You are on page 1of 11

Journal of Petroleum Science and Engineering 49 (2005) 212 222 www.elsevier.

com/locate/petrol

Application of artificial neural networks for reservoir characterization with limited data
K. Aminian *, S. Ameri
Petroleum and Natural Gas Engineering Department, West Virginia University, P.O. Box 6070, Morgantown, WV, USA Accepted 20 May 2005

Abstract The production performance of a heterogeneous reservoir cannot be realistically predicted without accurate reservoir characterization. Lack of detailed permeability data often hampers the reservoir characterization efforts. In this study, artificial intelligence techniques were successfully utilized to predict the missing information. More specifically, the objective of this study was to develop an accurate reservoir description with the aid of artificial neural networks (ANN) utilizing available geophysical well log data and limited core data. The superior ability of ANN for pattern recognition makes it a prime candidate for permeability prediction from well log data. The most complete reservoir description is generally provided through identification and characterization of the flow units. The presence of flow units suggests that separate permeability-log data relationship exists for each unit. The flow units in the wells with core data were first defined using self optimizing (Kohonen) neural network to cluster core-log data in a systematic manner. Utilizing the flow unit definition, a back-propagation artificial neural network to predict the flow units in the wells without core data was trained and tested. A second back-propagation artificial neural network was then trained and tested to predict the permeability profile within each flow unit. However, the use of limited permeability measurements to train and test the second network resulted in inconsistent predictions due to sensitivity to the arrangement of the data. An innovative approach for training and testing of the ANN was then developed to overcome the problem and to provide consistent and reliable predictions. The two networks which predicted the flow units and their attributes were utilized to develop the description of the reservoir. Primary and secondary production data were utilized to evaluate the reservoir description. The reservoir description substantially improved the simulation of the secondary recovery performance. The simulation results confirmed the presence of heterogeneities which have had profound impact on the performance of the reservoir. The methodology presented in this paper can serve as a guideline for the characterization of reservoirs with limited data. D 2005 Elsevier B.V. All rights reserved.
Keywords: Reservoir characterization; Flow units; Artificial neural networks

1. Introduction Significant amount of oil remains in many depleted or marginally productive oil reservoirs. Further develop-

* Corresponding author. E-mail address: khaminian@mail.wvu.edu (K. Aminian). 0920-4105/$ - see front matter D 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.petrol.2005.05.007

ment of these reservoirs is often hampered by the lack of sufficient data to evaluate their potential and predict their performance. The production performance of a reservoir is invariably influenced by the degree of heterogeneity present. Therefore, production performance of a heterogeneous reservoir cannot be realistically predicted without accurate reservoir characterization. While a number of techniques for reservoir characterization exist, the

K. Aminian, S. Ameri / Journal of Petroleum Science and Engineering 49 (2005) 212222

213

need for detail geological and reservoir data often preclude their applications in reservoir with limited data. The key parameter for reservoir characterization is the permeability distribution. In most reservoirs permeability measurements are rare and therefore permeability must be predicted from the available data. Generally, geophysical well logs are the most abundant source of data. Predicting permeability from well log data in heterogeneous reservoirs is a complex problem. Statistical methods often fall short of the accuracy needed for permeability prediction from well log data in heterogeneous reservoirs. Previous investigation (Mohaghegh, 2000; Aminian et al., 2000, 2001) indicated that artificial neural networks (ANN) are superior to statistical methods in predicting permeability from well log data because of their excellent pattern recognition ability. The ability of neural networks to accurately predict permeability in heterogeneous reservoirs can be enhanced by subdividing the formations into several zones that have distinct lithological, petrographical, or petrophysical attributes. The most complete reservoir description is generally provided through identification and characterization of the flow units. Flow unit is a reservoir rock volume within which geological and petrophysical properties that affect fluid flow are internally consistent and predictably different from properties of other rock volumes (Ebanks, 1987). Flow unit model integrates geological and petrophysical properties by subdividing the reservoirs into volumes that approximate the architecture of a reservoir at a scale consistent with reservoir simulations. The objective of this study was to evaluate the ability of neural networks to identify, predict, and characterize the flow units in a reservoir with limited core data. An oil reservoir that was discovered in 1895 in West Virginia was selected for this study. In 1981, a pilot waterflood was commenced in the field. Based on the pilot recoveries; development proceeded to a full-scale waterflood. Over 140 new wells were drilled for waterflooding purposes. Core and core analyses were available from 6 wells, which were drilled during the waterflood evaluation period. The production-injection data were available from several 5-spot patterns. This offered the unique opportunity to verify the reservoir description by history matching with a reservoir simulator. 2. Background The attempts to predict permeability from well log data has generally been in the form of empirical correlations between permeability, porosity, and water saturation (Timur et al., 1968). Prediction of permeability

