Hydraulic fracturing induced event prediction

Abstract

Methods and systems for predicting events during hydraulic fracturing of a target subterranean formation, inccluding: obtaining a time-series of measurements associated with the hydraulic fracturing; decomposing the measurements into a plurality of respective components that collectively represent nonstationary and nonlinear properties of the pressure measurements; extracting respective attributes that are representative of events occurring in the target subterranean formation from the respective components; and predicting, based on the extracted attributes, a likelihood of one or more future events being induced by the hydraulic fracturing.

Description

Hydraulic Fracturing Induced Event PredictionRELATED APPLICATIONS

[0001] This application claims priority to and the benefit of United States Provisional Patent Application No. 63 / 681 ,606 filed August 9, 2024, the contents of which are incorporated herein by reference.TECHNICAL FIELD

[0002] This disclosure relates to hydraulic-fracturing analysis and more particularly to systems and methods for predicting events that can occur due to hydraulic fracturing.BACKGROUND

[0003] Hydraulic fracturing is generally applied to formations which lack the permeability to flow fluids or gas on their own, after a borehole is drilled and a wellbore formed. Hydraulic fracturing employs fluid and material (e.g., proppant) to create or restore fractures in a geological formation in order to stimulate production from new and existing oil and gas wells. Hydraulic fracturing in the oil and gas industry may increase the flow and recovery of oil and / or gas from a well. Natural gas or crude oil may flow more easily from the stimulated formation up the well.

[0004] Operating wells may be subjected to hydraulic fracturing to remain operating. Fracturing may allow for extended production in older oil and natural gas fields.Hydraulic fracturing may also allow for the recovery of oil and natural gas from formations that geologists once believed were impossible to produce, such as tight shale, silt, and carbonate formations.

[0005] Hydraulic fracturing in development of an oil-and-gas well may involve injecting water, proppant, and chemicals under pressure through a wellbore into a geological formation in the Earth crust. This process may create new fractures in the rock as well as increase the size, extent, and connectivity of existing fractures and geological bedding planes within the rock formation itself. Thus, hydraulic fracturing(also called fracing or fracking) is a well-stimulation technique in which rock is fractured by pressurized fluids. The high-pressure injection of fracing fluid (also labeled as fracking fluid, frac fluid, etc.) into a wellbore generates cracks in the deep-rock formations through which natural gas, petroleum, and brine will flow more freely.

[0006] An example of fracing fluid is primarily water containing sand or other manmade proppants, such as ceramics. In some instances, water and proppant make up 98 to 99.5 percent of the fluid by volume used in hydraulic fracturing. In addition, chemical additives may be incorporated in the water to alter viscosity, reduce friction, negate negative rock-fluid and fluid-rock interactions such as the formation of precipitates or scale. The formulation may vary depending on the well, area and formation being treated. In some examples, the sand or other proppants may be suspended in the water with the aid of viscosity increasing agents. Other chemical additives may be added to the fracing fluid to reduce friction, such as in slickwater. Fracing jobs may direct completion hardware, proppant weights, and water volumes to place sand as proppant in the fractures, propping these fractures open to maintain fluid flow during production.

[0007] Hydraulic fracturing can, in some cases, induce events such as seismic events (e.g., induced seismicity), interaction between neighboring wells (e.g., fracture driven interactions, or FDI’s), casing deformation. These fracturing induced events can have negative side-effects that can result in physical damage and / or adversely impact production. Accordingly, there is a need for an accurate system of monitoring for and predicting the likely occurrence of fracturing induced events so that mitigating action can be taken prior to the occurrence of potentially damaging events.SUMMARY

[0008] According to a first example aspect of the disclosure, a method of predicting events during hydraulic fracturing of a target subterranean formation is provided. The method includes: obtaining a time-series of measurements associated with the hydraulicfracturing; decomposing the measurements into a plurality of respective components that collectively represent nonstationary and nonlinear properties of the measurements; extracting respective features from the respective components; and predicting, based on the extracted features, a likelihood of one or more future events being induced by the hydraulic fracturing.

[0009] In some examples of the first aspect, the measurements include a time-series of pressure measurements. In some examples of the first aspect, the measurements include a time-series of seismic wave measurements.

[0010] In one or more off the preceding examples, decomposing the measurements into a plurality of respective components comprises performing a decomposition in respect of the measurements, wherein each component comprises a respective intrinsic mode function (IMF).

[0011] In one or more off the preceding examples, the decomposition comprises an empirical mode decomposition (EMD).

[0012] In one or more off the preceding examples, the decomposition comprises a variational mode decomposition (VMD) and each component comprises a respective band limited IMF.

[0013] In one or more off the preceding examples, extracting the respective features from the respective components comprises computing an instantaneous frequency (IF) for each IMF.

[0014] In one or more off the preceding examples, computing the instantaneous frequency (IF) for each IMF comprises applying a Hilbert transform to each IMF.

[0015] In one or more off the preceding examples, decomposing the pressure measurements into a plurality of respective components and extracting the respective features collectively comprise performing a Hilbert-Huang Transform (HHT).

[0016] In one or more off the preceding examples, extracting respective features comprises obtaining time-frequency-energy distribution data based on the respective components.

[0017] In one or more off the preceding examples, extracting respective features comprises using one or more machine learning models to extract behavioral features based on one or both of the IF or IMF.

[0018] In one or more off the preceding examples, predicting the likelihood of one or more future events being induced by the hydraulic fracturing comprises using one or more machine learning models to map the extracted features to respective predictions of the one or more future events.

[0019] In one or more off the preceding examples, the method includes training the one or more machine learning models based on historic data obtained in respect of historic hydraulic fracturing of one or more subterranean formations.

[0020] In one or more off the preceding examples, the one or more machine learning models comprises deep neural networks.

[0021] In one or more off the preceding examples, training the one or more machine learning models comprises pre-training a first machine learning model of the one or more machine learning models based on historic data obtained in respect of the historic hydraulic fracturing of the one or more subterranean formations other than the target subterranean formation, and fine-tuning the first machine learning model based on historic data obtained in respect of historic hydraulic fracturing of target subterranean formation.

[0022] In one or more off the preceding examples, the one or more future events is selected from a group comprising: events that are expected to cause environmental or structural damage to or in the vicinity of the target subterranean formation; and events that can adversely impact extraction of a hydrocarbon through the hydraulic fracturing.

