaches described based on
PVT data. The reason we used the selected modeling approach is that it is more efficient to run and provides
prediction results reflecting the different model performance.
URTeC 383 7

Since we don’t have PVT samples corresponding to AMG data at a specific depth, we chose to use early
production GOR data as the true values. GOR prediction based on AMG data will be compared to the true
values from early production data.

4. Results and Discussions

We show the modeling results in this section, including radar chart analysis of PVT samples, machine
learning model results based on PVT samples, comparison of C1-C5 composition between AMG and PVT
data, and finally the GOR prediction results based on AMG compositional inputs.

4.1 Radar chart analysis of PVT samples

When we include PVT samples from multiple shale plays, it is important to verify that the composition of
C1-C5 for all PVT samples have a similar pattern before we develop a global GOR prediction algorithm for
all shale plays. We chose to use Radar chart analysis for this objective. As shown in Figure 2, the polygon
shapes for all samples from different shale plays demonstrate a generally similar shape. The size of the
polygon is corresponding to the GOR value of the PVT sample. A small size polygon (inner cycle) indicates
a low GOR oil sample and the large size polygon (outer cycle) indicates a high GOR gas sample.

4.2 Machine learning model results based on PVT samples

We screened different machine learning models for GOR prediction. Universal Kriging, Gaussian Process
and Random Forest models show the best GOR prediction accuracy, when using C1-C5 composition from
PVT samples. The data used for the machine learning model development includes PVT samples from all
shale plays. The MAPE of GOR prediction based on the machine learning model (Kriging) is around 17%,
while the measurement error of GOR in the PVT data is about 10%. As shown in Figure 3, the predicted
and measured GOR values (black dots) are closely spread around the reference line. Compared with the
empirical correlation based on wetness (red triangles, MAPE is around 34%), the accuracy is significantly
improved with the current machine learning algorithms including C1-C5 composition as input.

In addition, the prediction results based on multiple input parameters (C1-C5 composition) show the GOR
prediction is not shale dependent as the empirical correlation based on wetness. As show in Figure 4, the
machine learning algorithm (Kriging) based on only Eagle Ford PVT samples made good GOR predictions
(MAPE is 29%) for a different shale play. When the empirical correlation from Eagle Ford was used to
predict the GOR for a different shale play, very poor results were found (MAPE is 82%).

The machine learning algorithm with multiple input parameters successfully overcame the weakness of the
empirical correlation based on wetness. The algorithm does not only have improved accuracy but have the
important capability to apply to reservoir fluid system in a new shale play which is not included in the
model development.

4.3 Comparison of C1-C5 composition between AMG data and PVT data
Because it is not possible to take wireline fluid samples in shale reservoirs, we don’t have the opportunity
to compare the C1-C5 composition between AMG data and PVT sample at a specific depth like in
conventional reservoirs. However, we still need to perform an evaluation to prove we can apply a model
based on PVT data to AMG data. Fortunately, the samples from shale reservoirs show strong consistency.
URTeC 383 8

Figure 2 Radar chart of all reservoir fluid samples from different shale reservoirs

Figure 3 Comparison of the prediction performance between the machine learning model (Kriging) and
the empirical correlation based on wetness

Figure 4 GOR prediction of the reservoir fluids from a different shale play using machine
learning model (Kriging) and the empirical correlation based on wetness
URTeC 383 9

Instead of comparing with corresponding PVT sample at the same depth in the same well, we compare C1-
C5 composition from AMG data in each well with PVT analog data from other wells.

For AMG data, we selected C1-C5 composition at four different depths (measured depth). For PVT data,
we selected analog PVT samples from the fluid database. As we discussed before, all PVT samples show
similar a polygon shape in a Radar chart, but with different polygon sizes which are correlated to GOR
values. If the AMG data are consistent with analog PVT samples, the polygon shape of AMG data should
be similar to other PVT samples. We have AMG data from 11 wells, 7 wells show good agreement between
AMG data and analog PVT samples and 4 wells with poor agreements. Figure 5 to Figure 7 show data
comparisons for 3 wells with good agreement at the four depth levels (except well No.2 at depth 12 900 ft).
Figure 8 shows the data comparison from one of these wells with poor agreement. The four wells (No.1 to
No. 4) are selected to show Radar charts for compositional similarity. The same wells are also chosen for
demonstrating production GOR data and the final predicted GOR logs.

