Académique Documents
Professionnel Documents
Culture Documents
characterization case study presented in this paper Wolf. Core analyses are also available for two other cored
incorporates petroleum engineering, surface seismic, wells, the Behm #1 and #2 Federal Rabbit wells, for which
crosswell seismic, geologic, and neural network data. This the actual cores are not available for study. The dry hole in
study was conducted by the Montana University System SW1/4, SW1/4 Section 8 was omitted from analysis because
Petroleum Reservoir Characterization research team which of severe deviation in the hole. The Brown 7-34-20 Federal
was formed in November, 1993 with funding from the US well in SE1/4, SW1/4, SW1/4 Section 7 is included because
Department of Energy (DOE) and from the Montana Science there are production data for this well; however no logs are
and Technology Alliance (MSTA). available.
As a part of the reservoir characterization study, 3-D
A Reservoir Characterization Case Study: The NE surface seismic and crosswell seismic surveys were
Rabbit Hills Oil Field conducted. The 3-D surface seismic survey included the
The Rabbit Hills oil field lies low on the eastern flank of the south quarter of Section 7 and the north three-quarters of
Sweetgrass arch, north of the Bearspaw Mountains in north- Section 18 (Fig. 2). The crosswell seismic survey was
central Montana, USA. It is located in Blaine County, about conducted between the Meridian 42-18 Flynn and the Behm
eight miles north of the town of Chinook, Montana, in #2 Federal Rabbit over a horizontal distance of approximately
sections 7 and 18, T34N, R20E (Figs. 1 and 2). 950 ft (290 m). The Federal Rabbit #2 was the designated
injection well in the pilot polymer flood project that began in
Field Characteristics. The Rabbit Hills reservoir is of mixed March, 1996.
carbonate-siliciclastic composition and occurs in the upper A neural network approach was then used to correlate
part of the Bowes Member, upper member of the Middle well log porosities and permeabilities with seismic amplitude
Jurassic Sawtooth (Shaunavon) Formation. Gross reservoir data from the 3-D seismic survey. Input data for the network
interval thickness of the upper part of the Bowes varies from were derived from the five wells for which core analyses were
9 to 26 ft (2.7 to 7.9 m); pay thickness is strongly facies available as well as from porosity logs. The neural network
controlled within the interval and may vary from 2 to 20 ft porosity and permeability predictions were then utilized as
(0.6 to 6.1 m). The average drilling depth in the field is input data for the reservoir simulator.
4,090 ft (1,246 m) and the average reservoir subsea elevation The principal parts of the Rabbit Hills reservoir
is minus 1401 ft (427 m). To date, the field has produced characterization study being reported on in the present paper
slightly over 2 million barrels of 19-degree API-gravity oil. are the design, execution, and results of the reservoir
There is no associated gas but there is a strong water leg. simulation modeling. This simulation work has been ongoing
Seismic and subsurface data indicate a structural high in the over several years as parameters of the model were refined
field area and the field is considered a structurally enhanced based on the addition of geologic, seismic, and production
stratigraphic trap. data.
Field History. The field was discovered in 1972 with the Geology
Amoco #1 USA Erving Wolf well which had an IP of 366 The upper Bowes is interpreted as a valley fill deposited on a
BOPD. The Amoco well went on production in 1973 and was lowstand unconformity that was incised into the lower Bowes1
the only well in the field until 1984; except for 11 months of (Fig. 3). The location and orientation of this incised valley
production from the Norfolk Flynn 7-1 well during 1975- may have been in part structurally controlled, as suggested by
1976 before this well ceased production due to high water the mappable structural element (now a structural high)
recovery. In 1984, the Brown 7-34-20 Federal well went on beneath and on trend with the thickest valley fill (Figs. 4 and
production and 21 wells now comprise the field. Of these, 5). Sediments were deposited in shifting channels across an
currently 11 are producing, 3 are shut in, 1 is an injection aggradational coastal braid plain in response to the flooding
well, and 7 are dry holes (Fig. 2). Most of the producing of the Rierdon sea. It is unlikely that individual channels can
wells were drilled in 1991 and 1992 resulting in a sharp be distinguished at any level of evaluation. The
increase in oil production rate and cumulative oil. A pilot predominance of ooid and bioclastic grains indicates the
polymer flood was begun in March, 1996. sediment was supplied predominantly from a marine
environment. The quartz sand probably sourced both from
Study Area and Research Data. Twenty-one wells comprise the marine environment and from land. Sediment was
the NE Rabbit Hills field and characterization study area carried into a landward-shifting valley estuary by marine
(Fig. 2). Five cores are available; four from producing wells, waters that mixed with sediment moving down the valley.
