Surrogate Reservoir Model

Modeling and Analysis
Modeling pressure and saturation distribution in a CO2 storage project using a Surrogate Reservoir Model (SRM)
Alireza Shahkarami, Shahab D. Mohaghegh, Vida Gholami, Alireza Haghighat, and Daniel Moreno, West Virginia University, Morganstown, WV, USA Abstract: Capturing carbon dioxide (CO2) from large point sources and depositing it in a geological formation is an efcient way of decreasing CO2 concentration in the atmosphere. A comprehensive study is required to perform a safe and efcient CO2 capture and storage (CCS) project. The study includes different steps, such as selecting proper underground storage and keeping track of CO2 behavior in the storage environment. Numerical reservoir simulators are the conventional tools used to implement such an analysis. The intricacy of simulating multiphase ow, having a large number of time steps required to study injection and post-injection periods of CO2 sequestration, a highly heterogeneous reservoir, a large number of wells, etc., will lead to a complicated reservoir model. A single realization for such a reservoir takes hours to run. Additionally, a thorough understanding of the CO2 sequestration process requires multiple realizations of the reservoir model. Consequently, using a conventional numerical simulator makes the computational cost of the analysis too high to be practical. In this paper, we examine the application of a relatively new technology, the Surrogate Reservoir Model (SRM), as an alternative tool to solve the aforementioned problems. SRM is a replica of full-eld reservoir simulation models. It can generate outputs in a very short time with reasonable accuracy. These characteristics make SRM a unique tool in CO2 sequestration modeling. This paper proposes developing an SRM for a CO2 sequestration project ongoing in the SACROC unit to model pressure behavior and phase saturation distributions during different time steps of the CO2 storage process. 2014 Society of Chemical Industry and John Wiley & Sons, Ltd Keywords: CO2 sequestration and storage; fast track modeling; pattern recognition; surrogate reservoir models (SRMs)
Introduction
arbon dioxide (CO2) is the primary greenhouse gas (GHG) that has been contributing to global warming and climate change since the beginning
of the Industrial Revolution. CO2 comprises nearly 80% of global anthropogenic (produced by human activity) GHG emissions.1 The atmospheric concentration of CO2 recently reached a considerable level of 400 ppm in May 2013 an almost 100 ppm increase
Correspondence to: Department of Petroleum and Natural Gas Eng, Mineral Resources Building, Room 147, Evansdale Drive, Morgantown, West Virginia, 26506, USA. E-mail: Alireza.shahkarami@gmail.com Received August 14, 2013; revised January 9, 2014; accepted January 10, 2014 Published online at Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/ghg.1414
2014 Society of Chemical Industry and John Wiley & Sons, Ltd | Greenhouse Gas Sci Technol. 4:127 (2014); DOI: 10.1002/ghg
A Shahkarami et al. Modeling and Analysis: Modeling pressure and saturation distribution in a CO2 storage project using a Surrogate Reservoir Model (SRM)
since 1960.2 Fossil fuel use is considered the main source of CO2 emission. Even with considering new policies of CO2 emissions,3 it is unlikely that there will be a significant decrease over the next 25 years in the percentage of world energy produced by fossil fuels (81% in 2010). Therefore, mitigating the amount of CO2 coming from human activities is a major challenge in reducing the anthropogenic effects on global warming and climate change. The Intergovernmental Panel on Climate Change (IPCC) defines carbon capture and storage (CCS) as a process consisting of the separation of CO2 from industrial and energy-related sources, transport to a storage location and long-term isolation from the atmosphere.4 The geological storage of CO2 is the injection of the captured CO2 into appropriate deep geological formations. The geological sequestration of CO2 is not a new technology. In the early 1970s, CO2 was injected for the first time into subsurface geological formations in Texas in order to enhance the oil recovery.57 However, it was not until the 1990s when the geological storage of CO2 gained enough credibility to be applied in a large-scale project.8,9 In 1991, the Norwegian government instituted a tax on CO2 emission, which motivated Statoil to run the first commercial CCS project in order to capture CO2 from the Sleipner oil and gas field in the North Sea and inject it into a thick layer saline aquifer in 1996.10 In less than two decades, carbon storage in deep geological formations has emerged as one of the most important options for reducing CO2 emissions.4,11,12 CCS plays a critical role in the portfolio of technologies required to attain a considerable reduction of global GHG emissions in the most economically efficient manner. This technology has the potential to decrease nearly one-fift h of the emissions required to cut GHG emissions from energy consumption in half by 2050.13 CO2 leakage is one of the major risks in a CCS project; therefore, keeping CO2 in a safe and controlled environment for a long period of time is a main challenge.14,15 Consequently, the following tasks must be thoroughly accomplished: quality control of candidate underground storage, keeping track of CO2 plume conditions, and simulating the reservoir behavior (such as reservoir pressure, which is an appropriate indicator of potential leakage). Numerical reservoir simulators are the conventional tools used to perform the aforementioned tasks.1621 In order to have a comprehensive study of a CCS project, hundreds to thousands of realizations with
different reservoir characteristics and operational conditions are required. Although using a numerical reservoir simulator gives accurate results, it is very time-consuming and computationally expensive. Furthermore, due to the process of CO2 sequestration, a compositional simulator should be utilized, which generally leads to an even higher computational time.22 The reservoir simulation model in this work comes from the work done by Han.23 The original geo-cellular model in that work consisted of over nine million grid blocks. In order to simulate CO2 trapping mechanisms, he had to upscale the model and decrease the number of grid blocks to 13 600. However, with the time lapse required to run this study (1000 years), even this upscaled model requires a high computational cost and takes hours to run a single realization. The reservoir simulation in this study was conducted using Computer Modeling Group (CMG) simulator called GEM-GHGTM.24 GEM-GHGTM is specifically designed for simulating CO2 sequestration processes. The objective of this study is to examine the effect of the uncertainty involved in a reservoir parameter (permeability) and also the impact of operational constraints on the output of numerical reservoir simulators (pressure and phase saturations). The tool to accomplish the objectives of this study is a patternrecognition-based technology known as Surrogate Reservoir Models (SRMs). SRMs have been introduced as a tool for addressing many time-consuming operations performed with reservoir simulation models.25 SRMs attempt to reduce the dimensionality of the problem by using fuzzy pattern recognition techniques. The capability of SRMs to replicate full field models that run in fractions of a second makes them an efficient alternative tool to address many time-consuming operations performed with reservoir simulation models.25 The engines of SRMs are based on Artificial Neural Networks (ANNs). In order to develop the SRM, a few different realizations of the reservoir simulation model were created and run by numerical reservoir simulators. The inputs and outputs of these realizations generated a spatiotemporal database. The spatio-temporal database was used to train the SRMs (the pattern-recognition-based models, particularly ANNs). The SRM aims to replicate the results of traditional numerical reservoir simulators at the grid level in a matter of seconds. The SRM passed training, calibration, and validation steps to be qualified as a reliable replica of the reservoir
Modeling and Analysis: Modeling pressure and saturation distribution in a CO2 storage project using a Surrogate Reservoir Model (SRM) A Shahkarami et al.
simulation model. Further validation process is applied to verify the efficiency of the SRM on different realizations of the reservoir simulation model. These realizations were not seen during the training process; therefore they are referred to as blind realizations. At the end of this process the SRM is ready to reproduce the outputs (pressure and phase saturations) of numerical reservoir simulators at the grid block level. The time needed to accomplish each run and achieve the desired results using the SRM is in the order of seconds, whereas the time required to perform the process using a numerical reservoir simulator is in the order of hours and a day.
Potentials of pattern recognition techniques