[0023] In one or more off the preceding examples, the one or more future events are selected from a group comprising: an induced seismicity (IS) event that exceeds threshold event criteria; a fracture driven interaction (FDI) between different wells; a casing deformation; and a plug slip.

[0024] In one or more off the preceding examples, the hydraulic fracturing comprises: injecting a frac slurry comprising frac fluid and proppant through a wellbore into the target subterranean formation; and adjusting operating parameters of the hydraulic fracturing in real time based on the predicting to reduce the likelihood of the one or more future events.

[0025] In one or more off the preceding examples, the method includes generating a real-time display that indicates the predicted likelihoods of the one or more future events.

[0026] In one or more off the preceding examples, obtaining the time-series of measurements associated with the hydraulic fracturing is performed in respect of a first wellbore in the target subterranean formation, and the method further comprises obtaining further time-series of measurements in respect of one or more further wellbores in the target subterranean formation; wherein: decomposing the measurements into a plurality of respective components is performed in respect of each of the measurements from the first wellbore and the further time-series of measurements from the one or more further wellbores; and the predicting is based on extracted features from the respective components from both the first wellbore and the one or more further wellbores, a likelihood of one or more future events being induced by the hydraulic fracturing.

[0027] In one or more off the preceding examples, the method includes analyzing the extracted features to assess fracture effectiveness during the hydraulic fracturing; and based on the assessed fracture effectiveness, generating an optimization signal to guide stimulation operations in real time, wherein the optimization signal comprises one or more of: (i) a visual indicator displayed on an operator interface, (ii) a real-time alert recommending an adjustment to one or more operational parameters, and (iii) a machine-readable control signal for automatic adjustment of one or more of a pump rate, fluid viscosity, or proppant concentration. In one or more of the precedingexamples, the time-series measurements are pressure measurements obtained at a wellbore head associated with the hydraulic fracturing.

[0028] According to a further examples aspect, a computer implemented system is disclosed that is configured for performing the methods of any of the preceding examples of the first aspect.

[0029] The details of one or more implementations are set forth in the accompanying drawings and the description below. Other features and advantages will be apparent from the description and drawings, and from the claims.BRIEF DESCRIPTION OF DRAWINGS

[0030] FIG. 1 is a diagram of the real-time fracing process and a well site having a hydraulic fracturing system.

[0031] FIG. 2 is a diagram of a subterranean geological formation having fractures around a wellbore.

[0032] Fig. 3A is a diagram illustrating three separate atypical fracturing-induced events which are considered risks, but may be the result of the same horizontal bedding slip mode of failure.

[0033] Figs. 3B to 3E shows three-dimensional diagrams of a subterranean geological formation, representing a normal state and three separate atypical fracturing- induced events.

[0034] FIG. 4A is a flowchart of a process for predicting atypical fracing-induced events in a subterranean formation, according to example embodiments.

[0035] FIG. 4B is a flowchart illustrating a data collection operation of the process of FIG. 4A.

[0036] FIG. 4C is a flowchart illustrating a decomposition operation of the process of FIG. 4A.

[0037] FIG. 4D is a flowchart illustrating a feature extraction operation of the process of FIG. 4A.

[0038] FIG. 4E is a flowchart illustrating an event prediction operation of the process of FIG. 4A.

[0039] FIG. 4F is a flowchart illustrating a process for training one or more machine learning based models for the event prediction operation of FIG. 4E;

[0040] FIG. 5A illustrates decomposition of a signal into a plurality of intrinsic mode functions (IMFs).

[0041] FIG. 5B illustrates an IMF and its corresponding analytical signal corresponding to an operational pressure signal segment.

[0042] FIG. 5C illustrates a set of decomposed operational pressure signal segments corresponding to a pressure signal.

[0043] FIGs. 6A and 6B are respectively schematic perspective and side views illustrating a pre-existing fault plane in a subterranean formation at which a well site is located.

[0044] FIG. 7 is a set of plots that visually illustrate information that can be obtained from pressure data that includes a time series of pressure measurements.

[0045] FIG. 8 is a set of plots that have been extracted from the plots of FIG. 7.

[0046] FIG. 9 is a set of plots illustrating properties present in IMFs and sets ofInstantaneous Frequencies (IFs) obtained in respect of pressure measurements for periods leading up to an induced atypical seismic event and a period following the included seismic event.

[0047] FIG. 10A is a schematic diagram representing correlation of measurements from multiple wells in a subterranean formation.

[0048] FIG. 10B is a flowchart illustrating an offset well analysis process that can be performed to support correlation of measurements from multiple wells in a subterranean formation.

[0049] FIG. 11 is an example a graphical user interface display representing information that can be generated according to example implementations.

[0050] FIG. 12 is diagram of a computing system having a processor and memory storing code executed by the processor to receive an indicator of the amount of shear fracturing during hydraulic fracturing.

[0051] FIG. 13 is a block diagram of a tangible, non-transitory, computer-readable medium that can facilitate analysis and control of hydraulic fracturing.DETAILED DESCRIPTION

[0052] Methods and systems for predicting atypical events caused by hydraulic fracturing are disclosed. These events may include various undesirable events such as seismic events varying in size from microseism ic events all the way up to large seismic events which can cause damage to surface infrastructure as well as wells in the field, fracture-driven interactions (FDI’s), and wellbore casing deformations. In one example, the method includes obtaining, during hydraulic fracturing in a region, time-series measurements of a physical property that is indicative of the subterranean impact of the hydraulic fracturing. For example, the physical property can be pressure, and the timeseries measurements can be pressure measurements. Alternatively, the physical property can be a seismic wave (for example as represented in an acoustic signal) and the time-series of measurements can be seismic wave measurements. The measurements are decomposed into components representing nonstationary and nonlinear properties, extracting features indicative of the conditions in the subterranean formation, and predicting the likelihood of future adverse events based on these features.

[0053] In some examples, time-series measurements can be obtained for both pressure and seismic waves in the field.

[0054] In some scenarios, the disclosed method and systems may allow for realtime adjustment of fracturing operations to mitigate risks by altering active hydraulic fracturing parameters.