and one from the dry hole in SW1/4 SW1/4 Section 17. Abundant sediment and rising base level caused the valley
Routine core analyses of porosity and permeability conducted floor drainage system to braid rapidly back and forth across
by service companies are available for four of these cores; the valley, distributing and redistributing the aggradational
analyses are not available for the Amoco #1 USA Erving bed load.
SPE 38985 NEURAL NETWORK PARAMETER PREDICTION IN RESERVOIR CHARACTERIZATION AND SIMULATION 3
have a slight effect on the testing error and the above scale estimation during the survey and sonic logs were constructed
factors resulted in the lowest error. The network input data for the test wells using a sonic log from a nearby well and
thus consisted of scaled combinations of seismic amplitudes, gamma logs from the test wells. It was economically not
instantaneous seismic amplitudes, and grid coordinates. possible to perform deviation surveys in the test wells, so the
For the neural network to predict the porosity of the upper crosswell data were used to estimate the extent and direction
Bowes Member, the seismic data representing the Bowes had of well deviations.
to be extracted from the full seismic data set. Extracting the The bandwidth of the recorded crosswell data ranged from
data was accomplished by training a separate neural network approximately 250 to 1200 Hz. These frequencies are
to predict the time of the regional unconformity between the significantly higher than the frequencies in the 3-D data
Bowes Member and the underlying Firemoon Member. This volume (up to approximately 100 Hz). This high frequency
regional unconformity was chosen because it is a salient content makes it possible to image lithologic variations at an
feature in the seismic data that is located at the bottom of the order of magnitude greater resolution than possible for
Bowes reflection (Fig. 7). surface seismic measurements4.
The training data for this horizon-picking neural network Fig. 9 shows the resulting crosswell reflection image with
were created by hand picking the interface on three random well logs and the velocity tomogram overlay. The entire
sets of twenty adjacent CMPs. This trained network was then imaged interval covers 500 ft (152 m) in depth. Within this
used to estimate the time of the Bowes-Firemoon interface for interval, numerous reflection events occur illustrating the
the remainder of the data set. An algorithm which precisely capability of crosswell data to significantly increase
located the time of the peak representing the interface was resolution. The Bowes interval is evidenced by the two high
used to fine tune the neural network’s predictions. Thirteen amplitude reflection events. Although considerable lateral
seismic samples above the predicted Bowes-Firemoon variation is evident at the top of the events, interpretation
unconformity were selected for each CMP as input to the must be constrained by knowledge of the error introduced due
neural network. To train the neural network, seismic data to lack of a well deviation survey. A well deviation survey is
from the training well locations were input into the neural critical for accurate location of reflectors.
network in random order. Corrections were made for well The tomogram overlay is used as the velocity model for
drift at the producing depth using synthetic seismograms processing the reflection events. The logs shown on either
when possible. side show good agreement between the derived sonic
Because the Rierdon-Bowes interface is poorly defined velocities and tomographic velocities. The seismic velocity at
(Fig. 7), the time interval was limited to prevent inclusion of the level of the Bowes is approximately 12,500 ft/s (3810
data representing the Rierdon. If information from the m/s).
Rierdon were present in the input data, any porosity Efforts are ongoing to determine effective methods to
variations in the Rierdon would introduce additional error. integrate crosswell data into our integration work.
The Bowes Member is thin with respect to the seismic
wavelength so it was not possible to separate the lower and Neural Networks
upper Bowes seismic expression to isolate the upper Bowes. A neural network approach enables seismic data to be
However, since the lower Bowes is represented by a uniform correlated with porosity or permeability without the
depositional environment, lower Bowes porosity variations requirement of explicitly defining the relationship of
should be small relative to the expected variation in the upper parameters that affect the acoustic impedance, such as water
Bowes. content, lithology, or pore pressure. Additionally, neural
networks can incorporate an interval of seismic data rather
Crosswell Seismic than a single value for parameter prediction. Use of an
In addition to acquiring a 3-D surface seismic survey, a interval of seismic data increases the amount of available
crosswell seismic survey was acquired about two years later. information to estimate parameters. As might be expected,
The primary goal of the crosswell work was to detect lateral the relationship between seismic data and physical parameters
changes in formation properties in the Bowes using crosswell is not unique. Any number of combinations of values of
reflection imaging4. The thickness of the producing zone, porosity, pore pressure, lithology, water content, and other
less than 20 ft (6 m), places it at the resolution limit of the 3- parameters can give rise to the same acoustic impedance.
