HYDRAULIC FRACTURING, COMPLETION, AND DIVERTER OPTIMIZATION METHOD FOR KNOWN WELL ROCK PROPERTIES

Abstract

A method for optimizing hydraulic fracturing includes characterizing a fracture induced by pumping fracturing fluid into a subsurface formation. The characterizing includes analyzing properties of reflected tube waves detected in a well. Change in expected characterization of the subsurface formation is modeled with respect to a modeled change in at least one parameter of the pumping fracturing fluid. The modeled change is compared to a measured change in the characterization with respect to an actual change in the at least one parameter. The modeled change and the measured change are used to train a machine learning algorithm to determine an optimized change in the at least one parameter.

Description

CROSS REFERENCE TO RELATED APPLICATIONS

Continuation of International Application No. PCT/US2020/042825 filed on Jul. 20, 2020. Priority is claimed from U.S. Provisional Application No. 62/876,613 filed on Jul. 20, 2019. Both the foregoing applications are incorporated herein by reference in their entirety.

STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

Not Applicable.

NAMES OF THE PARTIES TO A JOINT RESEARCH AGREEMENT

Not Applicable.

BACKGROUND

This disclosure generally relates to multi-stage hydraulic fracturing and well completions. More particularly, the present disclosure relates to techniques for hydraulic fracturing treatment planning and optimization for oil and gas producing wells.

Multi-stage hydraulic fracturing is a well stimulation and completion technique to improve the productivity of the well by enhancing the connectivity of the wellbore to the adjacent fluid reservoir. Hydraulic fracturing is performed by high-pressure injection of fracturing fluids into the wellbore to create fractures within the reservoir rock formation. The composition of the fracture fluid is primarily water mixed with sand and/or other proppant. The fracture fluid may comprise other solutes such as chemical additives, polymers, acids or solids such as quartz or other sized particulates. The effectiveness of the hydraulic fracturing operation affects the hydraulic conductivity of the induced fracture network where natural gas, oil, and water flow into the well from the reservoir rock. The hydraulic conductivity of the fracture network in turn affects the ultimate hydrocarbon production and therefore profit. Service companies usually become more skilled at formulating effective fracture treatment parameters with each well or stage completed, yet achieving the optimal parameters for hydraulic fracturing remains to a considerable extent a practice of trial and error.

Completion engineers use relatively complex fracture models to design fracture treatment parameters, such as fracturing fluid composition and volume, pumping rate and pressure, casing/liner perforation types, perforation cluster type, and perforation cluster spacing. In addition to the foregoing parameters, some other treatment additives such as acid, gels, diverters, or breakers may also be used in hydraulic fracturing. Generally, water and sand are relatively inexpensive, while other additives and diverters may be more expensive and thus their use may be limited and/or closely monitored. The fracture treatment design intends to obtain a relevant type of fracture for the formation characteristics to optimize production and hydrocarbon recovery. For example, the completion engineer may prefer long and uniform fractures rather than a dense and interconnected fracture network emanating from the wellbore to cover a reservoir area of interest to be drained and produced. The desired length of the fractures may be affected by the cost/estimated production ratio.

SUMMARY

One aspect of the present disclosure relates to a method for optimizing hydraulic fracturing. A method according to this aspect of the disclosure includes characterizing a fracture induced by pumping fracturing fluid into a subsurface formation. The characterizing includes analyzing properties of reflected tube waves detected in a well. Change in expected characterization of the subsurface formation is modeled with respect to a modeled change in at least one parameter of the pumping fracturing fluid. The modeled change is compared to a measured change in the characterization with respect to an actual change in the at least one parameter. The modeled change and the measured change are used to train a machine learning algorithm to determine an optimized change in the at least one parameter.

In some embodiments, the at least one parameter comprises at least one of pad volume, fracture fluid viscosity, fracture fluid density, propping agent type, propping agent volume, fracturing fluid Injection rate and fracture fluid volume.

In some embodiments, the characterizing a fracture comprises analyzing a far field fracture system property and a near field fracture system property.

In some embodiments, the analyzing the far field fracture property comprises analyzing pressure decay after completing pumping the fracturing fluid.

In some embodiments, the analyzing the near field fracture property comprises analyzing at least amplitude and arrival time of the reflected tube waves.

In some embodiments, the tube waves are induced by inducing a pressure change in the well.

In some embodiments, the machine learning algorithm comprises a recursive feature elimination.

In some embodiments, the recursive feature elimination comprises a ridge regression model for each of a plurality of dimensions of hydraulic fractures.

In some embodiments, the plurality of dimensions comprises at least one of, fracture length, fracture height, near field fracture conductivity and far field fracture conductivity.

In some embodiments, the recursive feature elimination comprises a random forest model for each of a plurality of parameters of a fracture treatment.

In some embodiments, the plurality of fracture treatment parameters comprises at least one of fluid pumping rate, fluid pumping pressure, proppant type, proppant size, proppant size distribution and stage length of a fracture treatment state.

In some embodiments, the at least one parameter comprises at least one of diverter type and diverter amount.

A non-transitory computer readable medium according to another aspect of this disclosure includes logic operable to cause a programmable computer to perform actions comprising: characterizing a fracture induced by pumping fracturing fluid into a subsurface formation, the characterizing comprising analyzing properties of reflected tube waves detected in a well; modeling change in expected characterization of the subsurface formation with respect to a modeled change in at least one parameter of the pumping fracturing fluid; comparing the modeled change to a measured change in the characterization with respect to an actual changes in the at least one parameter; and using the modeled change and the measured change to train a machine learning algorithm to determine an optimized change in the at least one parameter.

In some embodiments, the at least one parameter comprises at least one of pad volume, fracture fluid viscosity, fracture fluid density, propping agent type, propping agent volume, fracturing fluid Injection rate and fracture fluid volume.

In some embodiments, the characterizing a fracture comprises analyzing a far field fracture system property and a near field fracture system property.

In some embodiments, the analyzing the far field fracture property comprises analyzing pressure decay after completing pumping the fracturing fluid.

In some embodiments, the analyzing the near field fracture property comprises analyzing at least amplitude and arrival time of the reflected tube waves.

In some embodiments, the tube waves are induced by inducing a pressure change in the well.