[0055] To provide context, FIG. 1 illustrates an example of an environment to which example implementations of methods and systems disclosed herein can be applied. In particular, FIG.1 is a schematic diagram illustrating a well site having a hydraulic fracturing system 100 that employs a fracing process with fracing slurry having fracing fluid 102 (e.g., slickwater) and proppant 104. The fracing fluid 102 may be held in a vessel(s) or surface pits. The fracing fluid 102 and proppant 104 may be stored in vessels or containers and including on vehicles (e.g., trucks, skids, etc.) in certain examples. In some implementations, the fracing fluid 102 is slick water which may be primarily water, such as 98.5% or more by volume. The fracing fluid 102 can also be gel-based fluids. The fracing fluid 102 can include polymers and surfactants. Other common additives may include hydrochloric acid, friction reducers, emulsion breakers, emulsifiers, and so on. At the well site, the proppant 104 can be provided from multiple railcars, hoppers, transport trucks, containers, sand castles or bins of proppant of differing mesh size (particle size).

[0056] The system 100 includes control devices 106 and 108 for the fracing fluid 102 and the proppant 104, respectively to prosecute the fracing process that are actions to realize fracturing. The control device 106 may include pumps as motive devices (and also as metering devices in some examples). The control device 106 for the fracing fluid 102 may also include a control valve in certain examples. The pumps may be, for example, positive displacement and arranged in series and / or parallel. In some examples, the speed of the pumps may be controlled to give desired flow rate of the fracing fluid 102. The proppant control device 108 may include, for example, a blender, feeder (e.g., rotary feeder, etc ), conveying belt, metering device, and so on. A blender, for example, may be a solid blender that blends proppant 104 of different mesh size. The proppant 104 may be added (e.g., via gravity, or mechanical feed) to a conduit conveying the fracing fluid 102 such as at a suction of a fracing fluid pump (e.g., of 106) to give a stream 110 that enters the wellbore 112 for the hydraulic fracturing. Thus, the stream 110 may be a slurry that is a combination of the fracing fluid 102 and proppant104. The stream 110 may be labeled as fracing fluid or as fracing fluid having proppant. For instances when proppant 104 is not added to the fracing fluid 102, the stream 110 entering the wellbore 112 for the hydraulic fracturing may be the fracing fluid 102 without proppant 104.

[0057] The wellbore 112 may be formed through the Earth surface 114 into a subterranean geological formation 140 in the Earth’s crust. The source of fracing fluid 102 and the source of proppant 104 may be disposed at the Earth surface 114. The wellbore 112 may be vertical, horizontal, or deviated. The wellbore 112 may be a cemented cased wellbore and have perforations for the stream 110 to flow (injected) into the formation. Balls that seat in the well, or other techniques, may also be utilized to control the fracing (fracturing) process.

[0058] The hydraulic fracturing system 100 may include a control system 116 to direct operation of the hydraulic fracturing system. The fracturing system 100 generally includes gauges or sensors to measure different operating parameters. For example, the system 100 may include a pressure sensor 118 disposed at a wellhead (not shown) of the wellbore, or in the subsurface of the wellbore 112 to measure the wellhead and / or bottom pressure during the hydraulic fracturing. In some implementations, the control system 116 may receive the measured pressure data and may also consider the wellhead pressure as the treating pressure of the hydraulic fracturing. In some examples, the fracturing system can include seismic sensors 124 that can be deployed at one or both of surface locations and down hole locations to measure seismic signals during the hydraulic fracturing. For example, the seismic sensors 124 cpoudl include one or more of: geophones, accelerometers, tilt-meters, distributed acoustic sensing systems, fiber-optic cables and / or arrays of the aforementioned configured to monitor seismic waves. The control system 116 may include a computing system to implement techniques described herein associated with analysis (including subterranean event prediction) and control. The computing device may be disposed within the control system 116. The computing device may instead be a field computer or remotecomputer. The control system 116 may include one or more controllers. The control system 116 (and the computing device) can include a processor(s) 120, such as a central processing unit (CPU) or microprocessor, or cloud based computing system. The control system 116 (and the computing device) can include memory 122 storing code (e.g., logics, instructions, etc.) executed by the processor 120. The memory 122 may include volatile memory, such as random access memory (RAM), cache, cloud computing, etc. The memory 122 can include nonvolatile memory, such as a hard drive, read only memory (ROM), solid state drives, or cloud servers, etc.

[0059] Hydraulic fracturing may create new fractures in the rock in the geological formation, as well as increase the size, extent, and connectivity of existing fractures and bedding planes in the geological formation. The producing formation is fractured open via hydraulic pressure. Then, proppants 104 (propping agents) are pumped into the oil or gas or other type of well with fracing fluid 102 to hold the fractures or fissures open so that energy (pressure) can be applied (e.g., via pumped fracing fluid) into the formation and converted to stress to enhance the breaking of the rock. Hydraulic fracturing is generally employed in low-permeability rocks, such as tight sandstone, shale, siltstone, carbonates, or mixed lithologies and some coal beds, to increase crude oil or gas flow to a well from petroleum-bearing rock formations. A beneficial application may be horizontal wellbores in low-permeability geological formations having hydrocarbons, such as natural gas, crude oil, or other marketable fluids or gases such as helium, etc. Massive hydraulic fracturing or high-volume hydraulic fracturing may be applied to formations containing the above fluids or gases where the rock formation or reservoir have low natural permeability (e.g., less than 0.1 millidarcy) and often in the nanodarcy range (e.g., less than 1 microdarcy) of permeability

[0060] FIG. 2 illustrates geological formation 140 having fractures 208 around a wellbore 112. Fractures may begin vertical along, or perpendicular to wells. The fractures may grow in exiting fractures, within bedding layers, or create their own pathways according to the principal stresses exerted on the formation. The wellbore112 may have perforations 204 for injection of fracing fluid into the geological formation 1 0, or may rely on a series of shifting, sliding, or ball activated sleeves or ports to allow fracturing energy to enter the formation from the wellbore. In the illustrated embodiment, the wellbore 112 is a horizontal wellbore, but may include vertical or deviated well trajectories. The fractures may include fractures along vertical (or near vertical) fracture planes 206 with respect to the wellbore 112. The vertical planes 206 may include a vertical plane (or near vertical plane) that is generally perpendicular to the longitudinal length of the wellbore 112 and may include a vertical plane that is parallel (in-line) with the longitudinal length of the wellbore 112 or any azimuth angle between the two. In other words, the fractures may begin vertical perpendicular to the wellbore 112 and may also begin vertical along the wellbore 112 or any azimuth angle between the two. The fractures may grow into existing fractures within bedding layers. These fractures 208 (shear, tensile, dilation or any other mode of opening fractures and / or connected existing fractures) may be in a horizontal bedding layer(s) or stratigraphic layer(s) 210. The fracture tips of the fractures 208 (or a portion of each fracture 208 toward or near the fracture tip) may receive proppant. The proppant may be fine sand or other material (e.g., generally composed of sorted sands of any mesh size) so that the proppant may reach further into the fractures and allow these fractures to remain open after fracturing operations are completed 208. These lines represent a skeletonized schematic of rock layer separation extending in two dimensions to the dotted line boundary.