CO2 sequestration in underground storage is one of the most viable methods for reducing GHGs. The petroleum industry has decades of experience injecting gas (CO2 or hydrocarbons) into different types of reservoirs. This leads to an overlap of issues in the petroleum industry and CCS, such as modeling, history matching, and uncertainty analysis and risk management. These issues can be managed by the capabilities of pattern recognition techniques. In pattern recognition concepts, the data analysis process deals with predictive modeling. By having a high dimensional database, the objective is to learn the underlying behavior in the data and forecast the performance of unforeseen validation database. The learning process refers to some form of algorithm to reduce the error on the set of training data.26 The learning procedures could be distinguished into (i) supervised learning or (ii) unsupervised learning.*27 Supervised learning generally represents a learning procedure which takes an available set of inputs and known outputs corresponding to these inputs. Efforts will be made to build a predictive model by matching the available responses with the inputs. This predictive model is then able to generate reasonable predictions for the response to novel data. The most important characteristic of this learning technique is that the responses (outputs) are recognized or labeled in the training database. On the other hand, unsupervised learning involves only unlabeled data, which makes the process more challenging than the previous one.
*Recently another set of learning was discussed in the literature which is referred as called semi-supervised.
In other words, unsupervised learning forms clusters or natural patterns underlying the structure of data. One of the most famous pattern recognition techniques, one that has a long history in a variety of scientific fields, is ANNs, usually called Neural Networks (NNs). The learning procedure in ANNs is supervised learning. ANNs were originally motivated by the goal of having machines that are able to mimic the brain. In fact, the structure of ANNs is very similar to that of the human neural system, as it includes an interconnected group of artificial neurons. ANNs are cellular systems capable of obtaining and storing information and using experiential knowledge. An ANN is an adaptive system that adjusts its structure based on output and input information that flows through the network during the learning phase.28 Although ANNs have been around for a long time, their popularity in petroleum engineering started only two decades ago.29 Since this time, the applications of ANNs in addressing the conventional problems of the petroleum industry have been widely studied. Some applications of ANNs in petroleum engineering literature include well log interpretation, 3032 well test data analysis,3336 reservoir characterization,3739 calibration of seismic attributes,40 seismic pattern recognition,41 inversion of seismic waveforms,42 prediction of PVT data,4346 fractures and faults identification,4750 hydrocarbons detection,50,51 and formation damage forecast.52,53 That said, it should be noted that the effective use of pattern recognition techniques in the petroleum industry is not a trivial process. It requires insight in both the domain of reservoir engineering as well as a substantial application of pattern recognition techniques; otherwise, the results could be quite disappointing.54
Surrogate reservoir models

