Full ﬁeld reservoir modeling of shale assets using advanced data-driven analytics
Soodabeh Esmailia, Shahab D. Mohagheghb,*
aAsset Development Team, North Operation, California Resources Corporation, California 90024, USA
bWest Virginia University, 345-E Mineral Resources Bldg., P. O. Box 6070, Morgantown, WV 26506, USA
Hydrocarbon production from shale has attracted much attention in the recent years. When applied to this proliﬁc and hydrocarbon rich resource plays, our understanding of the complexities of the ﬂow mechanism (sorption process and ﬂow behavior in complex fracture systems-induced or natural) leaves much to be desired. In this paper, we present and discuss a novel approach to modeling, history matching of hydrocarbon production from a Marcellus shale asset in southwestern Pennsylvania using advanced data mining, pattern recognition and machine learning technologies. In this new approach instead of imposing our understanding of the ﬂow mechanism, the impact of multi-stage hydraulic fractures, and the production process on the reservoir model, we allow the production history, well log, completion and hydraulic fracturing data to guide our model and determine its behavior. The uniqueness of this technology is that it incorporates the so-called “hard data” directly into the reservoir model, so that the model can be used to optimize the hydraulic fracture process.
The “hard data” refers to ﬁeld measurements during the hydraulic fracturing process such as ﬂuid and proppant type and amount, injection pressure and rate as well as proppant concentration. This novel approach contrasts with the current industry focus on the use of “soft data” (non-measured, interpretive data such as frac length, width, height and conductivity) in the reservoir models. The study focuses on a Marcellus shale asset that includes 135 wells with multiple pads, different landing targets, well length and reservoir properties. The full field history matching process was successfully completed using this data driven approach thus capturing the production behavior with acceptable accuracy for individual wells and for the entire asset.
Much of the success in turning the shale source rock into an economically viable and producible hydrocarbon reservoir is accredited to George Mitchell and his team of geologists and engineers at Mitchell Energy & Development1. The success in production of shale oil and shale gas dates back to 1981 when multiple combinations of processes and technologies where examined before ultimately succeeding in 1997 with the use of a “slick-water” frac that made production from Barnett Shale economical and changed the future of the US natural gas industry (NGW, 2011). Today horizontal wells that include multi-stage, multi-cluster hydraulic fractures and pad drilling are the norm in developing shale oil and shale gas assets in North America and expanding throughout the world.
Shale reservoirs are characterized by extremely low permeability rocks that have a number of unique attributes, including high organic content, high clay content, extremely ﬁne grain size, plate-like micro-porosity, little to no macro-porosity, and coupled Darcy and Fickian ﬂow through the rock matrix. Unlike conventional and even tight sandstone gas reservoirs where all the gas is in the free state in the pore space, the gas in shale is stored by compression (as free gas) and by adsorption on the surfaces of the solid material, either organic matter or minerals (Guo et al., 2012).
This combination of traits has led to the evolution of hydraulic fracture stimulation involving high rates, low-viscosities, and large volumes of proppant. The stimulation design for plays such as Marcellus Shale is drastically different than anything else that has been performed in the past. It takes large amounts of space, materials, and equipment to treat the Marcellus Shale to its fullest potential (Houston et al., 2009). Currently, the Marcellus shale, covering a large area in the northeastern US, is one of the most sought-after shale-gas resource play in the United States. It has presumably the largest shale-gas deposit in the world, having a potentially prospective area of 44,000 square miles, containing about 500 TCF of recoverable gas (Engelder, 2009).
Figure 1. Data available in the dataset that include location and trajectory, reservoir characteristics, completion, hydraulic fracturing and production details.
This geological formation was known for decades to contain signiﬁcant amounts of natural gas but was never considered economical. Uneconomic resources, however, are often transformed into marketable assets by technological progress (Considine et al., 2009). Advances in horizontal drilling and multi-stage hydraulic fracturing have made the Marcellus shale reservoir a focal point for many operators. Nevertheless, our understanding of the complexities associated with the ﬂow mechanism in the natural fracture and its coupling with the matrix and the induced fractures, impact of geomechanical properties and optimum design of hydraulic fractures is still a work in progress.
A vibrant and fast-growing literature that covers operational and technological challenges of production from shale oil and shale gas is currently thriving.