D surface seismic. This ambiguity means that control data is very important in
The wells used for the crosswell survey were the Flynn determining how valid solutions are.
Trust 42-18 and the Federal Rabbit #2. The surface The application of neural networks is becoming
separation of the wells was approximately 950 ft (290 m). increasingly common. A neural network approach is often
Data processing results were obtained without the benefit of useful for pattern recognition problems or where no simple
sonic logs or deviation surveys for either of the wells mathematical relationship exists for the parameters of
occupied. Gamma logs were used for depth control and dip interest. The type of neural network used is dependent upon
SPE 38985 NEURAL NETWORK PARAMETER PREDICTION IN RESERVOIR CHARACTERIZATION AND SIMULATION 5
the problem. Neural networks can be categorized into two porosity maps. The four maps with the lowest error were
main classes: supervised and unsupervised. For this study, a then averaged together. This ensemble approach helps
backpropagation feedforward network (supervised) was used alleviate the dependence of the network solution on the
for parameter prediction. network initial values.
Because there is not a well defined analytical relationship
Background. Artificial neural networks are fundamentally between seismic data and porosity or permeability, prediction
different from conventional programming approaches. ANN error cannot be easily determined. Therefore, the reliability
utilize sets of input and output training examples to allow the of the training process is estimated by inputting a seismic
network to “learn” relationships. Typically a supervised trace interval that corresponds to known porosity or
network consists of a layer of input nodes for training permeability and comparing the neural network prediction to
example inputs, one or more “middle-layers” of nodes which the known parameter value.
are fully connected to the input nodes, and a layer of output A “one-out” testing procedure was performed by leaving
nodes also fully connected to the middle layer(s). A one well out of the network training set and inputting values
connection consists of an adjustable weight value. In a from the remaining wells in random order. The test well’s
feedforward network, the inputs are multiplied by the weight seismic data were then input into the trained network and the
values of all of their connections. All of the values connected network predictions for that well were compared to the well
to a node in a following layer are then summed and modified log or core values.
by a transfer or threshold function whose purpose is to limit There were five wells with core analyses in the study area
the range of the summation. These node outputs are then and eight wells with porosity logs. The “one-out” approach
passed forward to the next layer using the same process. The described above was used to train porosity and permeability
output of the nodes in the output layer represent the ANN networks. Networks were trained using four of the five sets of
outputs which are compared with the true outputs in the core permeability values or seven of the eight sets of porosity
training phase. The error is used to adjust the weights log values; the remaining sets were used to test how well each
throughout the network and the process is repeated with network trained. This process was repeated once for each
multiple passes through the input values until the ANN meets well in either the porosity or permeability group. For the
suitable error criteria. At this point the network is “trained”. final prediction maps, all of the wells are included in the
The “testing” phase consists of presenting new examples training data set. Neural network parameters, network
to the trained network and comparing the ANN output with architecture, and input data for the neural network were then
the true value. This step determines if the network will be adjusted based on the testing error. The error criterion was
able to generalize to new data sets. If testing is successful, the absolute difference of the predicted and measured porosity
the network is ready to be applied to data sets for which no or permeability values.
true outputs exist. Because the difference in porosity between the upper and
Drawbacks inherent in the use of neural networks include lower Bowes is approximately seven percent, if the testing
the lack of established procedures for network construction, error is less than that value, then it might be possible to
the requirement of fixed numbers of network inputs and distinguish the high porosity bioclastic limestone from the
outputs, the necessity of constructing training sets, and the non-reservoir lithology.
lack of a convenient method for error propagation. The
essence of a neural network lies in its nodes and their Neural Network Predictions. Porosity maps were generated
associated connections strengths or weights3. The number of using core, neutron porosity (NPHI) and density porosity
nodes in a layer and the number of layers are determined (DPHI) data. Predictions for two layers were made by using
adhoc; except for the first and last layer whose dimensions averaged log porosity values over interpreted intervals
are determined by the input and target or training examples. corresponding to the upper and lower Bowes. Permeability
Initially, each connection weight between nodes is set to a maps were generated from the core analysis data using
small random number. These weights are then continually similar interpretations. The parameter values subsequently
updated via a learning rule as the network “learns” in the used for input into the simulator consisted of the sets of two-
training phase3. layer porosity and permeability prediction maps.