Surrogate Reservoir Models (SRMs) are approximations of the full-field 3-dimensional numerical reservoir models that are capable of accurately mimicking the behavior of the full-field models. Unlike statistically based proxy models that require hundreds of simulation runs,5557 SRMs can be created in a few simulation runs. In 2006, SRM was presented for the first time by Shahab D. Mohaghegh to solve the problem of time-consuming runs for an uncertainty analysis of a giant oil field with 165 horizontal wells in
the Middle East. The reservoir simulation model included about one million grid blocks and took 10 h to run using a cluster of 12, 3.2 GHz processors. In his study, SRM was used as an objective function for a Monte Carlo Simulation to build thousands of simulation runs in a very short time compared to numerical simulators. Mohaghegh describes SMRS as an ensemble of multiple, interconnected neuro-fuzzy systems that are trained to adaptively learn the fluid flow behavior from a multi-well, multilayer reservoir simulation model, such that they can reproduce results similar to those of the reservoir simulation model (with high accuracy) in real-time.58 Since 2006, applications of SRMs as an accurate and rapid replica of a numerical simulation model have been reviewed in different studies.5863
SRM development
SRMs are developed using the data extracted from the realizations of simulation model. These data are included in a spatio-temporal database. Building this database is the first step in developing Artificial Intelligence (AI)-based reservoir models. The main objective of this database is to teach the model the whole process of fluid flow phenomena in the reservoir. Therefore meticulous efforts should be made in this part. The quality and quantity of this database determine the degree of success in developing a successful AI-based reservoir model including an SRM. Not dedicating enough attention to this part is the main reason behind unsuccessful attempts at applying AI-based models in the literature.54 Mohaghegh thoroughly discussed this step of SRM development in his paper.64 In order to create the spatio-temporal database, the first step is to identify the number of runs that are required to develop the SRM. The purpose of having different realizations of a reservoir simulation model is to introduce the uncertainties involved in the model to the SRM. This is a common step in building SRMs and developing response surface methods; however, there is a key difference between these two methods: the functional forms behind these models. Response surface and other reduced models are developed using statistical approaches, which use predetermined functional forms. The output of reservoir simulation models are then fitted to these predetermined forms. In order to match these functional forms, hundreds of runs are needed. On the other hand, the pattern
recognition characteristics of SRMs help to develop these types of models by having a small number of simulation runs. However, there is no algorithm to find the optimum number of simulation runs to build an SRM. The common practice when choosing the best number to train the SRM is to use rules of thumb based on the intricacy and heterogeneity of the reservoir model, which might change. Nevertheless, it is obvious that if the number of simulation runs is too small, the SRM will not be able to reproduce the simulator results properly. Otherwise, if the number of simulation runs is too big, there is no reason to develop an SRM since the solution is close to the original problem, which is a high number of simulation runs. After running the realizations, the static and dynamic data are extracted in order to build the representative spatio-temporal database. The database includes different types of data such as static and dynamic reservoir characteristics, operational constraints, etc. Static data refer to properties of the reservoir that are not changing over time, such as permeability, porosity, top, and thickness. Dynamic refers to any data such as well constraints or pressure and phase saturation that change over time.58 The training process of an SRM includes three different steps: training (learning), calibration, and validation procedures. Based on that, the spatio-temporal database is divided into three sets: the training or learning set, calibration set, and validation or verification set. The training set is part of the data shown to the ANNs during the training process. The ANNs are adapted to this set to match the provided outputs (reservoir simulation results). On the other hand, the calibration set is not used to adjust the outputs. This set is utilized to assure that any increase in accuracy over the training data set will lead to an increase in accuracy over a data set that has not been seen by ANNs. This set of data is helpful in determining when the training should be stopped. Finally, the verification set is a part of the database used to verify the predictability of the trained ANN, and subsequently, this data set is not used to train the ANNs. It is worth mentioning that the elapsed time to perform the training process (learning, calibration, and verification) is negligible when compared to the reservoir simulation run-time. Another important point is an SRM may be a collection of several ANNs that are trained, matched and verified in order to generate different results.
A further validation step in the SRM development is utilized to assure its robustness. This step is referred to as blind verification because it is a set of realizations that has not been used during the training process. These blind testing sets are complete realizations of the reservoir, while the verification set used in the training process is a randomly selected portion of spatio-temporal database.
Reservoir simulation model

We received the base reservoir model from a work performed by Han.23 The original reservoir model was for a CO2 enhanced oil recovery (EOR) project that lasted for 200 years, from 1972 to 2172. The model utilized in this study covers the period of January 1, 2172 to January 1, 3172 after the reservoir has been depleted from oil. Therefore the simulation model is just considered for CO2 storage and sequestration. The model contains 25 simulation layers of 1634 grid blocks. There are 45 injection wells planned to inject CO2 at a constant rate (331801.9 m3/day) for 50 years starting in 2172. Each well is perforated in a single layer, although the perforated layers might be different for different wells. The perforations happen in layers 19 (one well), 20 (40 wells), 21 (one well), and 22 (three wells). It is assumed that there is no-flow boundary condition at the outer boundaries. Figure 1 shows a 3-dimensional view of the structure in this simulation model. The objective of this reservoir model is to track the distribution of pressure and phase saturations at the target layer (layer 18) during and after injection of CO2. The total number of grid blocks in this layer is 544, of which only 422 are active. The initial properties (pressure, water saturation and gas mole fraction [CO2], respectively) at layer 18 are shown in Figs 2, 3, and 4. The white grid blocks are null or inactive because they have a negligible thickness value. The initial condition is the condition of the reservoir after 200 years of EOR process (from 1972 to 2172), which comes from the original model.
Field background
The Kelly-Snyder field, discovered in 1948, is one of the major oil reservoirs in the USA, having approximately 2.73 billion bbls of oil originally in place. The early performance history of the field indicated its sole production mechanism as solution gas drive, which could result in an ultimate recovery of less than 20% of the original oil in place. The Scurry Area Canyon Reef Operations Committee Unit (SACROC Unit) was formed in 1953, and in September 1954 a massive pressure maintenance program was started. Water was injected into a center-line row of wells along the longitudinal axis of the reservoir.65 In 1968, a technical committee investigating potential alternatives recommended that a water-driven slug of CO2 be used to miscibly displace the oil in the nonwater-invaded portion of the reservoir; they also recommended that a pattern injection program be developed in this area to implement the slug process and improve ultimate oil recovery. CO2 injection began in early 1972. Investigations of alternative methods for improving recovery in the SACROC Unit showed that an inverted nine-spot miscible flood program consisting of injecting CO2 driven by water would be the most effective and economical option. Under such a scheme, the predicted ultimate recovery would be about 230 million barrels more than what was expected from the original water injection program. 65 The SACROC Unit, within the Horseshoe Atoll, is the oldest continuously operated CO2 enhanced oil recovery operation in the United States, having undergone CO2 injection since 1972. Until 2005, about 93 million tonnes (93,673,236,443 kg) of CO2 had been injected and about 38 million tonnes (38,040,501,080 kg) had been produced. As a result, a simple mass balance suggests that the site has accumulated about 55 million tonnes (55,632,735,360 kg) of CO2.66 Currently SACROC continues to be operated by the current owner/operator, Kinder Morgan CO2.
Uncertain properties and training realizations

