BestPractice_IntegratingHorizontalWellDataInPropertyModelling_6729346_01_6729346_01

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Integrating Horizontal Wells in Property Modeling

Using declustering solutions in Petrel to account for sampling bias

Industry challenge
Horizontal and highly deviated wells preferentially drilled along the high quality part of the reservoir are common in conventional and
unconventional plays nowadays. The sampled subsurface data from the long sections of horizontal wells are generally clustered both in a vertical
sense (as they target stratigraphic levels) as well as spatially (as the wells target reservoir facies or fracture corridors). This means that the data is
usually biased and therefore most likely will misrepresent the global spatial distributions of the facies and petrophysical properties over the entire
volume of interest. Nonetheless, in order to build reliable reservoir simulation models, geostatistical methods require de-biased input spatial
distribution models (facies proportions and petrophysical distributions) that are representative of the entire volume of interest.

Although several methodologies are currently available in Petrel property modelling to deal with bias caused by spatial trends in the petrophysical
properties (e.g. trend modeling and removal in Data analysis, collocated co-kriging with secondary data, facies conditioning, etc.), there are no
simple direct methods and not even a clear consensus on how to deal with the sampling bias caused by horizontal well data. A common practice in
property modeling is to avoid the sampling bias in the input sample statistics by excluding the data from the horizontal well sections in the data
analysis and model fitting steps. However, horizontal well data do provide useful information, e.g. reservoir lateral heterogeneities, permeability
barriers and compartments, so rather than ignoring large portions of expensive and useful information, they should be an integral part of the input
dataset. The fundamental challenge faced by geomodellers is therefore: how can we optimally incorporate into the property modeling workflows
the information obtained from horizontal well data without introducing sampling bias in the geostatistical model?

The purpose of this article is to describe a general workflow to efficiently integrate horizontal well data as conditioning input in Facies and
Petrophysical modelling without introducing any sampling bias in the geostatistical model. The main focus will be on the use of the declustering
techniques available in Data analysis in order to estimate debiased facies and petrophysical property distribution that are more representative of
the whole volume of interest and not limited to high quality areas of the reservoir. Once the declustered histograms and global summary statistics
have been computed, it is then straightforward to use them as input in any of the simulation algorithms available in property modeling in the Petrel*
E&P platform.

What is declustering?
In a nutshell, the declustering methods compute and assign a weight to each input datum such that data in densely sampled areas receive less
weight (shrinking the histogram bars for the preferentially sampled data) and data in sparsely sampled areas receive larger weights (expanding the
histogram bars for the sparse data). This allows alleviating the sampling bias in the histograms and summary statistics of facies and petrophysical
properties caused by horizontal data preferentially sampled in the high quality part of the reservoir.

Two declustering algorithms are available in the Petrel Data analysis tool:

 Cell declustering: a commonly used technique which partitions the geocellular grid into blocks of cells, and then computes declustering
weights inversely proportional to the number of upscaled cells falling in each partitioning block.

Page 1 of 10 Error! No text of specified style in document.


The cell declustering weights will depend on the declustering geometry defined by the size and shape of the partitioning blocks, which are
given by the number of grid cells forming the blocks in the I, J, K directions.
Fig 1. Illustration of 2D cell declustering weights estimate algorithm.

 Kernel declustering: based on a statistical technique known as kernel density estimation. In a nutshell, a data density function is estimated on
the geocellular grid by applying a convolution operation of a smoothing kernel function on the upscaled data. This boils down to placing a
kernel function with given smoothing parameters or bandwidths on each of the upscaled data cells and then adding all the kernel functions
together. This results in a kernel density estimate at all the geocellular grid cells. Declustering weights are then defined inversely proportional
to the local data density.
Here, the declustering geometry is defined by the smoothing parameters (bandwidth) in the I, J, K directions, giving by the number of cells in
each direction. The declustering geometry controls the amount and orientation of induced smoothing, which ultimately affect the computation
of the kernel declustering weights.

Fig 2. Illustration of 2D kernel declustering weights estimate algorithm.


In outline, the declustering methodology follows these steps:

1. Compute and assign the declustering weighs to the input upscaled data using either of the available declustering algorithms (cell based or
kernel density based).
2. Compute the weighted data statistics (e.g. the mean of a continuous property such as porosity or the global proportion of a facies in a reservoir
model).

Page 2 of 10 Error! No text of specified style in document.