The research includes all aspects of drilling, completion, and production as well as difﬁculties in formation evaluation/characterization, in modeling macro- and micro-scales of ﬂuid transport, and in developing reliable reservoir simulators. Understanding reservoir properties like lithology, porosity, organic carbon, water saturation and mechanical properties of the rock, which includes stresses, and planning completions based on that knowledge is the key to production optimization. Therefore, the ﬁnal objective is to increase our ability to integrate laboratory and petrophysical measurements with geochemical, geological, petrologic, and geomechanical knowledge, to develop a more solid understanding of shale plays and to provide better assessments, better predictions, and better models.
Figure 2. Marcellus shale AI-based Full-ﬁeld history matching process.
Reservoir simulation has played an important role towards achieving the above mentioned stated goals (Mohaghegh, 2013a). However there are still many challenges to overcome before reaching the stated goals. Firstly, the physics of ﬂuid ﬂow in shale rocks haven’t been fully understood, and are undergoing continuous development as the industry learns more (Lee and Sidle, 2010). Secondly, full reservoir simulation is resource intensive and time consuming. Thirdly, challenges encountered when applying conventional reservoir simulation to shale resources (Mohaghegh, 2013b) could be solved with pattern recognition technologies (Mohaghegh, 2000a,b,c).
In this paper, we developed an Artiﬁcial Intelligence-based model that is conditioned to all available ﬁeld measurements (e.g. production history, measured reservoir characterizations including geomechanical and geochemical properties) as well as measured hydraulic fracturing variables like slurry volume, proppant amount and sizes, injection rate etc. Such model has the potential to provide operators with an alternative to history-match, predict and assess reserves in oil and gas producing shale reservoirs. The pattern recognition approach not only has a much faster turnaround time compared to numerical simulation techniques, but also offers reasonable accuracy while incorporating all available data compared to analytical and numerical techniques that are very selective in the type of ﬁeld measurements that they use. The integrated framework presented in this paper enables reservoir engineers to compare and contrast multiple scenarios and propose ﬁeld development strategies.
Figure 3. Three well types from a single pad.
2. Top-Down modeling - pattern recognition based reservoir modeling
Artiﬁcial intelligence and data mining refers to a collection of tools and techniques that provide the means for ﬁnding patterns among non-linear and interdependent parameters involved in the shale oil and shale gas development process. Interest in the research of pattern recognition applications has spawned in recent years. Popular areas include: data mining (identiﬁcation of a ’pattern’, i.e., a correlation, or an outlier in millions of multidimensional patterns), document classiﬁcation (efﬁcient search of text documents), ﬁnancial forecasting, and biometrics.
Top-Down modeling, a recently developed data-driven reservoir modeling technology (ISI, 2014), is deﬁned as a formalized, comprehensive, multi-variant, full-ﬁeld, and empirical reservoir model, which takes into account all aspects of production from shale including reservoir characterization, completion and hydraulic fracturing parameters as well as production characteristics. Despite the common practice in shale modeling using a conventional approach, which is usually done at the well level (Strickland et al., 2011), this technique is capable of performing history matching for all individual wells in addition to full ﬁeld by taking into account the effect of offset wells.
Figure 4. Log-log plot of production rate as a function of time for one of 135 wells in the asset studied for this paper.
There are major steps in the development of a Top-Down shale reservoir model that is enumerated as follows:
- Spatio-temporal database development; the ﬁrst step in developing a data driven shale reservoir model is preparing a representative spatio-temporal database (data acquisition and preprocessing). The extent at which this spatio-temporal database actually represents the ﬂuid ﬂow behavior of the reservoir that is being modeled, determines the potential degree of success in developing a successful model. The nature and class of the AI-based shale reservoir model is determined by the source of this database. The term spatio-temporal deﬁnes the essence of this database and is inspired from the physics that controls this phenomenon (Mohaghegh, 2011). An extensive data mining and analysis process should be conducted at this step to fully understand the data that is housed in this database. The data compilation, curation, quality control and preprocessing is one of the most important and time consuming steps in developing an AI-based reservoir model.