In order to introduce the uncertainties involved in the reservoir model to an SRM, a small number of geological realizations were built and run using a commercial numerical reservoir simulator. The numbers of realizations used were 10 and 16 to train the SRM for predicting pressure and phase saturations, respectively. Moreover, three and two realizations were used at the end as the blind runs in order to validate the trained SRM for above-mentioned properties. The variable properties in the realizations consist of permeability distributions at nine layers of the reservoir (layers 1, 2, 19, 20, 21, 22, 23, 24, and 25) and flowing bottom-hole pressure at 45 injection wells.
Figure 1. A three dimensional view of simulation model.
Figure 2. Initial pressure (kPa) at the target layer (layer 18) for the base simulation model.
Figure 3. Initial water saturation at the target layer (layer 18) for the base simulation model.
Figure 4. Initial gas (CO2) mole fraction at the target layer (layer 18) for the base simulation model.
The reason behind varying the permeability distribution maps for only nine layers goes back to the base model. In the base model, the permeability variation is only noticeable in the named layers while it is consistently low in the other layers. Figure 5 depicts the permeability distribution for the layers which were not altered during the SRM development. In this figure the low permeability range (less than a mili-Darcy) is notable. To generate the permeability distributions for other layers, the range of permeability in the base model was used. Additionally, the range for varying flowing bottom-hole pressure is 60% to 100% of the litho-static pressure.
An experimental design method was utilized over the properties range to construct combinations of the input parameter values such that the maximum information can be obtained from the minimum number of simulation runs. Latin hypercube sampling (LHS) is the experimental design method in this study. Latin hypercube sampling has enjoyed popularity as a widely used sampling technique for the propagation of uncertainty in analyses of complex systems.67 Using the experimental design method, the range and average of permeability distribution is constrained to the base model. The distribution of permeability changes over different realizations. It is
Figure 5. The permeability distribution for the layers which permeability does not alter through different realizations.
Figure 7. Flow chart to generate different realizations by altering owing bottom-hole pressure at injection wells.
BHP at injection wells). The permeability distributions at different layers for training and validation realizations are shown in Figs 8 and 9. Each row in these figures represents a scenario; training and validation realizations have been marked. Also, each column shows the permeability distribution for a particular layer at different realizations. Figure 10 displays the flowing bottom-hole pressure at injection wells for training and validation realizations. Scenarios 1 to 10 are training realizations, and Scenarios 11 to 13 are validation realizations.
Figure 6. Flow chart to generate different realizations by altering permeability distribution.
assumed the permeability values at the well locations are available (in reality coming from the core data); therefore, using a geo-statistical method (Inverse Distance Estimation provided in CMG-Builder), a distribution of permeability can be generated. Figures 6 and 7 explain the process of generating new realizations (altering permeability distribution and
SRM development training, calibration, and validation of neural networks

In the path to develop the SRM, ANNs should be trained, calibrated and validated. In order to generate ANNs, IDEATM soft ware was used (Fig. 11). IDEATM
Intelligent
Data Evaluation & Analysis (IDEATM ) software is built by Intelligent Solution Inc. (ISI).68
Figure 8. Permeability distributions at layers 1, 2, 19 and 20 for ten training and three blind realizations.
is a soft ware application made for the development of general data driven, intelligent models. Figure 12 shows the inputs used to train the SRM. In addition, Fig. 13 demonstrates the outputs of the SRM in this study. IDEATM provides a random data partitioning algorithm to set the training, calibration and verification shares of the dataset. As mentioned, the spatio-temporal database was built based on the information from ten simulation runs. The training, calibration and verification included 80%, 10%, and 10% of the data in the database, respectively. After training the SRM, its robustness was verified using blind realizations. These runs were not used at any step of training, calibration or verification. Backpropagation was used as the training algorithm. More information on IDEA and building ANNs can be found in ISI.68
Results and discussion