from porosity begins with the relationship between core porosity and log porosity. This relationship is then generalized by calibration of well logs so that permeability can be predicted from log porosity throughout the reservoir. This technique has been used with some success in sandstone (Weber and Van Geuns, 1990) and carbonate (Lucia, 1983) reservoirs; however, the existing correlations are mainly for homogeneous formations that have fairly constant porosity and grain size. The prediction of permeability in heterogeneous formations from well log data poses a difficult and complex problem. A comprehensive approach (Molnar et al., 1994) for correlating permeability with geophysical well log data in heterogeneous formations was previously developed. This approach combined gamma ray, deep induction, and compensated bulk density well log responses and detailed core analysis to subdivide the formation into several zones. Then, statistically viable and verifiable correlations between permeability and bulk density were developed for each zone. Previous investigations (Mohaghegh et al., 1996) revealed that neural network is a powerful tool for identifying the complex relationship among permeability, porosity, fluid saturations, depositional environments, lithology, and well log data. Artificial neural systems, or neural networks, a biologically inspired computing scheme, process data and learn in a parallel and distributed fashion. Neural networks can discover highly complex relationships between several variables that are presented to the network. Mohaghegh (2000) noted pattern recognition as one of the neural networks strengths. Artificial neural networks learn the nature of the dependency between input and output variables. Learning is based on pattern recognition. The network classifies new patterns and predicts an output based on the learned patterns. Neural networks often have application when relationships of parameters are too complicated or require too much time to solve via conventional methods. There are two types of neural networkssupervised and unsupervised. The unsupervised or self-optimizing network classifies a set of data into a specified number of categories. The most common unsupervised network is Kohonen network, which has a simple architecture of one layer of input and output. As with all unsupervised networks, the Kohonen network is provided the input data and learns without being shown the correct output. Supervised networks classify patterns and make decisions based on the patterns of inputs and outputs they have learned. Providing correct answers is an integral part of the learning/training process. Back-propagation is a common supervised network type. A typical ANN has three

214

K. Aminian, S. Ameri / Journal of Petroleum Science and Engineering 49 (2005) 212222

layers (input, middle or hidden, and output layer). Each layer is connected to the next neuron layer. As a result, the output of the first layer becomes the input for the next layer. The ANN learns by repeatedly adjusting the weighting of the inputs until the results produced are similar to the correct outputs in the training set. As a result of the non-linear weighting the neural network can handle very complex problems. A useful ANN must be able to predict an output with good results for data not previously seen by the network. The database to be introduced to the neural network is broken down into three groups: training, test, and verification. The training set is used to train and create the network. The actual output of the training set data is used to develop the weights in the network. At established intervals, the test set is used to evaluate the predictive ability of the network. The test set also insures that network would not memorize the data. Training continues as long as the computed error between the actual and predicted outputs for the test set is decreasing. Typically 80% of the data is used for training and testing. The other 20% of the data is categorized as verification or the production set. The verification set is used to evaluate the accuracy of the newly built network by providing the network a set of data it has never seen. 3. Study area A reservoir characterized by severe heterogeneity due to a complex interplay of stratigraphic and diage-

netic effects was selected for detailed analysis in this study. Like many of the oilfields in West Virginia, oil is trapped downdip of gas accumulated along the structural highs. The producing horizon in this field is an Upper Devonian Sandstone. Gross thickness can be up to 50 ft. The pay thickness varies from 4 ft up to 20 ft. Permeability can range from less than 5 millidarcy (md) to values in excess of 250 md. The primary production was estimated at 20 million barrels of oil and the original oil in place was estimated at 88.5 million barrels. The primary drive mechanism was solution gas and gravity drainage. The initial production (IP) of these wells varied from 0 to 300 and averaged 72 barrels of oil per day. This widely varied initial production is an indicator of the reservoir heterogeneity. In 1981, a pilot waterflood was commenced in the field. The pilot consisted of an approximately 34 acre dual five-spot pattern. Based on the pilot recoveries, development proceeded to a full-scale waterflood. The fullscale waterflood began in 1990. Since 1990, more than 100 new wells have been drilled for water injection and 40 new wells drilled for production. To date over 2 million barrels of oil have been produced as a result of the full-scale waterflood. Stratigraphic and petrographic analysis of cores and correlation of logs were combined to describe various sedimentary bodies or bfaciesQ within the formation (Ameri et al., 2001). Productive sandstones in the field lie within Upper Devonian sandstones. These sandstones are interpreted as shoreline sandstones and are comprised of five different lithofacies stacked into