The reason for inconsistent polygon shape in Radar chart between AMG data and PVT samples can be due
to correction process of C1-C5 compositions from raw data to calibrated data reflecting in-situ reservoir
fluid composition. All AMG data from 11 wells are obtained with the Schlumberger Flair system. The
choice of Extraction Efficiency Calibration (EEC) parameters plays an important role to correct the raw C1-
C5 compositional data to be similar as reservoir fluid samples (Guerriero, et al., 2012). The machine learning
algorithm in this paper is highly sensitive to EEC corrections, especially the calibrations on C4 and C5.
When large inconsistency is found between the AMG data and PVT samples, the developed algorithm shall
not be applied for GOR prediction due to the quality issue with AMG data.

4.4 Production GOR data

We collected historic production data (including GOR) from the 11 wells. All wells show that the GOR
values derived from production data fluctuate over time. Late production GORs are not recommended to
use due to many other factors introduced. We define an early production GOR by taking the average of the
early production data for a relative stable GOR period for most wells.

Figure 9 shows the GOR derived from historic production data for the same 4 wells discussed earlier. The
GORs in these 4 wells are representative for the spread in the 11 available AMG wells. It is clear from these
figures that even during early production, it is difficult to pinpoint a representative GOR value. It is
important to realize that the ‘true’ GOR value we specified for each of the 11 wells has some uncertainty,
which is larger than the measurement error of GOR in PVT data, due to the nature of the production GOR
data in shale wells.

4.5 GOR prediction results based on AMG compositional inputs

Finally, the GOR predictions are computed for the 11 wells with AMG data, using machine learning models
derived from PVT data. The results for the 4 wells discussed earlier are shown in Figure 10 and Figure 11
(two wells in each figure). In these figures, the reported C1-C5 compositions of AMG data are shown on the
left side in the reservoir fluid logs. The predicted GOR logs are shown in the middle. The predicted GOR
log with Kriging model is labeled with red color, while the predicted GOR log with Random Forest model
is labeled in blue color. The objective of showing prediction results from two best models is to show the
model performance difference although the overall prediction accuracy of the two models are very similar.
The black dashed line is the GOR value based on early-production data; this ‘true’ GOR value is by
definition constant along the entire well. The QC metric value is plotted in a log side by side with the GOR
URTeC 383 10

prediction on the right side. If the QC metric>0.3, the quality column is filled with green color representing
good quality AMG data. If QC metric<0.3, the quality column is filled with red color representing poor
quality AMG data.

Similar as the comparison of C1-C5 compositions from AMG data and PVT data, we have 7 wells with good
GOR prediction results. The predicted GOR values are in large agreement with the early production GOR
values. The poor GOR predictions for four wells are mainly due to the poor compositional similarity
between AMG data and analog PVT samples or/and low QC metric values indicating low quality AMG
data. Although the GOR prediction will not be performed when the comparison of compositional similarity
between AMG data and analog PVT samples shows poor agreement, we performed GOR prediction for all
11 wells for comparison purpose.