In some embodiments, the machine learning algorithm comprises a recursive feature elimination.

In some embodiments, the recursive feature elimination comprises a ridge regression model for each of a plurality of dimensions of hydraulic fractures.

In some embodiments, the plurality of dimensions comprises at least one of, fracture length, fracture height, near field fracture conductivity and far field fracture conductivity.

In some embodiments, the recursive feature elimination comprises a random forest model for each of a plurality of parameters of a fracture treatment.

In some embodiments, the plurality of fracture treatment parameters comprises at least one of fluid pumping rate, fluid pumping pressure, proppant type, proppant size, proppant size distribution and stage length of a fracture treatment state.

In some embodiments, the at least one parameter comprises at least one of diverter type and diverter amount.

Other aspects and possible advantages will be apparent from the description and claims that follow.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flow chart for an example embodiment of a method described in this disclosure to improve hydraulic fracture treatment planning.

FIG. 2 shows a sample of a horizontal well through the formation in 3D to which this method was applied.

FIG. 3 shows a possible wellbore data acquisition arrangement corresponding with the completion operation setup.

FIG. 4 shows fracture systems correlated with the production data from Organic Tracer Data (OFT).

FIG. 5 is a flow chart of an example method workflow used to improve diverter use.

FIG. 6 shows the effect of diverter pumped with different volumes with respect to NF and FF conductivity.

FIG. 7 contains four different designs analysis across the lateral to identify the effect of methods according to this disclosure on the development of the fracture system and diverter use.

FIG. 8 shows the geomechanical properties of an Austin Chalk well as derived by open hole acoustic well logs.

FIG. 9 shows the measured types of fracture systems created due to the variation of rock properties as described in FIG. 8.

FIG. 10 shows the geomechanical properties of an Austin Chalk well as derived by open hole acoustic well logs superimposed over measured completion results.

FIG. 11 shows a summary of real time, progressing fracture optimization at the field using rock geomechanical properties and fracture measurements of a Lower Eagle ford formation well.

FIG. 12 shows change in pad (in bbls) design on a stage to account for formation properties.

FIG. 13 shows completion-engineer driven optimization of treatment based on various formation properties.

FIG. 14 shows an example computer system that can be used to implement machine learning and to automate the present invention

FIG. 15 shows an example embodiment of recursive feature elimination model.

FIG. 16 shows an example embodiment of real-time hydraulic fracturing design optimization process for a single well.

FIG. 17 shows an example of real-time hydraulic fracturing design optimization for a zipper well.

FIG. 18 shows an example of a real-time hydraulic fracturing design optimization process using Artificial intelligence models with Recursive Feature Elimination (RFE).

FIG. 19 shows an example of a real-time hydraulic fracturing design optimization process using Artificial intelligence models with Recursive Feature Elimination (RFE) for various mechanical specific energy (MSE) ranges (top right inserts).

FIG. 20 shows an optimization output curve for the amount of proppant pumped to achieve maximum half-length under given conditions.

DETAILED DESCRIPTION

An objective of methods according to this disclosure is to increase the precision of multi-stage hydraulic fracturing to create the desired hydraulic fracturing results. At the same time, such methods can substantially decrease the cost of fracture treatment by identifying the point of diminishing returns for each fracture treatment parameter.

The present disclosure provides a novel and improved method for planning and optimizing hydraulic fracture treatments of subsurface formations in which the steps in an example embodiment are shown in a flow chart in FIG. 1. A method according to this disclosure may make use of prior study of geomechanical properties in the proximity of the treated wellbore or in similar rock formations in other geographic locations. The most commonly studied geomechanical properties in oil and gas operations are listed below as:

Mechanical specific energy (MSE)

Poisson's ratio

Shear stress

Young's modulus

Stress environment,

Min/max horizontal stresses

Overburden stress

V_P/V_S, a ratio of compressional to shear acoustic velocity Rock brittleness (related to bulk and Young's modulus)

Many of the foregoing properties may be determined from common geophysical and petrophysical surveys, including but not limited to well logging measurements such as sonic (acoustic travel time), gamma-ray, neutron porosity, bulk density, drilling data (d-exponent, etc.), seismic data or geological maps. For instance, an acoustic log (V_P/V_S) can deliver a ratio of compressional to shear velocity, elastic moduli (Young's modulus, shear modulus, bulk modulus), Poisson ratio, and porosity. Geochemical data may also be evaluated to determine the TOC, hydrocarbon content. Examples of geochemical data include chemical logs, core, and fluid analysis. Assuming the same treatment parameters, all of these properties and their various combinations will affect the properties and geometry of the created fracture system in the treated formation(s).

Referring to FIG. 1, an example process according to the present disclosure will follow a general procedure comprising the following

At 101 in FIG. 1, obtain existing or available geophysical/geomechanical (e.g., well log) data, MSE logs, and define similar near wellbore regions (zones), or MSE-delineated facies. Other properties helpful in identifying zones comprise local lithology and hydrocarbon properties. As an initial step, existing geophysical/geomechanical data are collected and evaluated to categorize the reservoir rock properties along the well lateral into major intervals of common geophysical properties. Referring briefly to FIG. 2, the reservoir rock has been evaluated based on the analysis of geomechanical (MSE) and geochemical rock properties and is subdivided into 2 major zones (A, B) with 3 subzones (B′, B″, B′″).

Additionally, the data obtained in this step may be analyzed to evaluate zones with similar features such as composition or stress-strain relationship in a rock formation. The characterized zones are later matched with well-path locations such as stages and perforations, as shown in FIG. 2.

Dividing the wellbore into regions based on mechanical properties is important as those properties may, all other parameters being equal, affect the resulting fracture systems. Referring briefly to FIGS. 8 and 9, the figures demonstrate that a more brittle rock 802 in the near-wellbore region will result (as measured) in higher near-wellbore complexity 902 but shorter fracture half-lengths. Similarly, a more ductile region of the rock, 801, will result in longer fracture half-length but reduced near-wellbore complexity, 901.