- Simultaneous training and history matching of the reservoir model; in conventional numerical reservoir simulation the base model will be modiﬁed to match production history, while AI-based reservoir modeling starts with the static model and tries to honor it and not modify it during the history matching process. Instead, we will analyze and quantify the uncertainties associated with this static model at a later stage in the development. The model development and history matching in AI-based shale reservoir model are performed simultaneously during the training process. The main objective is to make sure that the AI-based shale reservoir model learns ﬂuid ﬂow behavior in the shale reservoir being modeled. The spatio-temporal database developed in the previous step is the main source of information for building and history matching the AI-based reservoir model.
Figure 5. Inside and closest outside distance.
- Simultaneous training and history matching of the reservoir model; in conventional numerical reservoir simulation the base model will be modiﬁed to match production history, while AI-based reservoir modeling starts with the static model and tries to honor it and not modify it during the history matching process. Instead, we will analyze and quantify the uncertainties associated with this static model at a later stage in the development. The model development and history matching in AI-based shale reservoir model are performed simultaneously during the training process.
The main objective is to make sure that the AI-based shale reservoir model learns ﬂuid ﬂow behavior in the shale reservoir being modeled. The spatio-temporal database developed in the previous step is the main source of information for building and history matching the AI-based reservoir model. In this work, an ensemble of multilayer neural networks is used (Haykin, 1999). These neural networks are appropriate for pattern recognition purposes in case of dealing with non-linear cases. The neural network consists of one hidden layer with different number of hidden neurons, which have been optimized based on the number of data records and the number of inputs in training, calibration and veriﬁcation process (Mohaghegh, 2000a). It is extremely important to have a clear and robust strategy for validating the predictive capability of the AI-based reservoir model. The model must be validated using completely blind data that has not been used, in any shape or form, during the development. Both training and calibration datasets that are used during the initial training and history matching of the model are considered non- blind. As noted by Mohaghegh (2011), some may argue that the calibration - also known as testing dataset -is also blind. This argument has some merits but if used during the development of the AI-based shale reservoir model can compromise validity and predictability of the model and therefore such practices are not recommended.
Figure 6. History matching result for entire ﬁeld by using the maximum possible combination of parameters.
- Sensitivity analysis and quantiﬁcation of uncertainties; during the model development and history matching that was deﬁned above, the static model is not modiﬁed. Lack of such modiﬁcations may present a weakness of this technology, knowing the fact that the static model includes inherent uncertainties. To address this, the AI-based reservoir modeling workﬂow includes a comprehensive set of sensitivity and uncertainty analyses. During this step, the developed and history matched model is thoroughly examined against a wide range of changes in reservoir characteristics and/or operational constraints. The changes in pressure or production rate at each well are examined against potential modiﬁcation of any and all the parameters that have been involved in the modeling process. These sensitivity and uncertainty analyses include single- and combinatorial-parameter sensitivity analyses, quantiﬁcation of uncertainties using Monte Carlo simulation methods and ﬁnally development of type curves. All these analyses can be performed on individual wells, groups of wells or for the entire asset.
- Deployment of the model in predictive mode; similar to any other reservoir simulation model, the trained, history matched and validated AI-based shale reservoir model is deployed in predictive mode in order to be used for reservoir management and decision making purposes.
Figure 7. History matching result for entire ﬁeld in optimum history matched model.
3. AI-based reservoir modeling in Marcellus shale
The study presented in this manuscript focuses on part of Marcellus shale that includes 135 horizontal wells within more than 40 pads. These horizontal wells have different landing targets, well lengths and reservoir properties. During the development of the Top-Down shale model all available data including static, dynamic, completion, hydraulic fracturing, operational constraint, etc. have been used for training, calibration, and validation of the model. A complete list of inputs that are included in main data set for development of the base model is shown in Fig. 1.
The data set includes more than 1200 hydraulic fracturing stages (approximately 3700 clusters of hydraulic fracturing). Some wells have up to 17 stages of hydraulic fracturing while others have been fractured with as few as four stages. The perforated lateral length ranges from 1400 to 5600 ft. The total injected proppant in these wells ranges from a minimum of about 97,000 lbs up to a maximum of about 8,500,000 lbs and total slurry volume of about 40,000 bbls to 181,000 bbls.
Figure 8. Data-driven full ﬁeld reservoir model training, calibration, and veriﬁcation cross plots.