3. Optionally, if the data are known to be clustered preferentially in either high- or low-quality areas, optimize the declustered data statistics by
repeating steps 1 and 2 for a different declustering geometries (partitions of different sizes and shapes if cell declustering is used or bandwidth
parameters of different sizes if kernel declustering us used), and find the optimal declustering geometry for which the desired statistic is either
minimized or maximized.

Fig 3. Illustration of cell declustering geometry optimization step.

Use and limitations of declustering techniques


Declustering methods should not be applied blindly as the declustering weights might produce erroneous histograms and summary statistics if
some required conditions are not satisfied.

First of all, before any sampling bias correction is applied to the data, it is important to assess if there is in effect sampling bias in the statistics and
property distribution caused by preferential sampling of the input well data. If both facies and petrophysical properties are to be modeled, the facies
and petrophysical log data should be, as usual, sequentially analyzed.

First, we should assess if there is any bias in the facies proportions caused by preferential sampling of the input data:
 Ensure that there is a good understanding of the sedimentological/depositional model of the reservoir. This should provide a quantitative
assessment of the shape, dimension and direction of continuity of the various facies over the area of interest.
 Check if the horizontals wells are in fact preferentially sampling a particular facies/rock type in the zone of interest. This will indicate that
statistics derived from the well data might be biased.
 Check if the target facies proportions estimated from the raw log data are overestimated or underestimated due to preferential sampling.
Compare to any known/assumed information regarding the depositional environment.

Page 3 of 10 Error! No text of specified style in document.


Fig 4. Assessing existence of bias in the facies proportions due to preferential sampling of the input data. Check facies thickness/proportions maps together with the wells tracks and
histogram of facies log data.

After assessing the existence or not of sampling bias in the input facies data, we should analyze if there is sampling bias in the petrophysical
properties distribution on a by-facies basis. Here we are interested in assessing whether or not the petrophysical property distribution estimated
from the input data is representative of the underlying facies. Even after conditioning by facies, sampling bias might still be observed in the
petrophysical input data due to i) spatial variability of the petrophysical property within the facies and ii) denser horizontal sampling on the low- or
high- valued areas of a particular reservoir facies.
 Ensure that there is a good understanding of the heterogeneity of the petrophysical properties within the volume of interest (a particular facies
of a particular zone). This should provide a quantitative assessment of the dimension and direction of continuity of the petrophysical property
within the facies/zone of interest.
 Check if the horizontals wells are in effect preferentially sampled in the low- or high- value areas within a particular facies of a particular zone.
This will indicate that statistics derived from the well data might be sampling biased.
 Check if the property distribution and summary statistics estimated from the raw log data are overestimated or underestimated due to
preferential sampling. Compare to any known/assumed prior information regarding the petrophysical property distribution within the facies.

Fig 5. Assessing existence of bias in the petrophysical property distribution due to preferential sampling of the input data.

Page 4 of 10 Error! No text of specified style in document.


Once we have established the existence of sampling bias in the input data, before applying declustering techniques we need to check that the
following conditions are satisfied for the declustering to be effective:

1. There are enough data covering the entire range of the target property distribution. Since the weights only adjust the influence of each
datum, declustering methods only work when there are enough data to assign weights in both the good and poor areas of the volume of
interest.

2. There is no evidence of geological trends causing spatial bias in the data. If a geological trend causing spatial bias can be observed in
the input data, declustering methods will not be able to correct the lack of representativeness in the property distribution. In this case,
trend modelling and debiasing methods must be applied first.

Fig 6. Illustration of scenarios where declustering methods might not be effective.

It is important to point out that, even if the two conditions above are satisfied, the declustering approach should be considered as a heuristic
estimation technique. That is, it cannot be guaranteed that the sampling bias caused by preferential sampling will be fully corrected.

If there is not strong evidence of sampling bias caused by preferential sampling in the input data, or not enough data are available covering the
entire range of the target property distribution, then it is better to keep it simple and use the naïve non-weighted histograms and statistics as a
proxy for the true but unknown property distribution.

Estimating and using declustered facies proportions


Once the facies log data have been properly upscaled into the geocellular grid model, we can use the declustering options available in the facies
Data analysis tool to alleviate any bias in the facies distribution caused by preferential sampling.
 Select the upscaled facies property and the Zone of interest in the 'Data analysis' tool.
The global facies proportions are computed by default using the facies segment lengths (along the well measured depth) as weights, as
opposed to cell counts only. When horizontal wells are included in the input data, the length weighted proportions can be significantly biased
towards the reservoir facies.
 Toggle off the length weighting button in the Proportion tab to compute the cell-count based proportions.
 In the Declustering tab, select the declustering settings and compute the optimal declustering weights from the input upscaled data:

Page 5 of 10 Error! No text of specified style in document.