The SRM was trained, calibrated and validated using a few simulation runs. In these realizations, the distributions of permeability (at nine layers) and flowing bottom-hole pressure for injection wells are the variable properties. In order to validate the robustness of the SRM, it was deployed on blind realizations of the reservoir model. The blind cases of reservoir simulation models were not used during the training process of the SRM. In this study, the SRM was trained and validated to reproduce the results of the reservoir simulation model (pressure, water saturation and CO2 mole fraction) at the target layer (layer 18) for different time steps during and after injection of CO2. Layer 18 is the first layer above the injection layers, and it was chosen
Figure 9. Permeability distributions at layers 21, 22, 23, 24 and 25 for ten training and three blind realizations.
Figure 10. Flowing bottom-hole pressure at 45 injection wells for ten training and three blind realizations.
10
Figure 11. Structure of ANNs built in IDEA.
to demonstrate the effect of changing the variable parameters on the pressure and phase saturation behaviors in this layer. The motivations behind this
study originate from the labor- and time-intensive characteristics of reservoir simulation models. A single realization of the reservoir simulation model in this study runs in 424 h (depending on convergence time) on a six processor computer with 24 GB RAM (random access memory). A typical analysis of a CO2 sequestration problem requires hundreds of realizations. On the other hand, a validated SRM (which was prepared using less than 20 realizations) runs in the order of seconds using the same computational power.
Figure 12. Inputs of SRM including static data, dynamic data and operational constraints.
Figure 13. Outputs of SRM.
11
Figure 14. Comparison between the results of simulation model (left) and SRM (right) for pressure distribution of a training realization at layer 18, nine years after injection starts. The gure below represents the relative error.
In addition to the high pace of the SRM, this AI model is able to accurately replicate the results of the reservoir simulation model. The SRM was developed to predict the distribution of pressure, water saturation and gas (CO2) mole fraction at layer 18 for 10 different time steps. In this paper, two time steps
including injection and post injection periods were selected as the representatives of the results. These two time steps consist of one injection and one post injection period. The injection time step is nine years after injection starts; note that total years of injection are 50 years. The second selected time step is in post
12
injection period and shows the results for 100 years after injection ends. For each time step three training realizations and one blind realization were chosen. The accuracy of the SRM to reproduce the results of the simulation model is illustrated in Figs 1437.
Figures 1416 demonstrate the pressure distribution at the target layer (layer 18) during the injection (nine years after injection starts) for three different realizations used to train the SRM. These images show the results of the simulator (left side) compared to the
Figure 17. Comparison between the results of simulation model (left) and SRM (right) for pressure distribution of a blind (validation) realization at layer 18, nine years after injection starts. The gure below represents the relative error.
13
Figure 18. Comparison between the results of simulation model (left) and SRM (right) for pressure distribution of a training realization at layer 18, 100 years after injection ends. The gure below represents the relative error.
SRM (middle). The relative error distribution between the simulator and the SRM is shown along the bottom of the images. The SRM predicts the pressure distribution very well, and the relative error distribution confirms this. There are a few blocks that are out of
the normal range due to numerical problems in the simulation model, causing issues with the pressure behavior. Although the SRM understands the general behavior at these blocks, it does not have a similar performance to the other blocks. The reason goes back
14
to pattern recognition characteristics of the SRM: it cannot learn a pattern that is out of the training range. Figure 17 is the results for the same property and time step (pressure distribution for nine years after injection starts) for a blind (validation) scenario. It is
obvious that the distribution of pressure is different in different realizations, although they are in a similar range. The main reason for such a behavior is altering the permeability distribution at the bottom layers (which are injection layers) for different realizations.
Figure 21. Comparison between the results of simulation model (left) and SRM (right) for pressure distribution of a blind (validation) realization at layer 18, 100 years after injection ends. The gure below represents the relative error.
15
Figure 22. Comparison between the results of simulation model (left) and SRM (right) for water saturation distribution of a training realization at layer 18, nine years after injection starts. The gure below represents the absolute error.
Figures 1820 compare the pressure results for training realizations after 100 years when the injection plan ends, and Fig. 21 displays the results for a blind realization. The general
relative errors for the pressure distribution are less than 10%. The results for the water saturation distribution are shown in Figs 2228. Figures 2224 display the results
16
of the numerical simulator (left) and the SRM (right) for three different realizations used in training, calibration and validation sets. The bottom of these figures shows the absolute error between the simulator
and the SRM outputs. Figure 25 demonstrates the same results and absolute error distributions for one blind realization. These figures (Figs 2225) are the results for nine years after injection starts. Although
Figure 25. Comparison between the results of simulation model (left) and SRM (right) for water saturation distribution of a blind realization at layer 18, nine years after injection starts. The gure below represents the absolute error.
17
Figure 26. Comparison between the results of simulation model (left) and SRM (right) for water saturation distribution of a training realization at layer 18, 100 years after injection ends. The gure below represents the absolute error.
the changes in the water saturation are not as great as the changes in the pressure (CO2 is the injected fluid and water does not tend to move due to low permeability values at this layer), the SRM performs well in these realizations. Figures 26 to 29 show the same