Referring once again to FIG. 1, at 102, design a primary fracturing model plan for the treatment operation. In this step, a completion engineer defines hydraulic fracture treatment parameters to obtain specific fracture network properties within the reservoir (e.g., fracture length). Completion designs may vary from one well to another based on, e.g., fracture stage axial length and casing/liner perforation schema, fracture fluid pumping rate, proppant (type(s) and amount(s)), additives (gels, breakers, diverters, etc.) among other parameters. For purposes of the present disclosure, the term “fracture treatment parameter” means any of the foregoing parameters that are directly controllable by the fracture treatment operator.

At 103, perform the hydraulic fracturing treatment. An hydraulic fracture treatment is performed on a particular longitudinal well axial interval (“stage”) by perforating the wellbore casing or liner, hydraulically isolating the interval, and pumping under high pressure a fluid comprising specific size proppant, chemicals, etc., according to the designed fracture treatment plan. Although the hydraulic fracture fluid pumping is not always performed exactly according to the intended design, the fluid concentrations are pumped as closely to those designed as practical. The hydraulic fracturing treatment parameters may be recorded based on operator decisions during the treatment operation and generally include but may not be limited to:

Pad volume

Fracture fluid viscosity and density

Propping agent (proppant) type, size distribution and volume

Fluid Injection rate

Treatment volume

At 104, perform fracture network measurements and inversions to obtain certain fracture properties or parameters. For purposes of the present disclosure, the terms “fracture properties” and “fracture parameters” are used to mean properties of fractures induced by pumping hydraulic fracture fluid into a subsurface formation. Such properties may include, without limitation, fracture length, fracture height, hydraulic conductivity. Such properties are a result of the fracture fluid pumping and are not within the direct control of the fracture treatment operator. The implementation, according to this example embodiment of a method, may include a measurement set up as shown in FIG. 3. A fracture pumping truck and/or other completion equipment used in completion 300 are connected to a well 301. Additionally, two modules used in this method act as an acoustic source 302 and an acoustic receiver 303 and the foregoing are also connected to or placed near the well 301 to perform data acquisition simultaneously and without any interference to the other aspects of the completion and fracture treatment. Each module may operate remotely or may be programmed to perform autonomously in real-time. The seismic energy source, at 302, includes a device able to generate an abrupt change in the rate of fluid injection, or a change in fluid pressure sufficient to excite water hammer or tube waves 304 in the well 301. For example, the source 302 can include a piston that quickly releases or inserts a slug of fluid into the wellbore.

The source 302 is designed to generate acoustic (or tube) waves within a frequency range of interest. The frequency range may be related to pre-defined completion properties such as the number of perforations, perforation diameter, perforation cluster spacing, length of the well, and acoustic propagation properties. In some cases, tube wave acoustic signals may be compressional or shear waves generated by other sources in or near the wellbore, such as by rate changes of a fluid pumping unit, devices located at the wellhead, from another wellbore, or downhole/reservoir noises. In real-time applications, if there are multiple sources present, additional filtering may take place to eliminate extraneous signals, or signal conditioning to utilize such pumping signals.

The acoustic energy traveling in the fluid-filled well experiences minor energy loss during propagation along well 301. Since the fluid-filled well is coupled to the surrounding formation, a tube wave reflects with varying properties based on obstacles or changes such as casing size or weight change, formation channels 307, the bottom of the wellbore or plugs positioned in the wellbore 308. The reflections 305 travel back to the location near the surface where they are sensed by one or more pressure transducers (e.g., hydrophones, as shown at 303. Tube wave reflections are influenced by acoustic source signature, acoustic wave velocity, wellbore attenuation, fluid properties, pressure, temperature, depth, and wellbore condition within or near within the targeted fracture treatment stage 306. The reflections 305 carry information related to the downhole conditions and travel back within the wellbore where they are sensed, recorded, and processed by the DAQ system 300. The recorded signals are then processed and analyzed, e.g., by an automated DAQ system that delivers insights in real-time. The recorded tube wave reflections from the downhole features usually contain frequency components ˜0.1-100 Hz.

Pressure and/or pressure time derivative may also be measured in a nearby (offset) well 309.

The automated DAQ system may comprise (none of the following shown separately) a seismic energy source controller, a seismic signal detector, a signal digitizer, computing chip, power supply/source, and a recording device to record the digitized detected seismic signals and the ground surface seismic sensors. The DAQ system may be in signal communication with the SRC and comprise an absolute time to reference recorded signals by using a global positioning system (GPS) satellite. The source controller (not shown) may be configured to actuate the seismic energy source SRC at selected times and cause the sensors to detect seismic signals at selected times autonomously, or substantially continuously. The time-series data is used by computer simulations, which solve partial differential equations governing tube wave propagation and fluid flow in the wellbore, with various objects such as casing change, plugs, perforations, and fractures. Each model is described by idealized models having minimal parameters.

Forward modeling may be embedded in nonlinear full-waveform inversion to estimate the near field fracture system characteristics such as a harmonic average of dimensions, boundary condition, and hydraulic conductivity of each cluster. Different models describing the fracture system as fractures and/or perforation, are designed to understand the downhole properties. A method to characterize the fractures is provided in U.S. Pat. No. 10,641,090 issued to Felkl et al., incorporated herein by reference. Other ways to determine the fracture network properties are also possible, such as using downhole tools, Fiber optic, pressure analysis, electromagnetic analysis, downhole cameras, and/or passive seismic emission tomography (Microseismic). The method, according to this disclosure, is not limited the fracture models described in U.S. Pat. No. 10,641,090. In some embodiment, other possible fracture models such as GDK, elliptical, radial, pseudo 3D, statistical, finite element analysis (FEA), or other custom models can be used to determine fracture closure dimensions (height, length, both, neither) by those skilled in the art.

The fracture characterization described in U.S. Pat. No. 10,641,090 is focused on the near-field (NF) and far-field (FF) analysis. Near-field measurements indicate fracture width and near wellbore conductivity, and far-field measurements indicate fracture length and the reservoir conductivity far from the wellbore. This modeling of reflection data can provide information such as fracture geometry, hydraulic conductivity of the fractures, and level of complexity of the fractures.

The far-field (FF) evaluation comprises measuring pressure using gauges located near the wellhead, an analyzing the pressure decay in the fracture treatment after shut in for a selected time. The pressure decay data is used to construct a model to determine far-field fracture conductivity, fracture complexity, reservoir connectivity for the first or any subsequent fracture treatment stage, or a stimulated portion of the wellbore in real-time.