Neural network predictions are often dependent upon Single layer porosity predictions gave testing errors less
these initial random weights. This occurs because multilayer than 4.0% porosity for the eight NPHI wells and less than
networks contain multiple minima in parameter space. Thus, 4.3% porosity for the seven DPHI wells. Errors for the two
depending on the starting point, different minima may be layer porosity predictions were less than 7.2% porosity for
reached; not always the global minimum. The number of layer 1 and less than 7.4% porosity for layer 2 using NPHI
minima is related to the complexity of the problem. In an porosities.
attempt to make the porosity predictions more reliable, ten Figs. 11 a and b show NPHI porosity prediction errors for
random seed values were used to generate ten different layer 1 and layer 2 using seismic amplitude, instantaneous
6 T. AHMED, C.A. LINK, K.W. PORTER, C.J. WIDEMAN, P. HIMMER, J. BRAUN SPE 38985
amplitude, and grid coordinates for input data. The root- cells containing them. They were generated using the core
mean-square error for layer 2 is slightly less than the error for lab analysis values for the five available wells and the “one-
layer 1 and the two layer errors are significantly larger than out” procedure. Predicted permeability values for layer 1 are
the error for one layer predictions. considerably higher than the predicted values for layer 2.
Figs. 12 a and b show the regridded porosity prediction One possible explanation for this could be the variable effect
maps using seismic amplitude, instantaneous amplitude, and due to the valley fill that formed the upper Bowes. Again, the
grid coordinates as input data. (Note: figures are oriented outlier values were removed from the permeability predictions
with north pointing to the right.) Pixels in the figures before input into the reservoir simulator.
represent predictions made at each CMP and then averaged The zones of high permeability do not correlate with the
over the simulator cells containing them. Fractional porosity predicted zones of high porosity. This observation is at least
is represented by the scale bar accompanying the figure. an indication that the network is responding to different
Histograms were used to identify outlier predictions which patterns in the seismic data than did the network that was
were then removed before input into the simulator. The trained on porosity, and should be expected.
number of outliers varied from one to three. Reasonable Testing errors for the permeability predictions were quite
porosity values were predicted for all CMP locations (the large. These results represent our first attempts at
outlier maximum was approximately 32% porosity). That the permeability prediction and we are exploring variations to
network predicted reasonable porosity values for all CMPs increase the accuracy of our predictions.
indicates that the network is recognizing significant
information in the input data. Reservoir Simulation
Using thirteen seismic samples effectively removes any A three-phase, three-dimensional black oil simulator was
information pertaining to thickness variation of the Bowes used to construct a model of the Rabbit Hills Field. A 24 by
member. This lack of thickness information could be a factor 30 by 2 Cartesian grid system was selected to simulate the
in reservoir predictions where tuning effects occur. In depletion process of the field. The spacing of the simulator
addition, because the seismic data represent the entire Bowes grid is not constant and is shown in Fig. 10 along with well
Member, any variation in the lower Bowes will introduce locations and the area covered by the 3-D seismic survey.
errors into the predictions. The increase in error from one to Simulator grid spacing varies from 150 ft (46 m) to 1500 ft
two layers is likely due to the limit in vertical resolution of (457 m) in the x-direction (north-south) and 250 ft (76 m) to
the seismic data. Layer 1 represents the average porosity over 1500 ft (457 m) in the y-direction (east-west).