results for the post injection time step (one hundred years after injection ends) as Figs 2225. The general absolute error for this property is less than 3%. Figures 3037 illustrate and compare the results of the simulator and the SRM for the gas (CO2) mole
18
fraction. Figures 3033 describe the results and the absolute errors of training (three realizations) and blind realizations for a time step during the
injection period (nine years after injection starts). Figures 3437 show the same property for a post injection time step (100 after injection). Although the
Figure 29. Comparison between the results of simulation model (left) and SRM (right) for water saturation distribution of a blind (validation) realization at layer 18, 100 years after injection ends. The gure below represents the absolute error.
19
Figure 30. Comparison between the results of simulation model (left) and SRM (right) for gas (CO2) mole fraction distribution of a training realization at layer 18, nine years after injection starts. The gure below represents the absolute error.
general absolute error for the gas mole fraction increases to 10%, the results of the SRM are satisfactory.
Figures 1437 prove the accuracy of the developed SRM in this study. The number of simulation runs required to train the SRM was surprisingly low. When
20
Figure 33. Comparison between the results of simulation model (left) and SRM (right) for gas (CO2) mole fraction distribution of a blind (validation) realization at layer 18, nine years after injection starts. The gure below represents the absolute error.
21
Figure 34. Comparison between the results of simulation model (left) and SRM (right) for gas (CO2) mole fraction distribution of a training realization at layer 18, 100 years after injection ends. The gure below represents the absolute error.
22
Figure 37. Comparison between the results of simulation model (left) and SRM (right) for gas (CO2) mole fraction distribution of a blind (validation) realization at layer 18, 100 years after injection ends. The gure below represents the absolute error.
23
it is compared to the computational power and the time needed for running the simulation model, the SRM shows its efficiency.
References
1. Intergovernmental Panel on Climate Change (IPCC), Climate Change 2007 Mitigation. Cambridge University Press, Cambridge, United Kingdom and New York, USA ( 2007). 2. The Keeling Curve. [Online] . ( 2013 ) Available at: http://sio. ucsd.edu/ [15 May 2013] . 3. International Energy Agency (IEA) . World Energy Outlook 2012. IEA , Paris, France ( 2012 ). 4. Intergovernmental Panel on Climate Change (IPCC), Carbon Dioxide Capture and Storage. Cambridge University Press, Cambridge, United Kingdom and New York, USA ( 2005 ). 5. Cramer TD and Massey JA , Carbon dioxide injection project Sacroc Unit, Scurry County, Texas, in Annual Meeting Papers. Divison of Production, American Petroleum Institute, Houston, Texas, USA (1972 ). 6. Dicharry RM, Perryman TL and Ronquille JD, Evaluation and design of a CO2 miscible ood project-SACROC unit, Kelly-Snyder Field. J Petrol Technol 25 (11):13091318 (1973 ). 7. Cameron JT, Sacroc Carbon Dioxide Injection: A Progress Report, in Annual Meeting Papers. American Petroleum Institute, Los Angelas, California, USA (1976 ). 8. Marchetti C, On geoengineering and the CO2 problem. Climatic Change 1: 59 68 (1977). 9. Baes CF, Beall SE, Lee DW and Marland G, The collection, disposal, and storage of carbon dioxide, in Interactions of Energy and Climate, Springer, The Netherlands, pp. 495 519 (1980 ). 10. Baklid A , Korbol R and Owren G, Sleipner Vest CO2 Disposal, CO2 injection into a shallow underground aquifer, in SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers, Denver, CO, USA (1996 ). 11. Sengul M, CO2 Sequestration A safe transition technology, in SPE International Health, Safety & Environment Conference. Society of Petroleum Engineers, Abu Dhabi, UAE ( 2006 ). 12. Kaarstad O, Experience from real CCS projects And the way forward, in 19th World Petroleum Congress. World Petroleum Congress, Madrid, Spain ( 2008 ). 13. International Energy Agency (IEA) , Energy Technology Perspectiv. IEA , Paris, France ( 2010 ). 14. Benson SM, Monitoring carbon dioxide sequestration in deep geological formations for inventory verication and carbon credits, in SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers, San Antonio, TX, USA ( 2006 ). 15. Emeka Eke P, Naylor M, Haszeldine S, and Curtis A , CO2 leakage prevention technologies, in Offshore Europe. Society of Petroleum Engineers, Aberdeen, UK ( 2011). 16. Pruess K, Xu T, Apps J and Garcia J, Numerical modeling of aquifer disposal of CO2, in SPE/EPA/DOE Exploration and Production Environmental Conference. Society of Petroleum Engineers, San Antonio, TX, USA ( 2001). 17. Krumhansl J, Pawar R, Grigg R, Westrich H, Warpinski N, Zhang D et al., Geological sequestration of carbon dioxide in a depleted oil reservoir, in SPE/DOE Improved Oil Recovery Symposium. Society of Petroleum Engineers, Tulsa, OK, USA ( 2002 ). 18. Cicek O, Compositional and non-isothermal simulation of CO2 sequestration in naturally fractured reservoirs/coalbeds: Development and verication of the model, in SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers, Denver, CO, USA ( 2003 ).
Concluding remarks
The consequences of the daily increasing CO2 concentration in the atmosphere have been shown as a real threat to life on this planet. CCS has showed the potential as a practical method to reduce the amount of CO2 coming from human activities. In order to secure the stability of a CCS project, a comprehensive study of fluid flow through porous media is required. The conventional tools used to perform such an analysis are numerical reservoir simulation models. Although numerical reservoir simulators are able to perform detailed analysis, they are highly time-consuming and computationally expensive. The pattern recognition based reservoir models are efficient alternative tools to address the aforementioned issues. The technology developed and utilized in this study is known as Surrogate Reservoir Models (SRMs). The capabilities of SRMs to be a fast and accurate replica of a reservoir simulation model make them an efficient tool to perform the conventional analyses in the petroleum industry. In this study, ten different realizations of the base model were designed to develop the SRM to predict pressure behavior in the reservoir. Sixteen realizations were considered in order to simulate the phase saturation behavior. The comprehensive spatio-temporal database was developed based on the data extracted from these realizations. The SRM was trained, calibrated and validated using a data driven and intelligent model developer soft ware. The robustness of the SRM was further validated using blind realizations.