Fig. 1. Distribution of main sandstone in parasequences A, B, and C in the reservoir.

K. Aminian, S. Ameri / Journal of Petroleum Science and Engineering 49 (2005) 212222

215

three parasequences. Two of lithofacies are shale and are useful for field-scale correlation. The other three lithofacies are sandstones. The upper sandstone unit (Uss) is Conglomeratic sandstone which has widely varied permeability values but a thin pay appears to be present. The middle sandstone (Mss) is the most common in the reservoir and contains the main pay zone. The lower sandstone (Lss) contains very thin pays that are discontinuous across the reservoir and their connection to the rest of the pay is questionable. The primary reservoir consist of three parasequences namely A, B, and C that are generally separated by thin shales interbeded with the sandstone. Parasequence A has all the lithofacies. Uss is most prevalent in the upper part of A. The other sandstones occupy the lower part of A. Mss is thicker in the west while Lss is thicker in the east. No Uss has been observed in Parasequences B or C. Parasequence B grades from shale in the west to interbeded sandstones in the east. Mss in Parasequence C is limited to thin zones in the eastern and northern portions of the field. Fig. 1 illustrates the thickness of Mss in parasequences A, B, and C. Strata of parasequences A and B, although at different stratigraphic positions, are present at the same subsurface elevation because of the structural plunge in the reservoir. This leads to an important heterogeneity within the reservoir due to presence of multiple pay units in some portions of the field. 4. Objective and methodology The goal of this study was to use artificial neural networks (ANN) to develop an accurate reservoir description when permeability data are limited. More specifically, the objective of this study was to develop a methodology for predicting the distribution and characteristics of flow units using available geophysical well log data. This study consisted of 4 parts namely: Flow Unit Identification, Flow Unit Prediction, Flow Unit Characterization, and Reservoir Characterization. Methodology, results, and analysis for each part will be presented in the following sections. 5. Flow unit identification The flow unit approach provides a means of uniquely subdividing reservoirs into volumes that within which petrophysical and fluid flow properties are internally consistent and predictably different from properties of other rock volumes. Generally, flow units are defined based on porosity-permeability relationships through statistical techniques. However in this study,

it was attempted to subdivide the formation into several bcategoriesQ or bclusterQ according to the similarities of permeability and all the available well log data using a self-optimizing neural network. Permeability profiles for 6 wells were available from the core analyses. These wells, which were drilled during the waterflood evaluation period, were selected as the primary study wells and are referred to here as wells B-18, B-19, H-8, H-9, H-11, and M-13. All the wells, drilled for waterflooding purposes, had gamma ray and density logs available. Only a limited number of wells had neutron porosity and induction logs available. The data from the six primary study wells were provided to the network as input. Four types of inputs were provided to the network. The first type included the digitized value of gamma ray and bulk density logs and corresponding depths. The selection of gamma ray and bulk density logs was primarily dictated by their availability in the majority of the wells in the field. The second type of input consisted of the first derivatives of the bulk density and gamma ray log values with respect to depth. Previous investigations (White et al., 1995) have revealed that ANN could not provide consistent and reliable predictions if only the log data are provided as input. In general, the well logs contain recorded information about rock properties and thus are subject to noise. The inclusion of the derivatives can alleviate this problem by allowing the network to recognize the changes in the shape of the various log responses. The third type of input data consisted of the well coordinates and log baselines. The measured permeability values from core analyses were the forth type of input. It should be noted that the scales of log and core measurement are different. However, self-optimizing network is capable of handling inputs at different scales. Several ANNs were considered to categorize the data into several different number of clusters. Stable results were obtained only with the 3-cluster network. These clusters are numbered in the order of the increasing depth here. In some of the study wells, the first cluster contained two sections: one in the upper and the other in the lower part of the formation. Since these two sections are not in direct communication with each other, they cannot form one flow unit. However, their characteristics appear to be similar. The unit, which is identified as the upper part of the cluster 1, was classified as flow unit 1 here. The second cluster, identified by ANN, was a thin discontinuous unit since it appeared to be present only in some of the study wells. Further evaluation of the formation characteristics in this cluster revealed that the formation properties