1. Select the target facies whose proportion is known to be biased (overestimated or underestimated) by preferential sampling. The
estimated proportions and histograms for the input upscaled data and declustering weighted data are displayed in the dialog.
2. Select either the 2D or the 3D declustering option. 3D declustering is recommended when highly deviated and horizontal well data is
present in the dataset.
3. Select the declustering algorithm from the dropdown menu. The Cell declustering algorithm is recommended for horizontal wells drilled
in a particular stratigraphic layer. The Kernel declustering is preferable for horizontal wells drilled through different stratigraphic layers.
4. Select either Minimize proportion or Maximize proportion optimization criterion depending if the proportion is known to be overestimated
or underestimated for the selected facies.
5. If we want to use available information regarding the directions of spatial continuity of the input data, (e.g. the modelled major, minor and
vertical variogram ranges), as prior constraint for the optimal declustering geometry, manually enter I, J, K values with aspect ratio that
correspond as the aspect ratio of the major, minor and vertical variogram ranges, and tick the Lock aspect ratio on.
6. Click on the ‘Estimate’ button to compute the optimal declustering geometry in the I, J and K directions, which translates into the
declustering cells size for the cell declustering algorithm or bandwidths for the kernel declustering algorithm.

The declustering weights are then automatically computed such that, when assigned to the input upscaled data, will produce the minimum or
maximum declustered facies proportion

Fig 7. Declustering tab in Data analysis tool for facies properties. Only available for upscaled and raw log data.

Note that the optimization step finds the declustering geometry for which the resulting declustering weights will produce the minimum/maximum
proportion for a single selected target facies that is known to be oversampled or under sampled. This means that the algorithm should perform well
for minimizing/maximizing the proportion of a single oversampled/undersampled target facies, provided there are ‘enough’ samples for the other
facies too.
However, when trying to maximize the proportion of an under sampled facies (e.g. shale) in order to simultaneously minimize the proportions of
many oversampled facies, the algorithm might or might not be able to produce a good result depending on the dataset. The effectiveness of the
declustering method is always limited by the amount of data and extent of the information about the facies proportions they provide.
 Once the optimal declustering weights have been computed, we need to toggle on the ‘Use declustering weights’ button in the Proportion tab
for the facies histograms and proportions to be re-computed using the declustering weights. The new declustered facies proportion should
now be corrected for the sampling bias caused by the horizontal well data.
 The next step is to complete any necessary facies data analysis. That is, fit any meaningful vertical trend model observed on the declustered
histogram and perform variogram analysis and modelling.

Page 6 of 10 Error! No text of specified style in document.


The general recommendation for variogram modeling is not to use highly deviated or horizontal wells for the computation of the experimental
variogram in the vertical direction. Non-vertical or deviated wells can be filtered out of the calculation of the vertical variogram by constraining
the angle tolerance not to exceed 20-30 degrees.
 Finally, apply to store the estimated facies proportions, variogram and trend models for subsequent use in Facies modelling.

Once the debiased facies proportions and trend model have been computed in Data analysis, it is then straightforward to use them as constraints
in any of the cell-based or object-based facies simulation algorithms available in property modeling in the Petrel* E&P platform.
 In the 'Facies modeling ' tool, select the upscaled Facies property and zone of interest.
 Select the desired stochastic Facies modeling method (Object modeling, Truncated Gaussian, Indicator simulation) and the facies to be
modelled. The selected facies must be consistent with the facies previously selected in data analysis.
 Toggle the ‘Use estimated facies proportions from data analysis’ button on. The estimated global facies proportions computed in data
analysis from the upscaled input data will be then fetched for all selected facies.
– If the ‘Length weighted’ button was toggled off in Data analysis, then the cell-count based global fractions computed in Data
analysis will be fetched.
– If the ‘Use declustering weights’ button was toggled on in Data analysis, then the ‘declustered’ global proportions computed in
Data analysis will be fetched.
 If necessary, use the vertical proportion curves and/or variogram model from data analysis.

Fig 8. Facies modeling tool. Proportion, trend and variogram models computed in Data analysis can be automatically used.

Page 7 of 10 Error! No text of specified style in document.


Estimating and using declustered petrophysical properties distribution
After estimating and using declustered facies proportions and assessing the need to also apply declustering techniques on the petrophysical input
data, we need to re-compute the optimal weights for the petrophysical data as the optimal declustering parameters most probably will not be the
same as the ones obtained for the facies data.