The near-field (NF) and far-field fracture characterization information combined with the fracture treatment stages located in characterized zones in the previous step may be used to optimize the hydraulic fracturing treatment plan based on the actual rock deformation, fracture characteristics, and hydrocarbon content characteristics. For example, in zones where the development of complex fracture systems are identified, changes to completion design may be implemented to improve the completion efficiency by promoting the development of the desired types of fractures, by modifying operational parameters (e.g., proppant, rate, cluster spacing, etc.) However, in case the complex fracture systems have a high Young Modulus (i.e., brittle rock), changes at the proppant and proppant concentration may be suggested.

Returning to FIG. 1, at 105, compare fracture network measurements and designed fracture network properties. FIG. 2 shows fracture network measurements along a reservoir treated by the same treatment parameters across the well “lateral” (high inclination interval intended to remain in a specific reservoir formation), which resulted in the development of distinctly different fracture types, as shown in zones labeled A, B′, B″, B′″. The well is highlighted based on rock hardness (MSE) information provided from drilling data and was used to divide the reservoir into zones with similar geomechanical properties. FIG. 2 shows several regions of the well categorized into two different zones A (stages 20-28) and Zone B (stages 1-19). Zone B is separated into stages 1-4, B′, and 5-17, B″, and highlighted for average fracture length, fracture height, and near wellbore hydraulic conductivity (NF conductivity). The apparent differences between Zone A and Zone B are in fracture length and near-well hydraulic conductivity. The fracture characteristics of more ductile rock in zone A, stages 20-29, would typically result in longer fractures and lower near wellbore conductivity. This particular example is unique in that NF conductivity in the ductile part (zone A) was much more significant in comparison to the rest of the well (due to a higher concentration of limestone in the rock formations). However, mineralogy of Zone A revealed that there was a high calcium content that caused the fractures to have a high conductivity near-wellbore and long fractures outside. In Zone B, a more brittle rock would be expected to exhibit shorter fractures.

Some of the foregoing information may be available before hydraulic fracturing treatment within the first stage and can aid the treatment plan as well as the prediction of the delivered fracture system. For example, a more brittle rock will result in more complex and shorter fractures, while a more ductile rock will result in more, longer fractures.

At 106, modify fracture treatment parameters to optimize hydraulic fracturing outcomes (reduce uncertainty design vs. actual fracture network parameters). The rock types may be unique for each formation, area and geological zone and may require recalibration on every well base design for the first few stages. The process of optimizing the hydraulic fracturing at a stage-by-stage basis is described in FIG. 15 and further below, while FIG. 16 shows an example of a single well real-time optimization workflow from one zone to another. Other implementation of real-time optimization can also be applied to “zipper” wells as shown in FIG. 17. Operators may choose to create a uniform fracture system or customize the average distributed density of a fracture system at a given interval within the reservoir. However, achieving these objectives may be challenging due to the varying reservoir rock geomechanical properties from one zone to another. FIGS. 10-13 show some practical “manual” modifications to the fracture treatments done to improve fracturing outcomes. FIG. 11 shows that after reduction in length for stage 23, 100 mesh size proppant was increased by 35 klbs. to achieve about 50% increase of fracture length to 300 ft. FIG. 12 demonstrates that based on similarity between stages 23, 24 and 34, the pad for stage 34 was increased by 250 bbls resulting in increased fracture length. FIG. 13 shows various adjustments and parameters made during an improved well completion.

Due to the many changes made to the completion base design, plus the high degree of reservoir heterogeneity, it is difficult for an analyst to detect and fully optimize the hydraulic fracture treatment based on the changes and to modify the treatment design accordingly to achieve desired fracture geometries. This is the main driver behind introducing machine learning (ML) and artificial intelligence (AI) to identify more subtle relationships and perform a more accurate calibration of the treatment to achieve the desired fracture properties in the future wells in a formation or a given area.

A method according to this disclosure identifies the reservoir behavior with regard to the fracture treatment operation based on historical completion data and determines the optimum treatment parameters to further calibrate the completion design for every future stage based on reservoir properties.

As the fracture characteristics (or parameters or properties) are substantially influenced by the rock type, data-driven models can be developed to consider the rock properties along with treatment parameters. The machine learning model only concerns input parameters and adjusts output accordingly. The customized models for each rock type identify areas that require recalibration along the well lateral (or among stages). A machine learning framework consisting of various modules may be used to examine all the possible parameter combination to predicting the treatment parameters based on the desired fracture geometries, and the means to achieve it.

An objective of a ML/AI process in methods according to the present disclosure is to understand the relationship between treatment fracture parameters, rock properties, and the characteristics of induced fractures using a data-driven approach. The target focus is then toward predicting optimized treatment parameters based on the desired fracture geometries, and the means to obtain such optimized parameters. FIG. 19 highlights important treatment parameters included in predicting modeling fracture geometries. FIG. 19 shows that data from multiple well completions can be combined with treatment variables among them to machine-predict expected NF, FF, fracture half-length, fracture height, and fracture width. Important treatment parameters included for modeling fracture geometries may comprise fracture fluid composition, proppant size and size distribution, fracture fluid pumping rate, perforation locations and numbers, axial length of each fracture state, whether the formation is to be treated with acid. Data from multiple well completions can be combined with treatment variables among them to machine-predict expected NF, FF, fracture half-length, height, and width.

Known data from at least one well on treatment parameters such as fluid (for example % pad, overflush, total clean fluid), proppant (size and amounts), pump rate, number and properties of perforation guns, stage length, pumped acid and other (for example gel) can be used as in input to predict behavior of another well. This behavior would include for example near-field complexity, far-field conductivity, fracture half-length, fracture height, and fracture width.

Because fracture characteristics are influenced by the rock type, individual models may be developed for each rock type. These rock type models may be unique for each formation, area and geological zone and require recalibration on every new well.

A module comprising feature extraction, feature selection, and an example of sub-modules, may be used to estimate and predict the fracture geometries given certain rock types and treatment parameters. This module of the machine learning framework enables the user to observe the expected result of a certain fracture treatment design given specific rock properties before a treatment stage is performed. A second module in this machine learning framework comprises a machine learning system that takes the output of the first module (e.g., a regression model), compares to the desired fracture geometries, and adjust the treatment parameters to obtain the optimal fracture network properties for successive stages in a fracture treatment.