the top half of the upper Bowes and layer 2 over the bottom History matching of the field covered initial field
half of the upper Bowes. As can be seen in Fig. 12, these production in June 1973 through January 1996. Several
maps predict that the high porosity regions are not continuous history match runs were necessary to ensure that the
throughout the upper Bowes. Using only the coordinates of simulation model closely matched the observed field
the CMPs as input data, a single layer porosity prediction performance and to improve its reliability for prediction. The
gave very low testing errors of less that 2% porosity. model was calibrated by matching the following historical
Table 1 shows the absolute difference between the field production data:
predicted and the actual single layer porosity values for the (a) monthly oil and water production for eleven
eight NPHI wells used in this study. Each column header producing wells,
identifies the data type input into the neural network to make (b) water-oil ratio,
the porosity predictions. C represents coordinate inputs, A (c) cumulative oil and water production.
instantaneous amplitude inputs, and S seismic amplitude Because the reservoir pressure was never below the oil
inputs. Inputting only coordinate data gives the lowest bubble-point pressure, the gas-oil ratio was not considered in
testing errors. This result shows that geologic control is the history match.
needed to determine which results are valid. Based on the The following are some of the issues that needed to be
geologic model of the reservoir, a homogenous model is addressed during the model construction and history
incorrect. By including instantaneous amplitude data, testing matching phase:
errors increase but are still within acceptable limits. (1) The actual areal extent of the NE Rabbit Hills Field.
Including both seismic and instantaneous amplitude data does (2) Early fluid performance which indicated the
not cause a significant increase in the error and may aid the presence of significant quantities of water phase attributed to
network in recognizing patterns in the data. For this reason it either high initial connate water saturation or the presence of
was decided to include all data sets for the final predictions. a small aquifer.
Figs. 13 a and b show the two layer permeability map (3) The slow decline in the production rates of the two
predictions. As in Fig. 12, figures are oriented with north original discovery wells indicating that withdrawals are being
pointing to the right and pixels in the figures represent balanced by the natural energy of a “large reservoir” or by a
predictions made at each CMP averaged over the simulator limited water influx augmented by fluid and rock expansion.
SPE 38985 NEURAL NETWORK PARAMETER PREDICTION IN RESERVOIR CHARACTERIZATION AND SIMULATION 7
Earlier attempts had been made to model the performance and well data were used to calibrate the model.
of the field without a water influx. These simulator
production and pressure decline rates were considerably Acknowledgments
below the recorded data. As a result, a source of edge water This work has been supported by US Department of Energy
influx was attached to the northwest and southwest flank of (DOE/EPSCoR) and Montana Science and Technology
the Rabbit Hills Field. Alliance (MSTA) research funds under Contract No. DEFC-
During the history matching process, none of the 02-91ER75681 granted to the Montana University System for
initialization data that were generated from the geological research in petroleum reservoir characterization. The
model, 3-D seismic, or neural network predictions were industrial partner in the research, Burlington Resources
altered to achieve matching. Only relative permeability and (formerly Meridian Oil Company), agreed to make the NE
well data were used to calibrate the model. Rabbit Hills field available, and has provided company data
Fig. 14 is a comparison of the actual and predicted flow and materials to the research team. Other operators in the
rates for the discovery well, Amoco #1 USA Erving Wolf, field, including L.E. Behm, United Meridian (formerly
from 1973 through 1995. Figs. 15 and 16 compare depletion Norfolk Energy), Amoco Production Company, and J. Burns
performance of the Rabbit Hills Field in terms of total flow Brown, have made cores, core analyses, and other
rates and cumulative oil production as a function of time. As information available. Crosswell seismic acquisition and
can be seen from the plots, the history matching was very processing were contracted by TomoSeis, Inc., Houston, TX.
satisfactory. The authors are grateful for the manuscript reviews
provided by colleagues, and the preparation of text figures by
Conclusions Pat Tamarin at MBMG and Jon Conaway at Montana Tech.
The Rabbit Hills Field characterization study has been
successful at integrating data of varying type and scale into References
parameter estimates used as input to a reservoir simulator. 1. Porter, K.W.: “Geologic data in a reservoir characterization,
The scale of the data types used ranged from core scale Sawtooth Formation, NE Rabbit Hills field, Blaine County,
measurements up through interwell seismic measurements, Montana,” Montana Bureau of Mines and Geology Open-File
Report 354, (in press).
surface seismic and regional geology measurements.