Figure 5 Radar chart comparison between PVT data and mud gas data (black line) at four different depths
(Well No.1)
URTeC 383 11

Figure 6 Radar chart comparison between PVT data and mud gas data (black line) at four different depths
(Well No.2)

Figure 7 Radar chart comparison between PVT data and mud gas data (black line) at four different depths
(Well No.3)
URTeC 383 12

Figure 8 Radar chart comparison between PVT data and mud gas data (black line) at four different depths
(Well No.4)

Figure 9 Production data from 4 wells and the defined well GORs at early production
URTeC 383 13

Table 1 Summary of GOR prediction results for all 11 wells


Wells AMG data GOR Comments on AMG data quality
quality prediction
results
1 Good Good
2 Good Good
3 Good Good
4 Poor Poor Radar chart inconsistency/Low QC Metric
5 Good Good
6 Good Good
7 Good Good
8 Good Good
9 Poor Poor Radar chart inconsistency
10 Poor Poor Radar chart inconsistency
11 Poor Poor Radar chart inconsistency

Table 1 shows the summary of GOR prediction results for all 11 wells with AMG data. Although the well
pool size is not optimal for statistics, around 70% of the wells have AMG data passing all QC procedures.
The prediction results based on good quality AMG data are close to the early production GOR data. The
AMG logs, the GOR prediction results, the constant production GOR, and the QC metric for the four wells
discussed earlier are shown in Figure 10 to Figure 11. In general, wells No.1, No. 2 and No.3 demonstrate
good prediction results. A deviation is observed in well No. 2 at the interval 12 800 ft – 13 900 ft. In that
interval, the predicted GOR is high and it is caused by the disagreement between the AMG data and the
PVT analog samples (see Figure 6). The GOR prediction in well No. 4 is consistently higher than the
production GOR. The large discrepancies between predicted and production GOR is expected in this well
due to the disagreement between the AMG data and the analog PVT samples (see Figure 8).

For all 11 wells, we found that most wells have rather stable predicted GORs along the long horizontal
well. This observation might be because all the wells are expected to stay in the same formation. It is very
difficult to define any gradual shift of GORs. However, we noticed that GORs have step-wise fluctuations
for some wells. To make a sensible evaluation on the step-wise fluctuation along the well, it is important to
make an integrated analysis with all other logs about reservoir quality including drilling events. Before
making any conclusion on the fluid heterogeneity along the well, lots of work needs to be done before we
understand the impact on the production. This is beyond the scope of the current paper. The work in this
paper provides an additional source of information, that could be useful for further studies on this subject.
URTeC 383 14

5. Conclusions

We developed a machine learning algorithm to predict continuous in-situ reservoir fluid properties using
advanced mud gas data. The algorithm has significantly improved accuracy of prediction compared with
the empirical correlation based on wetness. In addition, the model based on C1-C5 composition input
avoided the shale dependence we noticed for the wetness-based correlation. Therefore, the methodology is
more general than the previous work and can be applied to shale reservoirs with limited reservoir fluid data
available.

The results from multiple applications to 11 wells shows good prediction results when AMG data has
similar C1-C5 compositions compared with analog PVT samples and QC metric has a reading higher than
0.3. For this study, we achieved good results for 7 wells, for which AMG data are considered as high quality.
The results validate the methodology for further business applications.

The method based on AMG data provides a practical way for reservoir fluid data acquisition while drilling.
The approach provides continuous reservoir fluid data which is especially useful to understand the fluid
distribution for long wells and potential multiple layers in shale reservoirs. Such early access to the reservoir
fluid data provides great advantages for making decisions like identifying sweet spot and optimizing
fracking strategy.

6. Acknowledgement

The authors thank Equinor ASA management for permission to publish this paper. The support from
Equinor’s Digital Subsurface, especially Tina Todnem, Mark Thompson, Thibaut Forest, Mathias Alerini
are greatly appreciated. We wish to acknowledge the managerial and technical support from Equinor’s
Research and Technology (R&T) division, especially from Camilla Vavik Pedersen, Svein Tollefsen, Knut
Uleberg and Peter Eilsø Nielsen. Special thanks are extended to Andre Martins, Laura Froelich, Dharam
Kapila, and Jukka Ylitalo from Teradata.

7. References

Akihisa, K., Knapp, L., Uchida, S., Shimokawara, M., Akita, Y., Wood, J. M., … Sanei, H. (2017,
November 1). Sweet Spot Survey Based on Mud Gas and Cuttings Analyses - Montney Tight Gas
Reservoir. Society of Petrophysicists and Well-Log Analysts.