The wells are completed in both upper and lower Marcellus. The porosity of upper Marcellus varies from 5 to 10% while its gross thickness is measured to be between 43 and 114 ft. The total organic carbon content (TOC) of the upper Marcellus in this area is between 0.8 and 1.7%. The reservoir characteristics of lower Marcellus include porosity in the range of 8e14%, gross thickness between 60 and 120ft, and TOC of 2-6%.
4. Results and discussion
During the training and history matching process using AI- based modeling approach inclusion and exclusion of multiple parameters were examined in order to determine their impact on model behavior. Fig. 2 includes a ﬂowchart that shows the evolution process of developing the AI-based Marcellus shale full ﬁeld reservoir model. It starts from the base model (where most of the parameters are included as our ﬁrst shot) to converge to the best history matched model where optimum number of inputs are identiﬁed.
Figure 9. List of the inputs in optimum history matched model.
4.1. Impact of different input parameters
Base Model- As illustrated in Fig. 2, the base model was built by incorporating all available data that are listed in Fig. 1. This model consists of all ﬁeld measurements including well locations, trajectories, static data, completion, hydraulic fracturing data, production rates, and operational constraints.
Effect of Offset Well- In order to consider the effect of offset wells and taking into account for any well interference, all afore-mentioned properties for closest offset well were included in the modeling.
Effect of Different Well Types- Since drilling multiple wells from a pad is a common practice in the most shale assets and given the fact that the horizontal wells drilled from a pad experience different interaction with their offsets, three types of laterals have been deﬁned. Fig. 3 shows the conﬁguration of the three types of wells drilled from a single pad. Based on this deﬁnition a new parameter was added to the dataset marked as “Well-Type.” This parameter was assigned values such as 1, 2, or 3 in order to incorporate the “Well-Type” information.
Figure 10. The best (top) and worst (bottom) history matched wells in the optimum history matched model.
Following is a brief description of the three “Well-Types” used in the data-driven full ﬁeld model.
- Type one Lateral: This type of lateral has no neighboring laterals and does not share drainage area. It does not experience any “Frac Hits2” from wells in the same pad (it might experience Frac-Hit from a lateral from an offset pad) and its reach will be as far as its hydraulic fractures.
- Type two Lateral: The type two lateral has only one neigh-boring lateral and therefore, it shares part of the drainage area and “Frac Hits” are possible from laterals in the same pad (it might experience Frac-Hit from a lateral from an offset pad).
- Type three Lateral: The type three lateral is bounded by two neighboring laterals thus, the drainage area will be shared and “Frac Hits” are possible from both sides in the same pad. If a type three lateral experiences a Frac-Hit from an offset pad, it will be from a different depth.
Effect of Different Flow Regimes- There are usually two distinct ﬂow regimes that can be observed in all the wells. The ﬁrst ﬂow regime corresponds to the production of the initial free gas in the fracture/pore spaces, which is immediately available for production and it may last a few days to a few months (Flow regimetype one). Most of the wells have been observed to exhibit transient linear behavior as the main ﬂow regime (Flow regime type two).
Figure 11. History matching error histogram for optimum history matched model.
This transient linear behavior is characterized by a one-half slope on a log-log plot of rate against time. The transient linear ﬂow regime is expected to be caused by transient drainage of low-permeability matrix blocks into adjoining fractures (Bello and Wattenbarger, 2010). These two ﬂow regimes where introduced in the data-driven full ﬁeld model as dynamic property. Fig. 4 shows the two ﬂow regimes for one of the wells from the asset described in this case study.
Effect of Distances Between Laterals- In order to consider the impact of the location (distance from other laterals in the same pad and closest lateral from an offset pad), two distances were deﬁned and used in the data-driven full ﬁeld model: (1) distance between laterals of the same pad, (2) distance to closest lateral of a different pad. This concept is illustrated in Fig. 5.
Figure 12. Examples of history matching results-excellent matches.
As mentioned earlier, the base model was developed with maximum number of inputs that were available. The results of the history matching of the base model are shown in Fig. 6. This ﬁgure shows the history matching results for the entire ﬁeld. It must be noted that in the data-driven full ﬁeld reservoir model the history match is performed on a well by well basis. In order to generate the history match for the entire asset, as shown in Fig. 6, productions from individual wells are summed (the ﬁeld measurements as well as the Top-Down model) and are plotted.