2. Sheriff R.: Reservoir characterization investigations in
Neural network analysis was used to provide porosity and
Geophysics No. 7, Society of Exploration Geophysicists, Tulsa,
permeability estimates across the study area by training on OK (1992).
seismic amplitude information. Reasonable estimates were 3. Zurada, J.M.: Introduction to artificial neural systems, West
obtained that agreed with geologic and seismic trends. The Publishing Co., St. Paul, MN (1992).
neural network predictions were then used as input to a 4. Harris, Jerry M., Nolen-Hoeksema, Richard C., Langan, Robert
reservoir simulator. T., van Schaack, Mark, Lazaratos, Spyros K. and Rector, James
Simulator results showed very good agreement with W. III: “High-resolution crosswell imaging of a west Texas
production data. Throughout the history matching process, carbonate reservoir: Part 1. Project summary and
none of the initialization data that were generated from the interpretation,” Geophysics (1995), 60, no. 3, 667-681.
geological model, 3-D seismic, or neural network predictions
were altered to achieve matching. Only relative permeability
Well C A AC ASC
A 1.14 2.3 2.72 2.46
B 0.48 3.41 2.92 2.73
C 0.08 3.58 2.64 2.73
D 1.03 2.57 0.93 0.42
E 1.00 6.22 1.86 3.56
F 1.53 1.53 1.61 3.24
G 1.59 4.42 2.52 3.95
H 1.58 2.45 1.63 1.55
8 T. AHMED, C.A. LINK, K.W. PORTER, C.J. WIDEMAN, P. HIMMER, J. BRAUN SPE 38985
Fig. 1. Map showing location of NE Rabbit Hills oil field and other localities in north-central Montana.
Fig. 2. Map detail of NE Rabbit Hills oil field in sections 7 and 8, T34N, R20E, in northern Blaine County, Montana. Areas of the project and of 3-
D seismic survey shown. Twenty-one wells are included in the study; the well in SW 1/4 section 8 is excluded.
SPE 38985 NEURAL NETWORK PARAMETER PREDICTION IN RESERVOIR CHARACTERIZATION AND SIMULATION 9
Fig. 3. Three-dimensional diagram of coastal braid-plain valley-fill model for upper Bowes at NE Rabbit Hills field. North is toward upper right;
coastal braid plain is oriented generally east-northeast. Multiple channels are shown depositing aggradational valley fill across the braid plain.
Braid-plain valley fill overlies erosional unconformity incised into lower Bowes. On north side of area, a marine tongue is shown as observed in
the Norfolk 7-10 Flynn well (Fig. 6). An earlier lowstand unconformity of regional extent is shown incised into underlying Firemoon Member of
Sawtooth (Shaunavon) Formation.
Fig. 4. Conventionally prepared isopach map of upper Bowes interval containing the reservoir rock (bioclastic limestone) as well as other
channel-fill lithologies. Bold numbers are thickness values; lighter numbers are initial production (IP) values in barrels of oil per day (BOPD).
Compare with structure map for Bowes interval (Fig. 5). A similar map is produced by computer contouring of field data.
10 T. AHMED, C.A. LINK, K.W. PORTER, C.J. WIDEMAN, P. HIMMER, J. BRAUN SPE 38985
Fig. 5. Conventionally prepared structure contour map on top “Sawtooth marker” horizon in lowermost Rierdon Formation approximately fifteen
feet above top of Bowes Member (top of Sawtooth Formation). Calculations based on log depths and Kelly Bushing (KB) data, and expressed
relative to mean sea level. Compare with isopach map of upper Bowes (Fig. 4). A similar map is produced by computer contouring of field data.
The structure is also visible in 3-D seismic data across the field.
Fig. 6. Annotated gamma and resistivity log suite for Norfolk 7-10 Flynn well. Structure marker horizon is in lowermost part of the Rierdon
Formation. Oil production at NE Rabbit Hills field is from the upper Bowes interval. Annotated depth is ft below ground elevation.
SPE 38985 NEURAL NETWORK PARAMETER PREDICTION IN RESERVOIR CHARACTERIZATION AND SIMULATION 11
Fig. 7. Seismic in-line from the 3-D seismic data set. The line runs east-west and is approximately in the center of the survey. The arrow in the
figure indicates the varying nature of the top of the Bowes Member.
Fig. 8. Seismic amplitude map from the top of the Bowes Member. Zones of high seismic amplitude are darker gray at left center and top center
of the map. They correspond to high porosity zones in the layer 2 neural network porosity prediction map (Fig. 12 b).