[0061] FIG. 3A is a further representation of illustrative geological formation 140 illustrating a plurality of wellbores 112A, 112B, 112C that support a hydraulic fracturing process, illustrating atypical fracing-induced events that systems and method of the present disclosure may predict, according to example implementations. These atypical events may, for example, include an induced seismicity (IS) event 310 that exceeds threshold event criteria; an induced fracture driven interaction (FDI) event 312 between different wells; a casing deformation event 314; and / or a plug slip event (not shown). Byway of further example, FIGs. 3B, 3C, 3D and 3E is a set of four illustrative representative 3D plots of geological formation 140 with a preexisting fault 302, illustrating a parent wellbore 112A, and three child wellbores 112B1 , 112B2 and 112B3. FIG. 3B illustrates standard hydrofracturing (HF) operations with no bedding plane activation; FIG. 3C illustrates an occurrence of well-scale bedding plane activation together with fracture driven inter-well fluid interactions between child wells 112B3 and 112B2, and child well 112B1 and parent well 112A; FIG. 3D illustrates an occurrence of pad-scale bedding plane activation, aseismic slip and casing deformation; FIG. 3E shows large scale bedding plane activation and an induced seismicity event (IS). The atypical fracing-induced events may be driven by the activation of subsurface features driven by stress changes to, hydraulic connection to, or other interactions with the subsurface features or generation of new subsurface features. The atypical fracing- induced events may be preceded by aseismic slip processes that the method of the present disclosure may monitor for and use for prediction.

[0062] FIG. 4A is a flow diagram illustrating a process 400 that can be performed at a computer system, for example a computer system of the control system 116 of hydraulic fracturing system 100, to predict atypical fracing-induced events during realtime hydraulic fracturing in a target subterranean formation 140.

[0063] Process 400 includes an on-going data collection operation 402 during which a time-series of measurements is obtained during operation of the hydraulic fracturing system 100. The time-series measurements can include one or both of pressure measurements and seismic wave measurements. With reference to FIG 4B, data collection operation 402 includes a set of actions, including a measurement action 422 for obtaining the time-series of pressure and / or seismic wave measurements and outputting such measurements in the form of a signal 426. By way of example, pressure measurements can be obtained from pressure sensor 118 installed at a wellhead of a wellbore 112 that extends into the subterranean formation 140, and, seismic wave measurements can be obtained from surface located seismic wave sensors 124. Thesemeasurements are recorded throughout the hydraulic fracturing process, capturing the dynamic changes in pressure that occur as a fracturing fluid is injected. In some examples, pressure measurements may alternatively or additionally be obtained via downhole offset pressure sensors and seismic wave measurements obtained vis downhole sismic wave sensors. Thus, in example embodiments, measured signal 426 comprises a real-time single-dimensional time-series which contains the measurements obtained during the hydraulic fracturing process. In some examples, measured signal 426 is a real-time single-dimensional time-series representation of pressure measurements. In some examples, measured signal 426 is a real-time singledimensional time-series representation of seismic wave measurements. In some examples, multiple measured signals 426 can be obtained, including signals corresponding to well-head pressure measurements signals, down-hole pressure measurements, surface seismic wave measurements and downhole / subterranean seismic measurements. The operations described herein can be performed in respect of any pressure measurement derived signal or seismic wave measurement derived signal, and the terms “signal” 426, “operational signal segments” 430, “decomposed signal segments” 454, and “extracted feature” 470 as described below can refer to either a pressure measurement based signal or a seismic wave measurement signal unless otherwise stated or unless the context requires otherwise.

[0064] In the illustrated example, data collection operation 402 includes a segmentation action 428 for segmenting the signal 426 into a sequence of operational signal segments 430. Signal segmentation is a process through which a time series of measurements is divided into discrete segments. These signal “segments”, can each have associated segment time data (e.g., two or more of a segment start time, a segment duration, and a segment end time) and represent a set of measurements made within a respective time duration of the time-series of measurements.

[0065] Signal segmentation action 428 can be performed based on operational properties of the hydraulic fracturing process, such as slurry rate, proppantconcentration, chemical additives, or other properties, to provide operational pressure signal segments 430 that correspond to periods that have similar fracturing process operational properties.

[0066] In this regard, in example implementations, the data collection operation 402 also includes a set of actions 431 for identifying and classifying operational signal segments. An operational property measurement action 432 obtains a time-series of operational measurements to provide a real-time operational signal 434 contemporaneously with the signal 426. The real-time operational signal 434 can have as many dimensions as operational properties that are being tracked. For example, in a simple case, the real-time operational signal 434 can be a single dimensional signal that represents a time-series of slurry flow rate values. In the case where further operational properties (e.g., proppant concentration and additive concentration) are also tracked, additional dimensional elements can be added to each time step in the operational signal for the additional operational properties.

[0067] In some examples, at least some of the operational measurements are setpoint parameters that are recorded based on data obtained from one or more of control systems 106, 108 or 116 throughout the hydraulic fracturing process, capturing the changes in set operational parameters as they are prescribed during fracturing fluid injection. In some examples, at least some of the operational measurements may be obtained from process sensors that measure real-time operational properties.

[0068] In the illustrated example, operational signal 434 is processed to identify and classify operational signal segments (Action 436) to divide the operational signal 434 into operational signal segments 438 based on changes in the operational properties. In one example, identification and classification action 436 can be performed using a changepoint detection process that applies changepoint analysis algorithms to identify discrete operational signal segments 438 based on changes in the probability distribution of the operational measurements included in the operational signal 434. In some examples, machine learning models obtained from un-supervised or supervisedmachine learning algorithms may be used to process operational signal 434 to identify and classify the operational signal segments 438. In some implementations, unsupervised machine learning algorithms, including but not limited to K-Means, Gaussian Mixture Models, or others, may be used to classify groups or “clusters” of operational signal segments. In some implementations, unsupervised models may assign classes based on pre-defined thresholds or known statistical characteristics of particular classes. In other implementations, supervised models may be pre-trained based on a plurality of operational signal segments with previously determined classes. These supervised models may for example be based on supervised machine learning algorithms including but not limited to linear regression, logistic regression, decision trees, mixture density networks, sequence to sequence, transformers or others.