Acknowledgements
The authors wish to acknowledge the US Department of Energy (DOE) National Energy Technology Laboratory (NETL) for their support of this project and for providing the base simulation model. They also extend their appreciation to Computer Modeling Group (CMG) and Intelligent Solution Inc. (ISI) for providing the soft ware applications for reservoir simulation and for development of the SRM, respectively.
24
19. Nghiem L , Sammon P, Grabenstetter J and Ohkuma H, Modeling CO2 Storage in Aquifers with a Fully-Coupled Geochemical EOS Compositional Simulator, in SPE/DOE Symposium on Improved Oil Recovery. Society of Petroleum Engineers, Tulsa, OK, USA ( 2004 ). 20. Izgec O, Demiral B and Akin S, Experimental and numerical investigation of carbon sequestration in saline aquifers, in SPE/EPA/DOE Exploration and Production Environmental Conference. Society of Petroleum Engineers, Galveston, TX, USA ( 2005 ). 21. Farokhpoor R, Torster O, Baghbanbashi T, Mork A and Lindeberg EGB. Experimental and numerical simulation of CO2 injection into Upper-Triassic sandstones in Svalbard, Norway, in SPE International Conference on CO2 Capture, Storage, and Utilization. Society of Petroleum Engineers, New Orleans, LA, USA ( 2010 ). 22. PVTi and ECLIPSE 300 Manual. Schlumberger, Geoquest, Houston, TX, USA ( 2005 ). 23. Han WS, Evaluation of CO2 Trapping Mechanisms at the SACROC Northern Platform: Site of 35 Years of CO2 Injection, Doctoral dissertation. The New Mexico Institute of Mining and Technology, Socorro, NM, USA ( 2008 ). 24. Computer Modeling Group, Users Guide GEM, Advanced Compositional Reservoir Simulator, version 2012. CMG, Calgary, Canada ( 2012 ). 25. Mohaghegh SD, Quantifying uncertainties associated with reservoir simulation studies using a surrogate reservoir model, in SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers, San Antonio, TX, USA ( 2006 ). 26. Duda RO, Hart PE and Stork DG, Pattern Classication, 2nd edn. John Wiley & Sons Inc., Hoeboken, NJ, USA ( 2001). 27. Jain AK, Data clustering: 50 years beyond K-Means. Pattern Recog Lett 31(8) : 651 666 ( 2010 ). 28. Mohaghegh SD, Neural network: What it can do for petroleum engineers. J Petrol Technol 47(1) :42 (1995 ). 29. Ali JK, Neural networks: A new tool for the petroleum industry?, in European Petroleum Computer Conference. Society of Petroleum Engineers, Aberdeen, UK (1994 ). 30. Baldwin JL , Otte DN and Whealtley CL , Computer emulation of human mental processes: Application of neural network simulators to problems in well log interpretation, in SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers, San Antonio, TX, USA (1989 ). 31. Masoud N, Neural network knowledge-based modeling of rock properties based on well log databases, in SPE Western Regional Meeting. Society of Petroleum Engineers, Bakerseld, CA, USA (1998 ). 32. Jong-Se L and Jungwhan K, Reservoir porosity and permeability estimation from well logs using fuzzy logic and neural networks, in SPE Asia Pacic Oil and Gas Conference and Exhibition. Society of Petroleum Engineers, Perth, Australia ( 2004 ). 33. Al-Kaabi AU and Lee WJ, An articial neural network approach to identify the well test interpretation model: Applications, in SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers, New Orleans, LA, USA (1990 ). 34. Ershaghi I, Li X, Hassibi M and Shikari Y, A robust neural network model for pattern recognition of pressure transient test data, in SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers, Houston, TX, USA (1993 ).
35. Athichanagorn S and Horne RN, Automatic parameter estimation from well test data using articial neural network, in SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers, Dallas, TX, USA (1995 ). 36. Sultanp MA and Al-Kaabi A , Application of neural network to the determination of well-test interpretation model for horizontal wells, in SPE Asia Pacic Oil and Gas Conference and Exhibition. Society of Petroleum Engineers, Melbourne, Australia ( 2002 ). 37. Mohaghegh SD, Are R, Ameri S and Rose D, Design and development of an articial neural network for estimation of formation permeability. SPE Computer Applications. Society of Petroleum Engineers, Dallas, TX, USA (1995 ). 38. Ahmed T, Link CA , Porter KW, Wideman CJ, Himmer P and Braun J, Application of neural network parameter prediction in reservoir characterization and simulation A case history: The rabbit hills eld, in Latin American and Caribbean Petroleum Engineering Conference. Rio de Janeiro, Brazil ( 30 August 3 September 1997). 39. Singh V, Painuly PK, Srivastava AK, Tiwary DN and Chandra M, Neural networks and their applications in lithostratigraphic interpretation of seismic data for reservoir characterization, in World Petroleum Congress, Madrid, Spain ( 29 June3 July 2008 ). 40. David HJ, Seismic attribute calibration using neural networks, in SEG Annual Meeting. Society of Exploration Geophysicists, Washington, DC, USA (1993 ). 41. Yang HZ and Huang KY, Hybrid neural network for seismic pattern recognition, in SEG Annual Meeting. Society of Exploration Geophysicists, Houston, TX, USA (1991). 42. Roth G and Tarantoia A , Inversion of seismic waveforms using neural networks, in SEG Annual Meeting. Society of Exploration Geophysicists, New Orleans, LA, USA (1992 ). 43. Briones MF, Rojas GA , Moreno JA and Martinez ER , Application of neural networks in the prediction of reservoir hydrocarbon mixture composition from production data , in SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers, New Orleans, LA, USA (1994 ). 44. Gharbi RB and Elsharkawy AM, Neural network model for estimating the PVT properties of middle east crude oils, in Middle East Oil Show and Conference, Bahrain (1518 March 1997). 45. Osman EA , Abdel-Wahhab OA and Al-Marhoun MA , Prediction of oil PVT properties using neural networks, in SPE Middle East Oil Show, Bahrain (1720 March 2001). 46. Oloso MA , Khoukhi A , Abdulraheem A and Elshafei M, Prediction of crude oil viscosity and gas/oil ratio curves using recent advances to neural networks, in SPE/EAGE Reservoir Characterization and Simulation Conference. Abu Dhabi, UAE (1921 October 2009 ). 47. Thomas A and La Pointe PR, Conductive fracture identication using neural networks, in The 35th US Symposium on Rock Mechanics. USRMS, Reno, NV ( 57 June 1995 ). 48. Key SC, Nielsen HH, Signer C, Snneland L , Waagb KH and Veire H, Fault and fracture classication using artical neural networks Case study from the Ekosk Field, in SEG Annual Meeting, Dallas, TX, USA ( 27 November 1997). 49. Sadiq T and Nashawi IS, Using neural networks for prediction of formation fracture gradie, in SPE/CIM International Conference on Horizontal Well Technology. Petroleum Society of Canada, Calgary, Alberta, Canada ( 2000 ).