The top plot in this ﬁgure is the history match (monthly production, the left y axis and the cumulative production, the right y axis). The bottom bar chart shows the well count (number of producing wells). In the top plot, the orange dots represent the actual monthly rate (normalized to protect conﬁdentiality of data) for the entire ﬁeld while the green solid line shows the AI-based model results. The orange area represents the actual cumulative production (normalized to protect conﬁdentiality of data) while the green area corresponds to cumulative production generated by the AI-based model.
4.2. Optimum history matched model
Although, the history matching results drived by using the maximum combination of parameters (Fig. 6) have reasonable accuracy, it is preferable to reduce the number of input variables in data-driven models. As the number of inputs in a data-driven model increases, a certain level control over the model behavior may be compromised. Accordingly, a history-matching process was derived with minimum combination of parameters that can/should be used to achieve an acceptable history match results for individual wells and for the entire ﬁeld (the total number of inputs decreased from 103 to 38).
Figure 13. Examples of history matching results-good matches.
The inputs that were removed and the reason for removing these are as follows:
- The upper and lower Marcellus rock properties are drastically different so averaging them destroys a major geologic factor that affects both the fracing and the resulting the well performance. The lack of allocated production should not pose a problem and in fact the resulting AI model could be used to estimate the allocation of each zone and checking that result against the know reservoir properties. In most areas, you are supposed to see lower production from the upper Marcellus. This again could be used as an additional validation of the AI-based model.
- The perforated lateral length and total stimulated lateral length were included in the data set. Since the length of stimulated lateral is always 100 ft longer than the length of perforated lateral, therefore the total perforated lateral length was removed from the model.
- Instead of including stage based hydraulic fracturing data, the total values for slurry volume, proppant amount, etc. were used in the optimized case. In addition, the average injection rate and pressure that were not changing considerably were removed.
- Since the inside and closest outside distances from an offset were included for each individual well, therefore there was no need to include these two distances for offset well consequently they were removed from the input data.
The ﬁnal history match result for the optimized model was improved and showed acceptable match of monthly gas rate and the cumulative production for the entire ﬁeld (Fig. 7). For this case, 80% of the data was used for training and 20% for calibration and veriﬁcation (10% for each). Fig. 8 shows the cross plot of the training, calibration, and veriﬁcation of the data-driven full ﬁeld reservoir model, which shows R2 of 0.99, 0.97 and 0.97 for training, calibration and veriﬁcation, respectively. In this ﬁgure, the x-axis is the predicated monthly gas rate by the model while the y-axis is the actual gas production rate from the ﬁeld. Fig. 9 shows the list of inputs that were used in optimum history matched model. Fig. 10 shows two wells with the best and the worst history matching results in the optimum history matched model. This ﬁgure shows that the erratic behavior displayed by the well in the graph on the right could not be captured by the data-driven full ﬁeld reservoir model, even though the trend was followed.
Figure 14. Examples of history matching results-average matches.
4.3. Error Calculation
The error percentage of monthly gas production rate for all the 135 wells was calculated using the following equation:
where, YitTDM is the predicted production by TDM (AI-based model);
- Yi;t is the actual ﬁeld data;
- ΔYim is the measured maximum change in actual production data;
- Nt(i) is the number of month of production.
Fig. 11 shows the histogram of error for the optimum history matched model. In this model, 101 wells were matched with less than 10% error (excellent), 22 wells had errors between 10 and 20% (good), 6 wells had errors between 20 and 30% (average) and 6 wells had errors of more than 40% (poor). Several example of excellent, good, average and poor history matching results are illustrated in Figs. 12-15.
In this paper, development and results of a data-driven, full ﬁeld Marcellus shale reservoir model was discussed with the aim of overcoming the current issues associated with numerical simulation and modeling of shale gas reservoirs. The advantage of this technology is its capability of handling and incorporating hard data instead of a rigid representation of ﬂow and transport mechanisms in shale reservoirs. When dealing with complex non-linear systems such as ﬂow in shale reservoirs, the available hard data could identify its functional relationship using pattern recognition.
The full-ﬁeld history matching was performed with acceptable accuracy. This model can be used for Marcellus shale wells and reservoir performance prediction and ﬁeld development.
Figure 15. Examples of history matching resultsepoor matches.