[0069] By way of example, changepoint analysis can be performed in respect of an operation signal 434 that comprises a single-dimensional time series of operational parameters representing a slurry rate in order to perform rate-based segmentation that identifies operational periods corresponding to (i) “on”; (ii) “off” ; and (iii) “outlier” classifications, which respectfully refer to segments that (i) occur during a primary fracturing stage, (ii) occur outside of the fracturing stage, or (iii) are outliers.

[0070] In other implementations, the operational segment classes may include further categories including but not limited to “pressure testing”, “fracture initiation”, “pumping period(s)”, “shut-in”, “flowback”, or others. In other implementations, segmentation may be carried out on multi-dimensional time series of operational parameters such as the slurry rate, proppant concentration, or others to further identify signal segments associated with further types of operational phases.

[0071] In at least some examples, the operational segments may based on operational measurements obtained from a plurality of wells or well pads.

[0072] In the illustrated example, signal segment classification action 436 outputs segment time data (e.g., at least two of a segment start time, segment duration andsegment end time) for each identified operational segment 438, together with a segment classification (e.g., “pre-fracturing”, “fracturing”, or “post-fracturing”).

[0073] The operational signal segment time data and segment classifications associated with signal segments 438 can then be applied by signal segmentation action 428 to divide the signal 426 into corresponding signal segments 430. In some examples, the signal segments 430 each inherit the segment time data and operational segment classification as its corresponding operational signal segment 438.

[0074] In some examples, the signal 426 may also be provided as an input to operational signal identification and classification operation 436 to provide a further dimension of data for operational segment identification and classification purposes. In some examples, signal segments 430 may be identified and classified solely based on operational properties inferred from the signal 426.

[0075] In some implementations, as part of signal segmentation action 428, the signal segments 430 may be normalized based on the operational parameters within the operational segment to remove the influence of operational changes.

[0076] As will be explained in greater detail below, in some examples, time-series measurements 422 can also be obtained simultaneously from multiple hydraulic fracturing wells that run adjacent to each other in a subterranean formation. The plurality of signals 426 may be individually processed into operational signal segments 430 using the techniques described above.

[0077] Accordingly, the output of data collection operation 402 includes a sequence operational signal segments 430. Each operational signal segment 430 comprises a time-series of measurements and has associated segment time data and an associated segment classification.

[0078] Referring again to FIG. 4A, process 400 includes a decomposition operation 404, an example of which is shown in greater detail in FIG. 4C. As indicated in FIG. 4C, decomposition operation 404 can include a first decomposing action 440 for decomposing each of the classified operational signal segments 430 into a respectiveplurality of components that collectively represent nonstationary and nonlinear properties of the measurements that make up the signal segments 430. In an example implementation, the time-series of measurements that make up each operational signal segments 430 are subjected to an empirical mode decomposition (EMD) process, where each measurement is broken down into intrinsic mode functions (IMFs). These functions represent the various frequency components of the signal segment 430, highlighting the nonstationary and nonlinear characteristics inherent in the fracturing process. In an example implementation, an EMD process can include the following iterative process: (1 ) Identify Extrema: Detect all local maxima and minima of a signal ( e.g., the collected time-series of measurements); (2) Create Envelopes: Interpolate between all maxima to form an upper envelope, and all minima to form a lower envelope; (3) Compute Mean: Calculate the mean of the upper and lower envelopes; (4) Extract Detail: Subtract the mean from the original signal to obtain a first intrinsic mode function (IMF); and (5) Iterate: Repeat actions (1 ) through (4) on the residual (original signal - extracted IMF) to extract further IMFs, continuing until the residual becomes a monotonic function from which no more IMFs can be extracted. By way of illustration, FIG. 5A illustrates decomposition of an operational signal segment 430 (Signal) into a plurality of respective intrinsic mode functions IMF 1 to IMF 5, plotted as amplitude over time. Each IMF represents a different frequency component of the input signal, enabling for the separation and characterization of different frequency components and supporting analysis of the complex behavior seen in signal 426.

[0079] In an alternative implementation example of decomposition operation 404, the time-series of measurements are subjected to a variational mode decomposition (VMD) process, where each measurement is broken down into a set of band-limited intrinsic mode functions (BL-IMFs). These functions represent the various frequency components of the signal segment 430, highlighting the nonstationary and nonlinear characteristics with the additional constraint of having a limited frequency band. In an example implementation, a VMD process can include the following iterative process: (1 )Initialize Modes and Frequencies: Begin with an initial guess for each mode and its associated center frequency; (2) Transform and Update Modes: For each mode, transform the signal into the frequency domain and update the mode to minimize its bandwidth around the current center frequency; (3) Update Center Frequencies: Recalculate the center frequency of each mode based on its updated spectral content; (4) Enforce Reconstruction: Ensure the sum of all modes approximates the original signal by incorporating a constraint (e.g., through Lagrange multipliers); and (5) Iterate: Repeat actions (2) through (4) until convergence is achieved, typically when successive updates result in negligible changes, yielding a predefined number of band-limited modes that collectively represent the input signal. Each BL-IMF represents a different frequency component of the input signal, enabling for the separation and characterization of different frequency components and supporting analysis of the complex behavior seen in signal 426. The BL-IMFs may be utilized in the system in a similar way to IMFs, and the term IMF and BL-IMF are considered interchangeable for in this disclosure unless specified otherwise or the context requires otherwise.

[0080] A further processing action 444 is performed to extract information on the amplitude and frequency content of the input signal from each IMF 442. In example implementations, processing action 444 obtains an analytic signal 446 by applying a Hilbert transform to each IMF 442. The analytic signal 446 is a complex-function representation of the IMF 442 that embeds information about an instantaneous frequency (IF) 452 of the IMF 442 and an “envelope” 450 of the IMF 442. The envelope 450, which is represented in the real-number portion of the analytic signal 446, is uniquely suited to representing the change in frequency dependent amplitude of the signal over time. Processing action 448 extracts the IF 452 and envelope 452 from analytical signal 446. For example, with reference to FIG. 5B a Hilbert transform can be performed on each IMF in action 444 to derive a respective analytic signal (represented by envelope 450), and an instantaneous phase and instantaneous frequency (IF) can then be computed in action 448 for each respective analytic signal 446. Theinstantaneous frequencies (IFs) 452 represent Time-Frequency-Energy data. For example, a Time-Frequency representation can be obtained by mapping the instantaneous frequencies and their corresponding amplitudes for each IMF over time, and an Energy Representation can be obtained by calculating the energy of the signal at each time-frequency point using the amplitude squared.