Then, machine learning methods of a generally linear regression type, for example Ridge Regression, in combination with a generally non-linear machine learning model, for example, Random Forest, can be used to better understand the impact of treatment parameters on fracture characteristics and identify the point of diminishing returns. These are example methods and other machine learning algorithms and approaches may be used. An example of a regressional model that may be used in this method, Ridge Regression, to study the linear relationship between treatment parameters and fracture geometries, can be fit to the treatment data. A possible advantage of Ridge Regression over Ordinary Least Squares (OLS) is that Ridge Regression can differentiate important from less-important parameters in the model and eliminates those parameters that do not have significant contribution to the model output. Also, Ridge Regression is a good linear model option when there is multicollinearity in the data (i.e., there are highly correlated variables in the data). In addition, Ridge Regression forces the coefficients to spread similarly between the correlated variables, which is an advantage for feature importance interpretation and understanding the importance of treatment parameters and rock properties. However other linear regression models can be used.

Non-linear models such as Random Forest also present a powerful machine learning technique that randomly creates an ensemble of decision trees. Each tree picks a random set of samples (bagging) from the data and models the samples independently from other trees. Instead of relying on a single learning model, Random Forest builds a collection of decision models and the final decision is made based on the output of all the trees in the model. Random Forest can be used for both classification and regression. A Random Forest Regressor can be trained on the treatment parameters with the target output being a characteristic of the fractures (e.g., half-length, fracture height or width). Note that other non-linear regression models can be used.

High-dimensionality is one of the main challenges in the development of fracturing data-driven models. Various feature selection methods have been developed to reduce the dimension of the input data and eliminate variables that do not contribute to the model output. These methods are grouped into Filter and Wrapper methods. Filter methods measure the relevance of the features by analyzing their correlations, while wrapper methods evaluate the effectiveness of a subset of features. One of the wrapper feature selection methods used in this study is Recursive Feature Elimination (RFE).

For the following description, please refer to FIG. 18. Both Ridge Regression and Random Forest models, selected as examples here, have built-in feature selection capability which makes them capable of dealing with high dimensionality. However, for Random Forest, a Recursive Feature Elimination (RFE) method is used to help interpret the feature importance results, and their contribution to the model performance. In the presence of similar features, Random Forest maintains the one deemed as most important, and the rest are aggressively eliminated as they do not provide much further information than the selected feature. In the present application, the application of RFE helps identify equally important features in different runs, while finding a feature combination that yields to the most accurate outcome. The flowchart showing the data-driven modeling of feature geometries is shown in FIG. 18.

The inputs, at 1800, to the ML (AI) model comprise fracture treatment parameters such as fluid types and quantities, proppant type and quantity, stage design, etc. Additional inputs may comprise rock properties, such as gamma ray properties, carbonate content, clay and quartz amounts, rock brittleness, etc. Each row in the model input represents a fracture treatment stage, and the output corresponding to each row includes fracture geometries, and Near Field and Far Field complexities determined at the end of the respective fracture treatment stage. To train the models, at 1802, firstly, each parameter, from among both rock properties and fracture treatment parameters, may be normalized by subtracting the values from the parameter's mean and dividing it by the parameter's standard deviation. Then, the model data set is randomly split into training and test sets, for example, 70% and 30% respectively.

Using training data (which may comprise a combination of input parameters and calculated fracture parameter results), a Ridge Regression model, at 1804, may be developed for each fracture dimension and complexity, i.e., fracture half length, height, and width, as well Near Field and Far Field complexities. L2 is used as the regularization parameter for the Ridge Regression model, and the ridge parameter (k) is changed within a range of 0.1 and 1 to select the optimum value. Ultimately, the trained model may be able to take the fracture treatment parameters and rock properties as input and estimate the expected fracture properties (e.g., dimensions) at 1806. The coefficients in the Ridge Regression model may correspond to the importance of each input parameter.

In parallel to the Ridge Regression model, a Random Forest model at 1810, may be utilized as a more complex and robust alternative to the linear regression. The output of the Random Forest model comprises fracture geometries at 1812. Random Forest uses a subsample selection called bagging and develops individual “trees” for each subsample. In case of strong correlation between inputs, the Random Forest model selects the top one and aggressively eliminates others as they do not have significant additional contribution to the model. For understanding the true feature importance, the Random Forest model may be coupled with a recursive feature elimination (RFE) method encapsulated in boxes 1814 and 1816—to accurately model the fracture properties (parameters) and provide information about the importance of treatment parameters and rock properties for each fracture parameter. For Random Forest, RFE fits a model using all the features, and recursively removes each feature, as shown at 1808. In each iteration, the model accuracy (Mean Absolute Error) and the feature importance values are recorded. The implementation of RFE provides opportunity for all the features to express themselves in the model, and not be shadowed by the most important ones. In each step, the importance of features and the corresponding model accuracy is considered. In the end, the importance of the input parameters is determined based on the maximum contribution each input parameter has had during the process.

Once the training process is completed, test portion of the data is used to validate the models in terms of their generalization ability. The outcomes, at 1818, from the linear model at 1808, and non-linear model at 1810 are compared as follows. Each model is assigned an R-squared (R²) score at 1824 and 1820, respectively. Importance of each feature of the Ridge Regression coefficient model and Random Forest model, 1826, 1822, respectively are estimated. The results are cross validated at 1828, to ensure the validity of both models, even though their accuracy can be different depending on the underlying relationships in the data. The R² scores, or coefficients of determination, are used to observe an amount of variance in the data that the models were able to explain. Using this ML framework, the user is able to estimate the fracture parameters in a particular stage using the fracture treatment as-designed parameters, before any particular stage is fracked.

Each different rock facies (or type) may have a natural tendency to promote a different fracture system, given the same treatment input, due to the different rock properties and stress state. The introduction of AI/ML analysis is a way to diagnose and adjust the treatment to achieve the desired fracture properties for any given rock type. The results of this three-part AI/ML analysis (rock properties, fracture system and completion schedule) affords the opportunity for designing an optimum completion strategy for future wells drilled in the same Eagle Ford area and landing zone.