Berkman, T., Ellis, L., Grass, D., Abstract: Integration of mud gas isotope data with field appraisal at Horn
Mountain Field, Deepwater Gulf of Mexico, 86 (13) AAPG Bulletin, 2002.

Bjorum, M., Willberg, D. M., Davis, M., Akbarzadeh, K., & Williams, R. (2013, November 5). Novel
Conrolled Pressure Coring and Laboratory Methodologies Enable Quantitative Determination of Resource-
in-Place and PVT Behavior of the Duvernay Shale. Society of Petroleum Engineers. doi:10.2118/167199-
MS
URTeC 383 15

Ellis, L., A. Brown, M. Schoell, and M. Haught, “Mudgas isotope logging while drilling: A new field
technique for exploration and production”, 19th International Meeting on Org. Geochem., 6-10 September
1999, Istanbul, Turkey, Abstracts Part I, p. 67-68.

Ellis, L., A. Brown, M. Schoell, and S. Uchytil, “Mud gas isotope logging (MGIL) assists in oil and gas
drilling operations” Oil &Gas Journal, v. 101, Issue 21, 2003.

Guerriero, N., Lessi, J. (2012). Method of calibration for the use in a process of determining the content of
a plurality of compounds contained in a drilling fluid. European Patent, EP 2 703 597 A1.

Kandel, D., Quagliaroli, R., Segalini, G., and Barraud, B., “Improved Integrated Reservoir Interpretation
Using Gas While Drilling Data”, SPE 75307, 2001.

Melo, B., Ferroni, G., Pereira, M. (2016, Febuary). Formation fluid prediction through gas while drilling
Analysis. Técnico Lisboa.

Niemann M., Breviere J., Lessi J., Jaulneau P. & Shell FEAST Team (2007): Quantitative Hydrocarbon
Extraction from Drilling Mud Allows Accurate Evaluation of Formation Fluid Composition; IMOG,
Torquay, UK.

Sima Liqiang, Wu Feng, Ma Jianhai, Fang Guoqing, Yu Hang. (2014 December). Quantitative calculation
of GOR of complex oil-gas-water systems with logging data: A case study of the Yingdong Oil/Gas Field
in the Qaidam Basin. Natural Gas Industry B, Volume 1, Issue 2, Pages 172-177.

Whitson, C. H., Alqahtani, F. M., & Chuparova, E. (2018, August 9). Fluid Heterogeneity on a Well-Box
Scale in Tight Unconventional Reservoirs. Unconventional Resources Technology Conference.
doi:10.15530/URTEC-2018-2882502

Wright, A. C. (1996, January 1). Estimation of Gas/Oil Ratios And Detection Of Unusual Formation Fluids
From Mud Logging Gas Data. Society of Petrophysicists and Well-Log Analysts.

Yang, T., Areif, I., Nieman, M., Houbiers, M., Meisingset, K. K., Martins, A., and Froelich, L.: “A Machine
Learning Approach to Predict Gas and Oil Ratio Based on Advanced Mud Gas Data”, SPE Europec featured
at 81st EAGE Annual Conference & Exhibition, 3-6 June, London, UK, 2019.

Yang, T., Basquet, R., Callejon, A., Van Roosmalen, J. J., & Bartusiak, B. (2014, August 25). Shale PVT
Estimation Based on Readily Available Field Data. Unconventional Resources Technology Conference.
doi:10.15530/URTEC-2014-1884129

Yang, T., 2017-2018. Estimation of shale gas and oil properties based on field data. SPE Distinguished
Lecturer Presentation.
URTeC 383 16

Figure 10 Predicted GOR log based on AMG data, early production GOR and QC metric log
for well No. 1 (left) and well No. 2 (right)
URTeC 383 17

Figure 11 Predicted GOR log based on AMG data, early production GOR and QC metric log
for well No. 3 (left) and well No. 4 (right)

Potrebbero piacerti anche