25
50. Aminzadeh F and deGroot P, A neural networks based seismic object detection technique, in SEG Annual Meeting. Houston, TX, USA ( 611 November 2005 ). 51. Cheng-Dang Z, Wu SL, Mo ZC, Zhu DH and Xu ZV, Direct identication of hydrocarbon from well logs: A neural network interpretation approach, in Annual Technical Meeting. Petroleum Society of Canada, Calgary, Alberta, Canada (1994). 52. Nikravesh M, Kovscek AR, Johnston RM and Patzek TW, Prediction of formation damage during uid injection into fractured, low permeability reservoirs via neural networks, in SPE Formation Damage Control Symposium. Society of Petroleum Engineers, Lafayette, LA, USA (1415 February 1996 ). 53. Kalam MZ, Al-Alawi SM and Al-Mukheini M, Assessment of formation damage using articial neural networks, in SPE Formation Damage Control Symposium. Society of Petroleum Engineers, Lafayette, LA, USA (1415 February 1996 ). 54. Zubarev DI, Pros and cons of applying proxy-models as a substitute for full reservoir simulations, in SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers, New Orleans, LA, USA (47 October 2009 ). 55. Carreras PE and SEaWGT, Tahiti : Development strategy assessment using design of experiments and response surface methods, in SPE Western Regional/AAPG Pacic Section/GSA Cordilleran Section Joint Meeting. Society of Petroleum Engineers, AK, USA ( 810 May 2006 ). 56. Li B, A novel response surface methodology based on Amplitude Factor analysis for modeling nonlinear responses caused by both reservoir and controllable factors, in SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers, Dallas, TX, USA ( 912 October 2005 ). 57. Salhi MA and Van rijen M, Structured uncertainty assessment for Fahud Field through the application of experimental design and response surface methods, in SPE Middle East Oil and Gas Show and Conference. Society of Petroleum Engineers, Kingdom of Bahrain (1215 March 2005 ). 58. Mohaghegh SD, Liu J, Gaskari R, Maysami M and Olukoko O. Application of well-based surrogate reservoir models (SRMs) to two offshore elds in Saudi Arabia, case study, in SPE Western Regional Meeting. Society of Petroleum Engineers, Bakerseld, CA, USA ( 2123 March 2012 ). 59. Mohaghegh SD, Modavi A , Hafez H and Haajizadeh M, Development of surrogate reservoir model (SRM) for fast track analysis of a complex reservoir. Int J Oil Gas Coal Technol 2 (1) : 223 ( 2009 ). 60. Mohaghegh SD, Articial intelligence and data mining: Enabling technology for smart elds. SPE Way Ahead J 5 (3):14 19 ( 2009 ). 61. Mohaghegh SD, Surrogate Reservoir Model. European Geological Union General Assembly, Vienna, Austria ( May 2010 ). 62. Mohaghegh SD, Amini S, Gholami V, Gaskari R and Bromhal G, Grid-based surrogate reservoir modeling (SRM) for fast track analysis of numerical reservoir simulation models at the gridblock level, in SPE Western Regional Meeting. Society of Petroleum Engineers, Bakerseld, CA, USA ( 2123 March 2012 ). 63. Amini S, Mohaghegh SD, Gaskari R and Bromhal G, Uncertainty analysis of a CO2 sequestration project using surrogate reservoir modeling technique, in SPE Western Regional Meeting. Society of Petroleum Engineers, Bakerseld, CA, USA ( 2123 March 2012 ).
64. Mohaghegh SD, Reservoir simulation and modeling based on pattern recognition, in SPE Digital Energy Conference and Exhibition. Society of Petroleum Engineers, Woodlands, TX, USA (1921 April 2011). 65. Dicharry RM, Perryman TL and Ronquille JD, Evaluation and design of a CO2 miscible ood project SACROC unit, Kelly Snider Field. J Petroleum Technol 11(11) :1309 1318 (1973 ). 66. Raines M, Kelly-Snyder (Cisco-Canyon) Fields/SACROC unit, Oil and gas elds in west Texas. West Texas Geological Society, Midland, TX, USA , Vol. 8, pp. 69 78 ( 2005 ). 67. Helton JC and Davis FJ, Latin Hypercube Sampling and the Propagation of Uncertainty in Analyses of Complex Systems, Report No. SAND20010417. Sandia National Laboratories, NM, USA ( 2002 ). 68. Intelligent Solutions Inc. [Online] . ( 2013 ). Available at: http:// www.intelligentsolutionsinc.com/ [May 2013] .
Alireza Shahkarami Alireza Shahkarami is a PhD candidate in Petroleum Engineering at West Virginia University (WVU). Alirezas research mainly focuses on the implementation of articial intelligence and data mining techniques to address conventional and unconventional problems in petroleum engineering.
Shahab D. Mohaghegh Shahab D. Mohaghegh is the president and CEO of Intelligent Solutions, Inc. (ISI) and Professor of Petroleum and Natural Gas Engineering at West Virginia University. A pioneer in the application of articial intelligence and data mining in the exploration and production industry, he holds BSc, MSc, and PhD degrees in Petroleum and Natural Gas Engineering.
Vida Gholami Vida Gholami is a Research Associate in the PEARL (Petroleum Engineering and Analytics Research Laboratory) at West Virginia University (WVU). For the past six years she had been working on the application of articial intelligence and data mining (AI&DM) in the petroleum industry.
26
Alireza Haghighat Alireza Haghighat earned his MSc in Petroleum Engineering from Delft University of Technology in Netherlands. He is currently pursuing his PhD in Petroleum Engineering at West Virginia University. Alirezas research focuses on application of AI and smart eld technology in CO2 sequestration.
Daniel Moreno Daniel Moreno earned his BSc in Mechanical Engineering from the Metropolitan University in Caracas, Venezuela in 2004. He holds an MSc in Petroleum and Natural Gas Engineering from West Virginia University. Presently, he works as a Production Engineer at Chevron in Bakerseld, CA.
27

Surrogate Reservoir Model

Încărcat de

Informații document

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Surrogate Reservoir Model

Încărcat de

Drepturi de autor:

Formate disponibile

Modeling and Analysis

Potentials of pattern recognition techniques

Surrogate reservoir models

Reservoir simulation model

Uncertain properties and training realizations

Figure 1. A three dimensional view of simulation model.

SRM development training, calibration, and validation of neural networks

Results and discussion

Figure 11. Structure of ANNs built in IDEA.

Figure 13. Outputs of SRM.

S-ar putea să vă placă și