A method according to this disclosure aims to reduce the uncertainty in the actual fracture network created versus the expected one. In spite of the theoretical plans for hydraulic fracturing treatment, unexpected factors influence the final result. To further understand the deliverables, a fracture characterizing step can take place before (which may comprise natural fractures) and after the hydraulic fracturing treatment of a stage.

The performed optimization may be an adjustment in pumping rate, perforation schema (including hole and explosive charge sizes), cluster spacing, diverter use (if any), proppant type, completion fluids, and the pumping schedule. This step is used to enhance the fracture treatment operation to obtain the desired fracture type. At 107, update well profile as needed. The operator may choose to maintain a well profile of the treated well and compare with other well profiles in the same or similar region as needed. The complete treatment planning analysis of prior stages or/and other wells within the region can be collected as part of a well profile. The well profile may then be used to improve future fracturing treatment operations to deliver the most desirable fracture properties and treatment for a well, or a series of similar wells (similar properties and behavior zones). The well profile develops a treatment system that is appropriate to the characteristics of the reservoir fluids and their anticipated behavior during the production phase (this may not correspond to a maximum produced fluid flow rate). Note that rock-type profile along the wellbore will not change, however the measured fracture properties, conductivity, etc.) will change and deviate from the planned ones.

FIG. 15 shows how the treatment and input from measurements can be implemented in real-time during fracture treatment from stage to stage. The process of optimizing the fracture treatment based on rock properties/facies may be performed stage-by stage. At 1501 input rock properties for stage n, which may comprise gamma ray log measurements, carbonate content, clay and quartz amounts, rock brittleness, among other properties. At 1502, input the fracture treatment design for stage n, which may comprise fracture fluid types and quantities, proppant type and quantity, stage axial length, etc. At 1503, use the trained ML/AI model made as described with reference to FIG. 18. At 1504, use the ML/AI model to estimate fracture properties. At 1505, compare the modeled fracture properties with determined properties using, e.g., reflected tube wave characterization, for example, as set forth in U.S. Pat. No. 10,641,090. At 1506, compare the determined fracture properties to desired or otherwise optimized fracture properties. If the determined fracture properties are not optimized, the operator can the adjust one or more parameters of the fracture treatment at 1507. At 1508, finalize the treatment design for stage n. Between stages n and n+1, at 1509, update the ML/AI model based on the determined fracture parameters. At 1510, use the updated fracture treatment model to analyze stage n+1.

For treatment stage n+1, at 1511, input the rock properties in stage n+1, then at 1512, adjust one or more parameters of the fracture treatment design. At 1513, finalize treatment of stage n+1 and repeat the foregoing process for stage n+2 and any subsequent stages in the designed fracture treatment.

Returning to FIG. 1, at 108, an optional step, diverters may be evaluated if used. If diverters are used in the hydraulic fracture treatment, characterization can be performed in between the insertion of diverters to measure fracture system properties before and after diverter use. Additionally, a separate workflow such as shown in FIG. 5 can be used to optimize the amount and diverter type. At 501, select a zone of interest to evaluate diverter use. At 502, a fracture treatment stage or stages are pumped with an initial amount of diverter. Between or among stages, as many other parameters as possible should be maintained constant. At 503, for each fracture stage may have its near-field and far-field properties evaluated. At 504, one or more subsequent stages may be pumped, using a different amount or amounts or concentrations of diverter. At 505, the foregoing actions at 502 to 504 may be repeated as needed. At 506, using the determined NF and FF properties in each stage, an optimum amount or concentration of diverter may be determined.

In FIG. 4, different fracture systems can be correlated with production, based on production reports from Organic Tracer Data (OFT). The OFT shows higher hydrocarbon production from stages 15+ versus 0-15. This information provides an understanding of the production driving mechanism for these laterals in the Austin Chalk formation. Based on the measurements, the development of the NF conductivity, i.e. higher (3.1) NF value (permeability k times width, w: kw) system is driving the production of these wells during boundary dominated flow conditions as opposed to lower NF conductivity value of 2. Thus an operator interested in optimizing production not only in the short-term but in the long term as well may focus on optimizing the NF kw part of the fracture system.

Another example can take place in zones where a planar radial fracture system is developed and measured. In this case, an alteration to the completion design suggested improving the development of these types of fracture systems within the subsequent zones of the lateral.

Returning to FIG. 1, at 109, optimize diverter use. The optimization of diverter use can be performed independently of well-zone characterization and optimization previously described, e.g. FIGS. 1-2. The use of diverters can enhance hydraulic fracturing stimulation and subsequent production, but the operator needs to decide whether and how much of diverter should be used. FIG. 5 shows the workflow for improved use of diverters for a plug, perforations, and fluid flow operations. Both types of diverter application can be optimized; for example, Far-field diverters can be pumped to increase fracture complexity while nearfield diverters can be used to improve stimulating new sets of fractures in the near-wellbore region. FIGS. 5 and 6 together show how an optimal amount of proppant can be measured. FIG. 5. describes the process of diverter optimization by selecting a reasonably uniform formation wellbore, 501. In step 502, a fixed amount of diverter is pumped as planned in a series of stages. Fracture properties are then measured, 503, and pumping a stage with a different diverter amount is then performed on the same well, 504. steps 502-504 are then repeated as deemed appropriate or needed, 505. The measured fracture properties are then correlated to the amount of diverter to arrive at a desirable diverter amount to achieve objectives, 506. Diverter is expensive, so getting the best results with the smallest amount of diverter is important.

FIG. 6 shows results of fracture system properties using 3 different amounts of diverter. The only slight improvement in stages 35-53 did not justify the larger amount of diverter, and thus 180 lbs was selected as “optimal” for this well.

As outlined in FIG. 5, steps 501-506, performing determinations of near field and far-field properties of a fracture system while holding all parameters but the amount of diverter pumped as the same, can help a possible optimization of the amount of diverter. FIGS. 3, 5 and 6 are an example of a system and a method used to measure the effects of diverters on fracture propagation to control the expansion of a fracture system. In this scenario, different amounts of diverter were pumped as well as stages with little or no diverters. The fracture network characterized by its far-field system (FIG. 6) and the interpretation indicated as the volume of diverters increases the far-field fracture extension reduced. This event stimulates the near-field fracture region and increases the near-wellbore conductivity and complexity.