216

K. Aminian, S. Ameri / Journal of Petroleum Science and Engineering 49 (2005) 212222 Table 2 R 2-values for flow unit predictions by different networks Well Flow units 1, T, 2 and 3 B-18 B-19 H-8 H-9 H-11 L-13 0.70 0.45 0.50 0.50 0.70 0.70 1, 2, and 3 0.70 0.40 0.00 0.90 0.90 0.50 1, T, and 2 0.90 0.90 1.00 0.90 0.80 1.00 1 and 2 1.00 0.50 0.00 0.90 0.90 1.00

are intermediate between clusters 1 and 3. Therefore, this cluster represents a transition zone and cannot be classified as a flow unit since its continuity is questionable. Cluster 3 is the most prominent and persistent unit in the reservoir and is classified here as flow unit 2. The lower section of the cluster could potentially be classified as flow unit 3. Flow unit 2 is considered to be the major producing unit in the reservoir because of its thickness and permeability while flow unit 1 is a minor contributor to the total flow capacity. Flow unit 1 corresponds to the lower part of Uss and the upper part of the Mss. Flow unit 1 consists of the fine to medium grain-size sandstone with high porosity and average permeability. Flow unit 2 corresponds to the middle and lower parts of Mss. It consists of homogeneous fine-grained sandstone with high porosity and high permeability. These descriptions of flow units 1 and 2 indicate that flow units do not necessarily coincide with geological lithofacies. Flow unit 3 corresponds to the thin pay section in Lss. Its characteristics are similar to flow unit 1. It is however uncertain that this unit is truly a flow unit due to lack of continuity and low flow capacity. The results of the self-organizing network predictions and final assigned flow units are summarized in Table 1 for all the study wells. Theses definitions and boundaries of flow units are similar to those determined by statistical techniques (Aminian et al., 2002, 2003; Mustafa, 2003). However, the 3-cluster self-optimizing network revealed the presence of a transition zone between flow units 1 and 2 that could not be identified by statistical methods. Further analysis of data in this zone indicated that well log responses in this zone exhibit similar trends as those for flow unit 2 while the permeability profiles are similar to flow unit 1. As it can be seen from Table 2, two of study wells did not contain flow unit 2. Fig. 1 indicates that Mss is present throughout the reservoir. Therefore, presence of the Mss does not guaranty presence of flow unit 2. In fact, the Mss must achieve sufficient thickness for flow unit 2 to
Table 1 Summary self-organizing network results Well B-18 B-19 H-8 H-9 H-11 L-13 T = Transition. Clusters 1, 1, 1 1, 1, 1 T, 2, and 1 T, 2, and 1 T, and 2 2, and 1 Flow units 1, 1, 1 1, 1, 1 T, 2, and 3 T, 2, and 3 T, and 2 2, and 3

T = Transition.

be developed. This observation leads to conclusion that flow unit 2 in the reservoir is compartmentalized. 6. Flow unit prediction To develop a complete description of reservoir, it is necessary to predict the distribution of flow units in the reservoir. However, the flow units in most of the well in the reservoir cannot be directly identified without permeability data. To predict the flow units in wells without permeability data, it is first necessary to establish certain patterns in well log data corresponding to different flow units. To discover these complex patterns, the more robust supervised back-propagation networks were employed in this part of the study. Several artificial neural network architectures and paradigms were considered to obtain reliable flow unit predictions. Initially, three-layer back propagation architecture was utilized. However, the results with neural network architecture with two hidden slabs with different activation functions and a jump connection yielded better results. The inputs and output to the back-propagation network were similar to those used in self-optimizing networks, except flow units were provided as target outputs during the training and testing. After a series of preliminary runs, it was concluded that the third group of the inputs (the well coordinates and log baselines) diminished networks ability to predict the flow units. Therefore, they were eliminated as input. This resulted in significantly better predictions. To verify the accuracy of the flow unit predictions by ANN, one of the six primary study wells was set aside as a verification well. Subsequently, a network was trained by using the data from the remaining 5 wells. The back-propagation networks were trained by providing flow units as target outputs. Once the network was trained and tested, the log data from the verification (sixth) well were introduced to the network and the flow units were predicted by the network. The predicted flow units were then compared against the flow units identified in the previous section for verifi-