The authors would like to thank RPSEA and U.S. Department of Energy for partially funding this study. Authors would like to thank Range Resources Appalachia LLC for providing the data. We also thank and acknowledge Intelligent Solutions Inc. for providing IMagine and IDEA software packages for the development of the Top-Down models.
- Bello, R.O., Wattenbarger, R.A., December 2010. Modeling and analysis of shale gas production with a skin effect. Journal of Canadian Petroleum Technology 49 (12), 37-48.
- Considine, T., Watson, R., Entler, R., Sparks, J., 2009. An Emerging Giant: Prospects and Economic Impacts of Developing the Marcellus Shale Natural Gas Play. State College: EME Department, PSU.
- Engelder, T., 2009. Marcellus 2008: Report Card on the Breakout Year for Gas Production in the Appalachian Basin. Fort worth Basin Oil and Gas.
- Guo, J., Zhang, L., Wang, H., Feng, G., 2012. Pressure transient analysis for multistage fractured horizontal wells in shale gas reservoirs. Transport in Porous Media 93, 635e653. http://dx.doi.org/10.1007/s11242-012-9973-4.
- Houston, N., Blauch, M., Weaver, D., Miller, D., O’Hara, D., 23e25 September 2009. Fracture-stimulation in the Marcellus Shale-lessons Learned in Fluid Selection and Execution. SPE 125987, SPE Eastern Regional Meeting, Charleston, West Virginia, USA.
- Haykin, S., 1999. Neural networks: a comprehensive foundation, second ed. Prentice Hall, Upper Saddle Rever, New Jersey.
- ISI, 2014. Intelligent Solutions, Inc. www.IntelligentSolutionsInc.com.
- Lee, J., Sidle, R., 2010. Texas A&M “Gas-Reserves Estimation in Resource Plays”. SPE Economics & Management Journal 2 (2).
- Mohaghegh, 2000a. Virtual intelligence applications in petroleum engineering: part 1; artiﬁcial neural networks. Journal of Petroleum Technology, Distinguished Author Series, September 2000, 64-73.
- Mohaghegh, S.D., 2000b. Virtual intelligence applications in petroleum engineering: part 2; evolutionary computing. Journal of Petroleum Technology, Distinguished Author Series, October 2000, 40-46.
- Mohaghegh, S.D., 2000c. Virtual intelligence applications in Petroleum engineering: part 3; fuzzy logic. Journal of Petroleum Technology, Distinguished Author Series, November 2000, 82-87.
- Mohaghegh, S.D., 2011. Reservoir simulation and modeling based on artiﬁcial intelligence and data mining (AI&DM). Journal of Natural Gas Science and Engineering 3, 697-705.
- Mohaghegh, S.D., 2013a. Reservoir modeling of shale formations. Journal of Natural Gas Science and Engineering 12, 22-33.
- Mohaghegh, S.D., 2013b. A Critical View of Current State of Reservoir Modeling of Shale Assets. SPE 165713-SPE Eastern Regional Conference and Exhibition. Pittsburgh, PA. 20-22 August 2013.
- NGW, 2011. Genesis: the North Texas Barnett Shale Opens New Energy Era, North American Shale Revolution. www.energyintel.com. Natural Gas Week, Special report Sept. e Dec. 2011.
- Strickland, R., Purvis, D., Blasingame, T., 14e16 June, 2011. Practical Aspects of Re- serves Determinations for Shale Gas. SPE-144357, North American Unconventional Gas Conference and Exhibition, The Woodlands, Texas.
Received 19 July 2014, Received in revised form 26 November 2014, Accepted 6 December 2014, Available online 23 January 2015.
E-mail address: [email protected] (S.D. Mohaghegh).
Peer-review under responsibility of China University of Geosciences (Beijing).
1Mitchell Energy & Development. Mitchell sold his company to Devon Energy in 2002 in a deal worth $3.5 Billion.
2Frac Hit is a phenomena that is encountered when producing shale assets. It happens when production from a given well is disrupted as a function of hydraulic fracturing activities in an offset well. In a Frac Hit the water used during the hydraulic fracturing of an offset shows up and sometimes even completely shuts-in the producing well.
1674-9871/ 2015, China University of Geosciences (Beijing) and Peking University. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC- ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).