Further increase in diverter pumped negatively affected the development of the fracture system. Using the described analysis, the operator can identify a “sweet spot” (in FIG. 6 it would be 180 lbs.) for the desired combination of fracture extent and near-wellbore complexity. Based on the presented procedure or a system implemented in a computer, an operator can revise their completion program for as many subsequent wells. The revisions allow the operators to better control and manage fracture propagation and complexity while pumping less diverter. This can result in significant economic improvements and savings.

For example, in FIG. 7, four different pumping schedules, wherein a “schedule” may comprise any or all of maximum fluid pumping rate, total fluid volume, over flush volume, proppant particle size and/or distribution, and whether Acid was used. Two pumping schedules are shown for shale and two are shown for limestone formations, including different pumping rates and acid volumes. The foregoing led to the development of fracture systems with different Near/Far-field kw properties and fracture length (shaded vs white bars), although the same volume of fluid and proppant used in every stage of each zone. In Zone 1 (design 1), the different pumping schedules planned resulted in lower NF kw and lower FF kw values. The low NF indicates a more significant fracture system half-length, while lower FF kw values indicate lower efficiency of proppant distribution in the fracture system. Zone 2 (design 2) was not modified based on this method and resulted in a fracture system with higher NF/FF kw values and normalization of the fracture length, and proppant distribution, based on the fracture property measurements. Zones 3 (design 3) and 4 (design 4) were tested to compare the diverter use efficiency according to this method. Stages in zones 3 and 4 were both located in a limestone formation. However, only zone 3 should use diverters to increase efficiency according to this method. The results indicated an enhanced NF kw system in the stages in which the diverter was pumped within zone 3 compared to zone 4. Therefore, this method resulted in a more conductive fracture system in the near well bore zone area by suggesting the use of diverters. The results as described above can be used in the offset well in the same field to calibrate the completion designs in a well/field development plan, improving completion efficiency and optimizing completion designs.

FIG. 20 shows an example result of optimization for the amount of proppant pumped to obtain a specific fracture half-length under given conditions. As is apparent from the output, if the desired fracture half-length is approximately 400 ft., the amount or proppant required would be about 160,000 lbs. If the desired fracture half-length is 350 ft., that can be obtained with approximately 130,000 lbs. of proppant, wherein all other parameters remain unchanged. Although about 200,000 lbs. of proppant could result in about the same half length of 350 ft., it would be at a cost of additional 70,000 lbs. of proppant and thus not economically favorable.

FIGS. 16, 17 are extensions of FIG. 15 applied to another well (not just stage to stage) and in a case of a zipper frac (fracture treatment). FIG. 15 shows how the treatment input from measurements can be implemented in real-time during pumping from stage to stage. FIG. 16 shows a process of implementing the treatment improvements from well to well. FIG. 17 shows a process of improving treatment among at least 2 zipper fracked wells. In FIG. 16, at 1610 the process starts with each treatment stage having its own design, however the first few (3) stages may use the same treatment parameters to establish a baseline in at first well at, 1620. At 1630, measurements (e.g., analysis of reflected tube wave properties) of the fracture system in a second, similar well, well 2, are evaluated with reference to desired results. At 1640, if necessary, the fracture treatment parameters are fine-tuned with the objective to obtain determined fracture parameters (e.g., from tube wave measurements) closer to target fracture parameters. At 1650, the objective of the remaining fracturing treatment and modifications to subsequent stages are to establish a uniform fracture system across the entire second well. Such uniformity of fracture properties may show increased fluid production as contrasted with non-uniform fracture system in the same reservoir rock.

FIG. 17 shows a process as in FIGS. 15 and 16 expanded to a zipper/pad frac. At 1710, as at 1610, a first few fracture treatment stages are pumped as designed. Each of those stages is evaluated for fracture system properties (NF, FF) at 1720 using methods as explained previously. At 1730, the fracture treatment is adjusted (e.g., one or more fracture treatment parameters) based on fracture system properties determined at an adjacent zipper fracked well. At 1740, the measurements and adjustments are repeated from stage to stage along both (or multiple) laterals to obtain uniform and optimal reservoir coverage (Stimulated Rock Volume, SRV).

FIG. 14 shows an example computing system 1400 in accordance with some embodiments. The computing system 1400 may be an individual computer system 1401A or an arrangement of distributed computer systems. The individual computer system 1401A may include one or more analysis modules 1402 that may be configured to perform various tasks according to some embodiments, such as the tasks explained with reference to FIG. 14. To perform these various tasks, the analysis module 1402 may operate independently or in coordination with one or more processors 1404, which may be connected to one or more storage media 1406. A display device 1405 such as a graphic user interface of any known type may be in signal communication with the processor 1404 to enable user entry of commands and/or data and to display results of execution of a set of instructions according to the present disclosure.

The processor(s) 1404 may also be connected to a network interface 1408 to allow the individual computer system 1401A to communicate over a data network 1410 with one or more additional individual computer systems and/or computing systems, such as 1401B, 1401C, and/or 1401D (note that computer systems 1401B, 1401C and/or 1401D may or may not share the same architecture as computer system 1401A, and may be located in different physical locations, for example, computer systems 1401A and 141B may be at a well drilling location, while in communication with one or more computer systems such as 1401C and/or 1401D that may be located in one or more data centers on shore, aboard ships, and/or located in varying countries on different continents).

A processor may include, without limitation, a microprocessor, microcontroller, processor module or subsystem, programmable integrated circuit, programmable gate array, or another control or computing device.