[0081] Accordingly, in an example implementation, the decomposition operation 404 (e.g., the collective actions 440, 444 and 448 of decomposing an operational signal segment 430 into a plurality of respective IMFs in action 440 and extracting the respective envelope and IFs features in actions 444 and 448) comprises performing a Hilbert-Huang Transform (HHT). The Hilbert-Huang Transform (HHT) is useful for analyzing time-series data that is both nonlinear and non-stationary. The combination of EMD with the Hilbert Spectral Analysis (HSA) that to adaptively decompose a signal into intrinsic mode functions (IMFs) and provide instantaneous frequency data can offer superior resolution and accuracy in signal analysis compared to traditional methods such as Fourier transform methods that assume data linearity and stationarity. The combined output of the decomposition operation 404 that is generated in respect of an operational signal segment can be referred to as a decomposed signal segment 454, and can include the IMFs 442, envelopes of the IMFs 450 and the IFs 452 of the IMFs 442.

[0082] FIG. 5C illustrates example of a set of decomposed operational signal segments 454 (individually labelled as decomposed operational signal segments 454-1 to 454-11 ) generated in respect of operational signal 434. As shown in FIG. 5C, each decomposed operational signal segment 454 includes data that represents, for the segment: IMF 442, IMF envelope 450 (and smoothed envelope); mean amplitude, and extreme.

[0083] Referring again to Figure 4A, the process 400 includes a feature extraction operation 406 that extracts features 470 in respect of operational signal segments 430 and decomposed signal segments 454. Feature extraction operation 406 can includemultiple parallel feature extraction processing paths, with each path configured to extract features corresponding to a respective behavioral category (e.g., a target behavioral category / ) from a respective input signal segment. In this regard, FIG. 4D illustrates an example of a generic feature extraction processing path 406- / that represents one of one or more feature extraction processing paths supported by feature extraction operation 406. In the illustrated example, the feature extraction processing path 406- / receives input signal segments 458 that can include operational signal segments 430, decomposed signal segments 454, or both operational signal segments 430 and decomposed signal segments 454, depending on the target behavioral category / of the feature extraction processing path 406- / . In some examples, the feature extraction processing path 406- / can be configured to process only input signal segments that have a specific operational classification. For example, where the target behavioral category / is related to an objective of identifying discrete events which occur during hydraulic fracturing, then processing may be restricted only to input signal segments that have bene previously associated with / assigned an operational classification of “on” (e.g., related to the primary fracturing period). In other implementations, the objective may be to characterize a breakdown period or post shut- in behavior of a fracturing stage, and therefore may be carried out on segments having operations classes related to those respective operational periods.

[0084] In the illustrated example, processing path 406- / incudes a behavioral segmentation action 460 for forming the input signal segment 458 into a respective set of behavioral segments 462. In some implementations where the input signal segment 458 is a decomposed signal segment 454, behavioral segmentation is performed in respect of each of the components of decomposed signal segment 454 (e.g., IMFs 442, IMF envelopes 450, and IFs 452). In some implementations, behavioral segmentation may be carried out on a subset of the decomposed signal components, including any combination of individual IMFs 442, IMF envelopes 450, and / or IFs 452.

[0085] In other implementations, decomposed signal components may be combined based on statistical properties, frequency behavior, signal stability, amplitude, or other metrics depending on the context.

[0086] Pre-processing may be carried out on the input signal segment(s) 458 prior to behavioral segmentation action 460, including but not limited to gaussian filtering, seasonal trend decomposition, scaling, or others. Statistical properties of the input signal segments 458 may be computed such as the rolling mean, rolling variance, derivative(s), or others. Time-frequency analysis such as Wavelet transforms or short- time Fourier transforms may be calculated to capture time-frequency components of the input signal segment(s) 458.

[0087] In some examples, behavioral segmentation action 460 may be carried out using techniques such as the aforementioned changepoint detection (mentioned above in respect of operational segmentation). In other example implementations, behavioral segmentation action 460 may be carried out using dynamic time warping (DTW). DTW is particularly suited for time-series analysis and is able to identify similarity between signals occurring at different speeds. DTW may be used with various techniques, including but not limited to similarity-based segmentation techniques such as template matching, sliding windows, time-series clustering, or others. Behavioral segmentation action 460 may utilize machine learning models obtained using un-supervised or supervised techniques. Supervised training can, for example, be based on a training dataset of labelled behavioral segments.

[0088] In some examples, in some feature extraction processing paths 406- / , multiple behavioral segmentation techniques may be applied to combinations of input signal(s) based on the context of the desired analysis, and the outputs may be combined to define a set of behavioral segments 462. For example, in a feature extraction processing path 406- / , a first behavioral segmentation action 460 may be performed only on IMFs 442 which have relatively stable instantaneous frequencies in order to correlate frequency behavior between wells. A second behavioral actionsegmentation action 460 may be performed based only on residual components of the decomposed signal segments 454 that correspond to operational segments classified as “after shut-in” to analyze the long-term fall-off trend in a well. The combined output of the first and second behavioral segmentation actions collectively form a set of behavioral segments 462. Each of the behavioral segments has associated behavioral segment time data (e.g., a time stamp) that includes information that indicates a segment start time, segment end time and segment duration.

[0089] Feature extraction processing path 406- / can include an action for extracting behavioral segment parameters 466 from identified behavioral segments 462 of the input signal segment 458 to obtain behavioral segment parameters 466 corresponding to the behavioral segments 462. These behavioral segment parameters 466 may include statistical properties of the input signal segment 458 such as average amplitude, standard deviation, distribution, skewness, kurtosis, duration, peak and trough locations, number of zero crossings, or others. These parameters may also include waveform parameters for one or more of the signal components represented in the input signal segment 458 such as root mean square (RMS), peak-to-peak value, peak factor, shape factor, crest factor, impulse factor, or others.

[0090] Together, behavioral segment parameters 466 can describe the waveform of the behavioral segments 462 in a quantitative, parametric way which can be utilized to describe patterns in behavior through further analysis or as input to one or more machine learning models.