K. Aminian, S. Ameri / Journal of Petroleum Science and Engineering 49 (2005) 212222 Table 3 R 2-values for well H-11 permeability predictions with different arrangement of data Well 1 2 3 4 5 R2 Run 1 H-8 H-9 B-19 M-13 B-18 0.3305 Run 2 H-9 B-19 M-13 B-18 H-8 0.5727 Run 3 B-19 M-13 B-18 H-8 H-9 0.7558 Run 4 M-13 B-18 H-8 H-9 B-19 0.6266 Run 5 B-18 H-8 H-9 B-19 M-13 0.3974

217

cation and a correlation coefficient (R 2-value) was determined. This process was repeated by assigning each of the primary study wells to be the verification well. As was discussed in the previous section, self-optimizing ANN identified four different zones. Two of these zones, classified here as flow units 1 and 2, correspond to main pay sections in the reservoir. The other two zones, classified here as transition and flow unit 3, were discontinuous and uncertain. Therefore, four different sets of flow unit designations as target outputs were considered in this part of the study. The first set consisted of 4 flow units (1, transition, 2, and 3). The second set consisted of three flow units (1, 2, and 3) for which the transition zone was incorporated into flow unit 1. The third set consisted of two flow units (1 and 2) for which the transition zone and flow unit 3 were incorporated into flow unit 1. It should be noted that flow unit 3 was originally identified as flow unit 1 due to its similarity to flow unit 1. Therefore, designating the data points in flow unit 3 as flow unit 1 could improve the network predictions because it creates a larger database for training. Finally, the fourth set consisted of 3 flow units (1, transition, and 2) for which the flow unit 3 was incorporated into flow unit 1. The results of flow unit predictions by various ANN are summarized in Table 2. As it can be seen from by R 2-values in Table 3, consistent predictions were only obtained by the 3-flow unit designation (1, transition, and 2). As was discussed earlier, the btransition zoneQ exhibit well log patterns similar to unit 2 but permeability values are similar to unit 1. In the absence of the permeability data, the network is not capable of correctly identifying the flow units when transition zone is present. This clearly indicates that presence of the transition zone adds more complexity to the prediction flow units from the patterns of well log data. 7. Flow unit characterization To develop a complete description of reservoir, it is not only necessary to predict the distribution of flow

units but it is necessary to predict their characteristics mainly permeability from the well log data. Previous investigations (Aminian et al., 2000) indicated that statistical correlations were not accurate enough for permeability prediction from the well log data. In this study, back-propagation ANNs were utilized for predicting permeability from well log data because of their superior ability over statistical methods for pattern recognition. The network architecture and evaluation procedures employed in this section are similar to those used for flow unit predictions in the previous section. The presence of flow units in the reservoir indicates that the permeability-log data relationships for various flow units are different. Therefore, it is critical to include flow unit designations as one of the input for ANN. Initially, 3-flow unit designation (1, transition, and 2) that provided the best results for flow unit predictions, was used but did not produce good results. The 2-flow unit designation (1 and 2) resulted in significantly better permeability predictions. However, consistent results could not be achieved. The review of the results revealed that the accuracy of the predictions greatly depended on the selection of test set. The common practice is to select the test set randomly. When number of data used to train the network is limited, this random selection could be sensitive to the arrangement of the data. This leads to development of different networks by re-arranging the data. To illustrate this problem, consider the case where study well H-11 was selected as the verification well while the remaining five wells were arranged in just five different ways as shown in Table 3. The R 2-vlaue determined by comparing the actual permeabilities to the ANN predictions for each case is also given in Table 3. As it can be seen, the results are not consistent. It is possible to re-arrange the data until a reasonable R 2-vlaue is obtained. However, this approach will lead to different arrangements of data to predict the permeability accurately in each of verification wells. As a result, 6 totally different ANNs (one for each verification well) for permeability predictions will be developed. However, it is not possible to determine which one of these six networks provides the best predictions for the wells without permeability data. To develop a procedure that can be applied for permeability prediction field-wide, it was proposed to use a common test set for all the networks. This set must capture the overall characteristics of the permeability profiles. A methodology was devised for common test selection that provided consistent and reliable results. The first step is to identify the input (log data) that has strong influence on the output (permeability). In this study, bulk density appeared to have the highest