Once the petrophysical log data have been upscaled into the geocellular grid model, we can then use the declustering options available in the
Petrophysical modeling Data analysis tool to alleviate any observed sampling bias in the property distribution.
 Select the upscaled petrophysical property of interest, the Zone of interest and the conditioning Facies (if necessary) in the 'Data analysis'
tool. In the Transformations tab, you can display the data distribution as histograms, identify trends within the data, and apply different data
transformations such as Normal score.
 Identify and model any observed trends. Since declustering techniques are ineffective in the presence of spatial trends in the sample data,
any observed trends should be removed first.
 Select Normal score transformations to transform the distribution of the (residual) data into a standard normal distribution.
 Tick on the ‘Decluster’ option to display the declustering settings and compute the optimal declustering weights. Similar to the facies case:
1. Select either the 2D or the 3D declustering option. 3D declustering is recommended when highly deviated and horizontal well data is
present in the dataset.
2. Select the declustering algorithm from the dropdown menu. The Cell declustering algorithm is recommended for horizontal wells drilled
in a particular stratigraphic layer. The Kernel declustering is preferable for horizontal wells drilled through different stratigraphic layers.
3. Select either Minimum mean or Maximum mean optimization criterion depending if the input sample mean is known to be overestimated
(data preferentially sampled on high values) or underestimated (data preferentially sampled on low values).
4. Information regarding the directions of spatial continuity of the input data, (e.g. the modelled major, minor and vertical variogram ranges),
can be used as prior constraints for the optimal declustering geometry by manually entering I, J, K values with aspect ratio that
correspond as the aspect ratio of the major, minor and vertical variogram ranges, and ticking the Lock aspect ratio on.
5. Click on the ‘Estimate’ button to compute the optimal declustering geometry in the I, J and K directions, which translates into the
declustering cells size for the cell declustering algorithm or bandwidths for the kernel declustering algorithm.

The declustering weights are then automatically computed such that, when assigned to the input upscaled data, will produce the minimum or
maximum declustered mean value.

Fig 9. Transformation tab in Data analysis tool for petrophysical properties. Declustering option is available for the Normal score transform.

Page 8 of 10 Error! No text of specified style in document.


 Finally, perform variogram analysis and model fitting if necessary and ‘Apply’ to store the estimated statistics, variograms and trends for
subsequent use in Petrophysical modelling.

Once detrending and normal score transform has been performed on the declustered data in Data analysis, it is then straightforward to use them
as constraints in any of Petrophysical properties simulation algorithms available in property modeling in the Petrel* E&P platform.
 In the ' Petrophysical modeling ' tool, select the upscaled petrophysical property of interest, zone of interest and conditioning facies (if any).
 Select the desired stochastic Gaussian modeling method: Gaussian random functions simulation, Sequential Gaussian simulation.
 Click ‘Use transformation from data analysis’ button. The transformed property distribution computed in data analysis from the upscaled input
data will be then fetched in the Distribution tab.
 ‘Use variogram from data analysis’ if necessary.

Fig 10. Petrophysical modeling tool. Variogram and property distribution models computed in Data analysis can be automatically used.

Quality checking the results


As stated before, it cannot be guaranteed that the declustering techniques will fully correct the bias caused by preferential sampling. We should
always QC the results by comparing the declustered solutions to any prior known/assumed information regarding the facies and petrophysical
properties distribution. Both qualitative and quantitative quality control procedures must be performed.
 Qualitative QC. Simulate and compare the facies and petrophysical models with and without applying declustering. Which one is more
consistent with the known/assumed model/geology of the reservoir?
– Show 3D displays of the models: intersections and stratigraphic sections, filtered by facies/zones/segments/I,J,K index/values.
– Generate and display QA attributes maps (for each facies/zone).

Page 9 of 10 Error! No text of specified style in document.


Fig 11. Quality assurance maps for facies and petrophysical models simulated with and without declustered statistics.

 Quantitative QCs: Compare the histograms, facies proportions, and summary statistics of the simulated facies petrophysical models with and
without using declustering. Which one is more consistent with the known/assumed information regarding the facies and petrophysical
properties distribution (within the facies/zone of interest)?

Fig 12. Histograms of facies models simulated with and without declustered facies proportions estimates.

Fig 13. Histograms of petrophysical property simulated with and without declustered distributions.

Page 10 of 10 Error! No text of specified style in document.

You might also like