12 T. AHMED, C.A. LINK, K.W. PORTER, C.J. WIDEMAN, P. HIMMER, J. BRAUN SPE 38985
Fig. 9. Plot showing the results of crosswell reflection imaging overlain on the velocity tomogram produced from the crosswell seismic data set
recorded between the wells Flynn Trust 42-18 and the Federal Rabbit #2. Well separation is approximately 950 ft (288 m). The location of the
reflection events is heavily influenced by any well deviation present.
SPE 38985 NEURAL NETWORK PARAMETER PREDICTION IN RESERVOIR CHARACTERIZATION AND SIMULATION 13
6.28
Flynn_trust_7-10
Flynn_trust_7-9
6.27
Flynn_trust_7-15
Federal_7-12
Flynn_trust_7-1 Austad_1
6.26
Federal_7-34-20Flynn_trust_7-14 Flynn_trust_7-16
6.25
northing
Fed._Rabbit_1
Flynn_trust_41-18
Federal_17-4
Fed._Wolf_13-12
USA_E._Wolf_1Flynn_MOI_21-18
6.24
Malsam-Fed._1 Flynn_trust_42-18
Federal_17-5
Fed._Rabbit_2
Fed._Wolf_22-18
6.23
Flynn_trust_18-2
6.22
6.21 Federal_1-18
6.2
2.076 2.077 2.078 2.079 2.08 2.081 2.082 2.083 2.084 2.085 2.086
easting x 10 6
Fig. 10. Map showing the locations of the wells in the Rabbit Hills field overlain on the reservoir simulator grid (irregularly spaced lines). Also
shown are the bin locations for the 3-D seismic survey. For the neural network analysis, a boundary of 7 CMPs was removed from the seismic
grid to eliminate the low-fold areas.
layer 1 layer 2
predicted predicted
actual actual
0.25 0.25
0.2 0.2
frac. porosity
frac. porosity
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
test well no. test well no.
(a) (b)
Fig. 11. Plots showing neutron porosity prediction error for (a) layer 1 (upper) and (b) layer 2 (lower) predictions. The RMS error for layer 2, 0.
0389 porosity units, is less than the RMS error for layer 1, 0.0486 porosity units.
14 T. AHMED, C.A. LINK, K.W. PORTER, C.J. WIDEMAN, P. HIMMER, J. BRAUN SPE 38985
25 0.2 25
0.2
20 20
simulator row. no.
0.05 0.05
5 5
-----> ----->
N N
0 0 0 0
0 5 10 15 20 25 0 5 10 15 20 25
simulator col. no. simulator col. no.
(a) (b)
Fig. 12. Plots showing the neural network predicted porosity distribution for (a) layer 1 and (b) layer 2 merged with contoured values outside the
prediction area. Layer 2 porosity highs correlate with seismic amplitude highs taking into account that column and row simulator grid spacing is
not regular as suggested in the figures. North is to the right. These values were used for input into the simulator.
350
25 500 25
300
20
simulator row. no.
400 20
simulator row. no.
250
15 300 15 200
150
10 200 10
100
5 100 5
-----> -----> 50
N N
0 0 0 0
0 5 10 15 20 25 0 5 10 15 20 25
simulator col. no. simulator col. no.
(a) (b)
Fig. 13. Plots showing the neural network predicted permeability distribution for (a) layer 1 and (b) layer 2 merged with conventionally
contoured permeability values outside the prediction area. Measured permeability values were obtained from core analyses on five wells. North
is to the right. Note, simulator row and column spacing is not regular as suggested in the figures.
SPE 38985 NEURAL NETWORK PARAMETER PREDICTION IN RESERVOIR CHARACTERIZATION AND SIMULATION 15
Fig. 14. Actual (points) and simulator predicted (solid line) flow rates for the discovery well, Amoco #1 USA Erving Wolf, from 1973 through
1995.
Oil flowrate, STB/month
Fig. 15. Actual (points) and simulator predicted (solid line) total production rate for the Rabbit Hills Field from 1973 through 1995.
Cumulative Oil Produced, MSTB
Fig. 16. Actual (points) and simulator predicted (solid line) overall cumulative oil production for the Rabbit Hills Field from 1973 through 1995.
The predictions overly the actual production values.