[0091] In various examples, feature extraction processing path 406- / can include a behavioral segment classification action 468 for classifying behavioral segments 462 based on the extracted behavioral segment parameters 466. By way of example, behavioral segment classification action 468 can be implemented using one or more machine learning models obtained from un-supervised or supervised machine learning algorithms. For example, characteristic patterns or anomalies in time-frequency-energy distribution data that is represented in the behavioral segments 462 and / or extractedbehavioral segment parameters 466 can serve as indicators of the real-time physical state within the subterranean formation 140, reflecting conditions that could precede the onset of significant seismic events or other atypical events. In this regard, behavioral segment classification action 468 can be configured to assign respective behavioral classifications to behavioral segments 466. In various examples, such behavioral classifications may be performed based on any combination or subset of the behavioral segments 462 (which include data from operational signal segment 430 and / or the decomposed signal segment 454) and the extracted behavioral segment parameters 466. The set of candidate classes that are assignable by segment classification action 468 are pre-selected based on the desired features to be extracted. In one example, behavioral segment classification action 468 is implemented using a machine learning model that has been configured to assign classifications that represent ground motion preceding induced seismic events. In such an example, each behavior segment 462 may be assigned a classification from a set of classification categories that include “stick-slip”, “stick”, “active sliding”, among other possible classifications.

[0092] The behavior segment classification generated by the behavioral segment classification action 468, together with the respective behavior segment parameters 466, can collectively form a set of extracted behavioral features 470- / for a behavioral segment 462, which can represent types of events occurring in the target subterranean formation 140 (e.g. hydraulic fracture propagation, sliding, etc.) along with the characteristics of the event (e.g. the duration, magnitude, etc.).

[0093] Referring again to FIG. 4A, the process 400 includes an event prediction operation 408 that is configured to map extracted behavioral features 470 to event predictions 480 that indicate a likelihood of one or more future events being induced by the hydraulic fracturing. Event prediction operation 408 can include multiple parallel prediction processing paths, with each path configured to process a unique group of extracted features 470 and make event predictions in respect of a respective type or types of event category(ies). In this regard, FIG. 4E illustrates an example of an eventprediction processing path 408-j that represents one of one or more event prediction paths supported by event prediction operation 408.

[0094] In the illustrated example, event prediction processing path 408-j includes an event probability prediction action 472 that is configured to output one or more probabilities indicating a likelihood of one or more types of atypical seismic events occurring, based on extracted behavioral features 470. In example implementations, event probability prediction action 472 may employ one or more machine learning models that utilize the extracted behavioral features to forecast the likelihood of induced atypical events. That forecast may, for example, take the form of an event probability 474 that provides percentage value indicating the likelihood of an event occurring. These models, which may include NN based models, may be trained using historical data from previous fracturing operations, which include documented outcomes and conditions. This training allows the models to learn the complex correlations between measured features and the occurrence of atypical events.

[0095] By way of example, FIG. 4F shows a flowchart of an event prediction model training process 482 that can be applied to obtain a trained event prediction model 492 that can be used to implement event probability prediction action 472. In the illustrated example, event prediction model training process 482 is provided with a training dataset 470T that includes behavioral features that have been extracted in respect of historical data and that have been labelled to indicate associated historical known events. The training dataset 470T can include multiple datasets, each being customized for different levels and types of training that can be performed as part of event prediction model training process 492. In one example, a base model pre-training action 484 is performed using a broad dataset for various events and various fracturing operations to provide a base ML model 485 that can predict the probability of a generic subterranean event. Event specific tuning 486 can then be applied to tune the base model 485 for specific event types (e.g., respective models that are specific to FDI event, CD events, IS events, etc., for each event prediction path 408-j). Further, region specific tuning 490can then be applied such that the event specific models 488 are each fine-tuned based on specific historic behavioral event data obtained from previous operations in the target subterranean formation, enhancing the model accuracy in predicting conditions relevant to the specific geological characteristics of the site. The trained models 492, tuned for specific events and regions, can then be used for real time inference in event prediction operation 408.

[0096] The above methodology can similarly be adapted to train the various other ML models that can be used to implement the operations of process 400.

[0097] Various model architectures are suitable for processing the extracted features, including but not limited to the Long Short-Term Memory (LSTMs) networks, a variation of Recurrent Neural Networks (RNNs). LSTMs are often used to analyze sequential data such as the behavioral segments discussed previously and can identify patterns such as the precursor patterns of behavior associated with IS events. LSTMs are also capable of handling sequences of variable length, particularly applicable to cases where different quantities of fracturing stages, operational segments, and behavioral segments are present. The neural network model may be trained on a plurality of historical data, including periods leading up to known events and periods without events. For example, in the context of induced seismicity, the NN may be trained to predict the “probability” that any seismic event will occur over a certain threshold magnitude or may be trained to predict the “probability” and magnitude of a seismic event.

[0098] Referring again to FIG. 4E, in some examples, the event prediction processing path 408-j can include an event risk level computation action 476 that is configured to convert the event probability 447 to a risk score (e.g., event risk level 478). The risk score may, for example, represent a level of hazard associated with the detected behavioral patterns, and may be tuned based on regional information, historical data, or other hazard considerations.

[0099] Referring again to FIG. 4A, in example implementations, the event predictions 480 (e.g., one or both of event probability(ies) 474 and event risk levels 478 computed by event prediction operation 408 are processed by one or more further analysis operations 410 that can function to advise operators and other parties of the hydraulic fracturing process and / or take actions. For example, the hydraulic fracturing process can include injecting a frac slurry comprising frac fluid and proppant through a wellbore into the target subterranean formation, and an analysis operation 410 can include adjusting operating parameters of the hydraulic fracturing in real time based on the event predictions 480 to reduce the likelihood of one or more future predicted events.

[0100] In some examples, analysis operation 410 can include using real-time stream of predicted event probabilities 474 in respect of behavioral segments together with historical data from previous stages of the same well, previous wells of the same pad, previous nearby wells, or other combinations of live and historical data to generate a historic risk assessment. In a further example, an analysis operation 410 can include using predicted event probabilities 474 to determine an event susceptibility score. For example, one or more algorithms and / or machine learning based models can be configured to perform a spatial analysis that includes determination of the event susceptibility of a well based on features of the well including but not limited to the region, subterranean formation the play, the depth, the stress regime, the lithology, nearby faults, historical event data, historic risk assessment data or others properties.

[0101] In some implementations, an event susceptibility score could be provided as a further input to the event risk level computation action 476 for calculating event risk level 478.