218

K. Aminian, S. Ameri / Journal of Petroleum Science and Engineering 49 (2005) 212222

Fig. 2. Test set selection method for training permeability prediction ANN.

impact on permeability. The second step involves a plot of the desired output (permeability) versus the input with highest impact (density). Fig. 2 illustrates the plot of permeability versus bulk density on a semi-log scale. A trend line representing the linear regression is also illustrated. It should be noted that this linear relationship is not a requirement for this methodology. Any reasonable pattern can be utilized. In fact, this linear trend is not statistically significant. Third, the data are divided into segments based on the clustering of the points around the trend line (or any other pattern). In each segment, the highest and the lowest permeability points above and below the trend line as well as one point on or close the trend line were selected for the test set. This approach provided the ANN with test set that had wide variation in output (permeability) values. This approach resulted in excellent permeability predictions for all of the study wells. Table 4 summarizes the
Table 4 R 2-values for permeability predictions by ANN Well B-18 B-19 H-8 H-9 H-11 L-13 R2 0.976 0.89 0.90 0.89 0.90 0.93

results of the permeability predictions. The detailed aspects of permeability predictions with ANN can be found elsewhere (Oyerokun, 2002). 8. Reservoir characterization To characterize the reservoir, it is necessary to predict the distribution and characteristics of flow units throughout the reservoir. Flow unit 1 corresponds to the lower part of Uss and the upper part of Mss. Therefore, flow unit 1 is expected to be always in parasequence A since Uss is only present in parasequence A. Flow unit 2 corresponds to the middle and lower parts of Mss. Consequently, flow unit 2 could be present in all three parasequences where Mss is present, provided it achieves sufficient thickness. It should be noted that the flow units 1 and 2 in all the primary study wells, except B-19, were in parasequence A. The data from study well B-19 does not reflect a significant difference between the properties of the flow unit 2 in parasequences A and B. Therefore, it was assumed that the flow unit ANN that was developed mainly for parasequence A can be extended to all parasequences. However, wherever Mss in parasequences A and B overlaps (see Fig. 1), flow unit 2 could be present in both parasequences. The core and log data indicate that parasequences A and B are separated by shales throughout most of the reservoir. This could lead to

K. Aminian, S. Ameri / Journal of Petroleum Science and Engineering 49 (2005) 212222

219

Fig. 3. Distribution of flow capacity of flow units in the reservoir.

the presence of two separate flow units in parasequences A and B. Study well B-19 was the only well in which flow unit 2 appeared to be present in both parasequences. However, presence of shale between the two parasequences in well B-19 is difficult to confirm. Since it is not possible to use ANN predictions to identify separately the flow units in parasequences A and B, the impact of overlapping parasequences will be investigated by production performance predictions. The final flow unit prediction and permeability prediction ANNs were developed utilizing all the six study wells and the common pre-selected test set. The log data from 125 wells in the field were used as input for flow unit network to predict the flow units in each of 125 wells. The results were then used as the input for permeability prediction network to predict the permeability profile for each well. The permeability network that provided the best predictions had been developed by assuming two flow units. The first unit consisted of flow unit 1, transition zone and flow unit 3. The second unit was the same as the main pay, flow unit 2. The key issue in application of permeability prediction network was the transition zone. The transition zone even had log responses similar to flow unit 2; its permeability characteristics were similar to flow unit 1. Therefore, the flow unit network predictions were modified before applying permeability prediction network. This was achieved by reclassifying the transition zone as flow unit 1 prior to applying the permeability prediction network. The predicted permeability values were combined with the flow unit thickness data to determine flow

capacity (permeability-thickness product) for each flow unit in each well. The results were utilized to generate field maps showing the distribution of flow capacity for each flow unit. This provided a detailed description of reservoir. Flow unit 2 is the major pay in the reservoir and accounts for most of the flow capacity in the reservoir. Fig. 3 illustrates the distribution flow capacity in flow units 1 and 2.

Fig. 4. Distribution of average per well primary recovery in the reservoir.

220

K. Aminian, S. Ameri / Journal of Petroleum Science and Engineering 49 (2005) 212222

Fig. 5. Pilot and ball area flood patterns.