The storage media 1406 may be implemented as one or more computer-readable or machine-readable storage media. Note that while in the example embodiment of FIG. 14 the storage media 1406 are shown as being disposed within the individual computer system 1401A, in some embodiments, the storage media 1406 may be distributed within and/or across multiple internal and/or external enclosures of the individual computing system 1401A and/or additional computing systems, e.g., 1401B, 1401C, 1401D. Storage media 1406 may include, without limitation, one or more different forms of memory including semiconductor memory devices such as dynamic or static random access memories (DRAMs or SRAMs), erasable and programmable read-only memories (EPROMs), electrically erasable and programmable read-only memories (EEPROMs) and flash memories; magnetic disks such as fixed, floppy and removable disks; other magnetic media including tape; optical media such as compact disks (CDs) or digital video disks (DVDs); or other types of storage devices. Note that computer instructions to cause any individual computer system or a computing system to perform the tasks described above may be provided on one computer-readable or machine-readable storage medium, or may be provided on multiple computer-readable or machine-readable storage media distributed in a multiple component computing system having one or more nodes. Such computer-readable or machine-readable storage medium or media may be considered to be part of an article (or article of manufacture). An article or article of manufacture can refer to any manufactured single component or multiple components. The storage medium or media can be located either in the machine running the machine-readable instructions, or located at a remote site from which machine-readable instructions can be downloaded over a network for execution.

It should be appreciated that computing system 1400 is only one example of a computing system, and that any other embodiment of a computing system may have more or fewer components than shown, may combine additional components not shown in the example embodiment of FIG. 14, and/or the computing system 1400 may have a different configuration or arrangement of the components shown in FIG. 14. The various components shown in FIG. 14 may be implemented in hardware, software, or a combination of both hardware and software, including one or more signal processing and/or application specific integrated circuits.

Further, the acts of the processing methods described above may be implemented by running one or more functional modules in information processing apparatus such as general purpose processors or application specific chips, such as ASICs, FPGAs, PLDs, or other appropriate devices. These modules, combinations of these modules, and/or their combination with general hardware are all included within the scope of the present disclosure.

The data collected and analyzed in this example can be used in machine learning model building to improve the decision-making process. The methods described in the disclosure can be automated in a microprocessor as a system to provide the data and recommendations or pumping treatment adjustments in near real-time.

Although only a few examples have been described in detail above, those skilled in the art will readily appreciate that many modifications are possible in the examples. Accordingly, all such modifications are intended to be included within the scope of this disclosure as defined in the following claims.

Claims

1. A method for optimizing hydraulic fracturing, comprising:

characterizing a fracture induced by pumping fracturing fluid into a subsurface formation, the characterizing comprising analyzing properties of reflected tube waves detected in a well;

modeling change in expected characterization of the subsurface formation with respect to a modeled change in at least one parameter of the pumping fracturing fluid;

comparing the modeled change to a measured change in the characterization with respect to an actual changes in the at least one parameter; and

using the modeled change and the measured change to train a machine learning algorithm to determine an optimized change in the at least one parameter.

2. The method of claim 1 wherein the at least one parameter comprises at least one of pad volume, fracture fluid viscosity, fracture fluid density, propping agent type, propping agent volume, fracturing fluid Injection rate and fracture fluid volume.

3. The method of claim 1 wherein the characterizing a fracture comprises analyzing a far field fracture system property and a near field fracture system property.

4. The method of claim 3 wherein the analyzing the far field fracture property comprises analyzing pressure decay after completing pumping the fracturing fluid.

5. The method of claim 3 wherein the analyzing the near field fracture property comprises analyzing at least amplitude and arrival time of the reflected tube waves.

6. The method of claim 5 wherein the tube waves are induced by inducing a pressure change in the well.

7. The method of claim 1 wherein the machine learning algorithm comprises a recursive feature elimination.

8. The method of claim 7 wherein the recursive feature elimination comprises a ridge regression model for each of a plurality of dimensions of hydraulic fractures.

9. The method of claim 8 wherein the plurality of dimensions comprises at least one of, fracture length, fracture height, near field fracture conductivity and far field fracture conductivity.

10. The method of claim 7 wherein the recursive feature elimination comprises a random forest model for each of a plurality of parameters of a fracture treatment.

11. The method of claim 10 wherein the plurality of fracture treatment parameters comprises at least one of fluid pumping rate, fluid pumping pressure, proppant type, proppant size, proppant size distribution and stage length of a fracture treatment state.

12. The method of claim 1 wherein the at least one parameter comprises at least one of diverter type and diverter amount.

13. A non-transitory computer readable medium comprising logic operable to cause a programmable computer to perform actions comprising:

characterizing a fracture induced by pumping fracturing fluid into a subsurface formation, the characterizing comprising analyzing properties of reflected tube waves detected in a well;

modeling change in expected characterization of the subsurface formation with respect to a modeled change in at least one parameter of the pumping fracturing fluid;

comparing the modeled change to a measured change in the characterization with respect to an actual changes in the at least one parameter; and

using the modeled change and the measured change to train a machine learning algorithm to determine an optimized change in the at least one parameter.

14. The non-transitory computer readable medium of claim 13 wherein the at least one parameter comprises at least one of pad volume, fracture fluid viscosity, fracture fluid density, propping agent type, propping agent volume, fracturing fluid Injection rate and fracture fluid volume.

15. The non-transitory computer readable medium of claim 13 wherein the characterizing a fracture comprises analyzing a far field fracture system property and a near field fracture system property.

16. The non-transitory computer readable medium of claim 15 wherein the analyzing the far field fracture property comprises analyzing pressure decay after completing pumping the fracturing fluid.

17. The non-transitory computer readable medium of claim 15 wherein the analyzing the near field fracture property comprises analyzing at least amplitude and arrival time of the reflected tube waves.

18. The non-transitory computer readable medium of claim 17 wherein the tube waves are induced by inducing a pressure change in the well.

19. The non-transitory computer readable medium of claim 13 wherein the machine learning algorithm comprises a recursive feature elimination.

20. The non-transitory computer readable medium of claim 19 wherein the recursive feature elimination comprises a ridge regression model for each of a plurality of dimensions of hydraulic fractures.

21. The non-transitory computer readable medium of claim 20 wherein the plurality of dimensions comprises at least one of, fracture length, fracture height, near field fracture conductivity and far field fracture conductivity.

22. The non-transitory computer readable medium of claim 19 wherein the recursive feature elimination comprises a random forest model for each of a plurality of parameters of a fracture treatment.

23. The non-transitory computer readable medium of claim 22 wherein the plurality of fracture treatment parameters comprises at least one of fluid pumping rate, fluid pumping pressure, proppant type, proppant size, proppant size distribution and stage length of a fracture treatment state.

24. The method of claim 13 wherein the at least one parameter comprises at least one of diverter type and diverter amount.