History matching the production performance is the main technique for evaluation of the reservoir description. The confirmation of the reservoir description will also verify the methodology developed for flow unit and permeability predictions. The annual primary production record was available on a lease (farm) basis for the entire life of the field. Lack of detailed well production histories precluded production history matching with a reservoir simulator. However, a qualitative assessment of reservoir description was obtained by comparing the distribution of the average cumulative primary production and distribution of flow unit flow capacity. Fig. 4 illustrates a primary production map that has been generated using the available lease cumulative production data. Fig. 4 correlates well with flow capacity distribution for flow unit 2 in Fig. 3. This qualitative comparison provides a preliminary verifica-

tion of the methodology used to characterize the reservoir. It further indicates that flow unit 2 is the major contributor to the production while flow unit 1 is minor contributor and flow unit 3 has no significant contribution to primary production. Further verification was attempted by history matching secondary recovery performance. The secondary recovery production-injection data were available from several 5-spot patterns. This offered a unique opportunity to evaluate the reservoir description by simulating the production history. Initially, the simulation study was focused in pilot area (see Fig. 5) where a significant amount of data was available. The pilot consisted of two 5-spot patterns (P1 and P2) as shown in Fig. 5. Two of the primary study wells, H-9 and H-11, are the injection wells in these patterns. The two patterns have similar sizes and water injection histories. However, P2 pro-

Fig. 6. Pilot flood patterns oil production history and simulation results.

K. Aminian, S. Ameri / Journal of Petroleum Science and Engineering 49 (2005) 212222

221

duced less than 15,000 barrels of oil while P1 produced over 25,000 barrels of oil over the same time period as illustrated in Fig. 6. This is clearly due to distribution and characteristics of flow units in these patterns. Therefore, they serve as excellent cases for evaluation of the reservoir description. The collected well records, well logs, and core analyses from the patterns were utilized to generate the necessary input for the flow unit and permeability prediction ANNs. The results of ANNs predictions were then utilized to generate the description of the reservoir in these patterns. The reservoir description was based on two flow units. Flow unit 3 was not considered to be contributing to production since the injection wells were not perforated in this zone. Transition zone was incorporated in flow unit 1 due to its permeability characteristics. The reservoir description and injection-production data were then used in conjunction with a reservoir simulator to predict the waterflood performance in these patterns. The simulation results are summarized in Fig. 6 for both patterns. The results of simulation closely matched the production history, which indicates flow unit predictions provided accurate description of the reservoir. It should be noted that traditional approach of using statistical methods to characterize the pilot area with core and well log data failed to match the production history. The detail infor-

mation regarding the pilot area simulation studies are documented elsewhere (Gil, 2000). To investigate the impact of overlapping parasequences A and B on flow unit distribution, two adjacent five-spot patterns were selected that were situated on a NESW trend over which Mss achieves maximum thickness (see Fig. 1). The patterns, referred to here as Ball 18 and Ball 21, are located northeast of the pilot area as illustrated in Fig. 5. The primary study well B18 is the producer in one pattern while primary study well B-19 is the injector in the other pattern. These patterns are among the most prolific oil producers in the field. The core descriptions from study wells B-18 and B-19 have revealed that flow unit 2 in B-18 is entirely in parasequence A and in B-19 in both parasequences. The collected well records, well logs, and core analyses from the 8 wells in these patterns were utilized to generate the necessary input for the flow unit and permeability prediction ANNs. The results of ANNs predictions were then utilized to generate the description of the reservoir in these patterns. Two separate models were utilized for simulation purposes. The first model (2-flow unit) combined the flow unit 2 in two parasequences into one continuous unit. The second model (3-flow unit) assigned two separate flow units, with no vertical communication, in each parase-

Fig. 7. Ball area flood patterns oil production history and simulation results.

222

K. Aminian, S. Ameri / Journal of Petroleum Science and Engineering 49 (2005) 212222 Ameri, S., Aminian, K., Avary, K.L., Bilgesu, H.I., Hohn, M.E., McDowell, R.R., Matchen, D.L., 2001. Reservoir Characterization of the Upper Devonian Gordon Sandstone, Jacksonburg Stringtown Oil Field, Northwestern West Virginia, Final Report, Performed under Contract No. DE-AC26-98BC15104, October 2001. Aminian, K., Bilgesu, H.I., Ameri, S., Gil, E., 2000. Improving the simulation of waterflood performance with the use of neural networks. SPE Paper 65630, Proceeding of SPE Eastern Regional Conference, October 2000. Aminian, K., Thomas, B., Bilgesu, H.I., Ameri, S., Oyerokun, A., 2001. Permeability distribution prediction. SPE Paper, Proceeding of SPE Eastern Regional Conference, October 2001. Aminian, K., Thomas, B., Ameri, S., Bilgesu, H.I., 2002. A new approach for reservoir characterization. SPE 78710, Proceeding of SPE Eastern Regional Conference, October 2002. Aminian, K., Ameri, S., Bilgesu, H.I., Alla, V., Mustafa, R., 2003. Characterization of a heterogeneous reservoir in West Virginia. SPE 84830, Proceeding of SPE Eastern Regional Conference, September 2003. Ebanks, W.J., 1987. Flow unit concept-integrated approach for engineering projects. Abstract, AAPG Annual Convention. Gil, E., 2000. Improving The Simulation of a Waterflood Recovery Process Using Artificial Neural Networks, MS thesis, West Virginia University, December 2000. Lucia, F.J., 1983. Petrophysical parameter estimated from visual description of carbonate rock: a field classification of carbonate pore space. JPT 35 (3). Mohaghegh, S., 2000. Virtual-intelligence applications in petroleum engineering: Part 1. Artificial Neural Networks. J. Pet. Technol. Distinguished Author Series, September 2000, pp. 6473. Mohaghegh, S., Ameri, S., Aminian, K., 1996. A methodological approach for reservoir heterogeneity characterization using artificial neural networks. J. Pet. Sci. Eng. 16, 263 274. Molnar, D., Aminian, K., Ameri, S., 1994. The use of well log data for permeability estimation in a heterogeneous reservoir. SPE 29175, Proceedings of SPE Eastern Regional Conference, pp. 167 180. Mustafa, R., 2003. Application of Statistical Methods for Flow Unit Identification and Characterization of Reservoir Using Well Log and Core Data, MS thesis, West Virginia University, May 2003. Oyerokun, A., 2002. A New Approach for Training and Testing Artificial Neural Networks for Permeability Prediction MS thesis, West Virginia University, May 2002. Timur, A., 1968. An investigation of permeability, porosity, and water saturation relationship for sandstone reservoirs. Log Anal. 9 (4). Weber, K.J., Van Geuns, L.C., 1990. Framework for constructing clastic reservoir simulation model. JPT 42 (10). White, A.C., Molnar, D., Aminian, K., Mohaghegh, S., Ameri, S., Esposito, P., 1995. The application of ANN for zone identification in a complex reservoir. SPE 30977, Proceeding of SPE Eastern Regional Conference, October.

quence. Available well logs were utilized to determine the distribution of flow units in each parasequence. The production history and simulation results are illustrated in Fig. 7. The detail information regarding these simulation studies are documented elsewhere (Alla, 2002). The results in Fig. 7 indicate that 3-flow unit model provides a better match for production history in both patterns. This conclusion confirms the presence of two separate flow units in areas where parasequences A and B overlap. Further, it can be concluded that the predicted permeability for flow unit 2 in parasequence B is reliable. This is significant because permeability prediction ANN was trained using permeability data from flow unit 2 in parasequence A. 9. Conclusions The following conclusions were reached during this study: 1. Self optimizing (Kohonen) neural network was able to identify two flow units and a transition zone between the two flow units in the reservoir. 2. The flow units do not correspond to lithological units in the reservoir and appear to be compartmentalized. 3. The flow unit definitions, based on self optimizing neural network, were instrumental in development of a back-propagation artificial neural for predicting the flow units from well log data. 4. Flow unit designations are significant input for backpropagation artificial neural network for permeability from well log data. 5. An innovative methodology for test data selection was developed which significantly enhanced the abilty of the neural networks for permeability prediction. 6. A complete description of reservoir model was obtained by combining neural networks predictions of flow unit and permeability from well log data. 7. Reservoir model developed based on the flow unit descriptions was verified by comparison against primary and secondary history production history. 8. Simulation of waterflood performance indicated the presence of two separate but similar flow units in areas where parasequences A and B overlap. References
Alla, V., 2002. Using Waterflood Performance to Characterize FlowUnits in A Heterogeneous Reservoir, MS thesis, West Virginia University, December 2002.

You might also like