[0102] As explained in greater detail below, in some examples, analysis operation 410 can include advisory generation actions such as generating reports or user interface outputs that provide one or more of: (1 ) a live view of a risk assessment level associated with one more future subterranean events; (2) a live view representing spatiotemporal evolution of events in the subterranean formation; (3) monitoring offsetwell geo-spatiotemporal features to update risk assessments; and (4) characterizing geospatial extents of event(s) based on offset well geo-spatiotemporal features.

[0103] A description of subterranean characteristics that are represented in measurement data will now be described in greater detail to provide an understanding of the features that can be extracted from such data and the atypical event predictions that can be inferred from such data.

[0104] Induced Seismic Event Prediction

[0105] In example implementations, the process 400 can be applied to predict an atypical event such as induced seismic event (IS event) that exceeds threshold criteria. In such implementations, feature extraction processing operation 406 and event prediction operation 408 can include respective processing paths that are configured to enable prediction of an atypical IS event. An example will now be described wherein process 400 is configured to predict hydraulic fracturing induced seismic events that occur primarily a result of aseismic shear along weak fault planes. Hydraulic fracturing may induce bedding plane slip by fracturing and lubricating low-strength or pre-sheared layer boundaries. This may allow hydraulic connection to faults near the wellbore or in the far field, which, in this context, refers to faults located outside of the immediate fracture zone. The far field encompasses areas beyond the traditionally considered fracture zone, extending the impact of the hydraulic fracturing to more distant faults.This hydraulic connection can result in slip along both beds and faults and eventual fault rupture (e.g., an atypical induced seismic event such as an earthquake). This slip can produce a specific tremor signature in measured data.

[0106] By way of context, FIGs 6A and 6B conceptually represent a hydraulic fracturing system 100 interacting with subterranean formation 140. A pre-existing fault 650 applies a horizontal force (Snmax) on the stratigraphic layers (e.g., high stiffness and low stiffness layers) in the subterranean formation 140. In the illustrated example, an atypical induced seismic event 310 occurs in the far-field (e.g., 2-3+ km away from the borehole location of hydraulic fracturing system 100) during stimulation. By way ofexample, induced seismic event 310 may be a 4.6 magnitude IS event. Aseismic shear along weak planes (e.g., along pre-existing fault 650) may be the primary failure mechanism. Weak planes in this context include pre-existing fractures, pre-existing faults, rubblized zones, shear zones, stratigraphic contacts, bedding planes, or other heterogenous features in the subsurface. For example, when hydrofracturing fluid reaches a stratigraphic discontinuity within shale, the interface may act as a preferential failure plane which shears aseismically, driven by the high horizontal force Smnax, which allows fluid to propagate and reach the pre-existing fault plane 650. Over-pressurization of the fault plane 650 results in a reduction of effective normal stress, resulting in slip and rupture of the fault plane, thereby causing atypical seismic event 310.

[0107] In an example implementation, during the time leading up to atypical seismic event 310, there are few discrete seismic events. Rather, a constant slow slip occurs, that may have a distinct vibrational signature known as episodic tremor and slip (ETS). The ETS signature includes a long period and low frequency rumbling resultant of shear movement. The characterization of an ETS signature has the capacity to give an early warning for atypical IS events.

[0108] For purposes of illustrating an example of an ETS signature, FIG. 7, 8 and 9 provide sets of time plots that visually illustrate information that can be obtained from pressure data such as a time series of pressure measurements acquired at a wellhead during a hydraulic fracturing process. FIG. 7 and 8 illustrate a first set of plots 602, 604, 606 and 608 corresponding to a first time duration prior to atypical seismic event 310. FIG. 9 shows the first set of plots 602, 604, 606 (taken from FIG.6), along with a second set of plots 902, 904, 906 and a third set of plots 912, 914, 916. The second set of plots 902, 904, 906 correspond to a pressure signal acquired during a second duration leading up to atypical seismic event 310. The third set of plots 912, 914, 916 correspond to a pressure signal acquired during a third duration following the atypical seismic event 310.

[0109] Referring to FIGs. 6 and 7, plot 602 represents measured pressure data versus time. Plot 604 is a spectrogram plotting frequency v time, and may for example be obtained based on a Short-Time Fourier Transform (STFT) process that includes dividing a signal into short segments and applying the Fourier transform to each segment. The STFT process provides a compromise between time and frequency resolution and is suitable for signals whose frequency content changes over time but is less effective for highly nonstationary signals such a pressure measurement data.

[0110] Plot 606 is a plot of amplitude versus time for three intrinsic mode functions (IMF_1 , IMF_2 and IMF_3) that have been obtained by applying EMD to the pressure measurements represented in plot 602. The amplitude the IMFs is in pressure measurements units (e.g., physical units such as MPa as illustrated, but can also be other physical units such as psi). These physical units have physical meaning when compared to Fourier transform energy as represented in a spectrogram. The representation of IMF data in actual physical units can enable measurement data to be more effectively processed to identify underlying physical processes. Among other things, the use of physical unit based data may enable for more accurate normalization between different wells and may be able to account for friction, stage length and other shear plane variables.

[0111] Plot 608 is a plot of frequency versus time for sets instantaneous frequencies IF_1 , l_2, IF_3 that have been respectively computed for (IMF_1 , IMF_2 and IMF_3).

[0112] FIG. 7 shows spectrogram plot 604 together with band-specific IMF and IF components of plots 606 and 608. In particular: Plot 606_1 shows IMF_1 in the absence of IMF_2 and IMF_3; Plot 606_2 shows IMF_2 in the absence of IMF_1 and IMF_3; Plot 608_1 shows the set of instantaneous frequencies IF_1 in the absence of sets I F_2 and IF_3; and Plot 608_2 shows the set of instantaneous frequencies I F_2 in the absence of sets IF_1 and IF_3.

[0113] As represented in the example plots of FIG. 6 and 7, each IMF has its own Instantaneous Frequency (IF) function, which defines how the frequency changes overtime. Together, the IMFs capture the oscillatory modes present in the measured pressure signal. Each IMF may characterize a frequency interval as seen in the spectrogram, as illustrated in FIG. 7, enabling variable signals to be tracked over both frequency and time. In contrast to a spectrogram, the instantaneous frequency allows each IMF to represent signals which span multiple frequency intervals, allowing for the identification and characterization of non-linear and non-stationary signals. The decomposition of the pressure signal into IMFs and computation of sets of IFs for the respective IMF’s enables the tracking of variable signals over both frequency and time, enabling recognition of patterns in frequency behavior. Breakdown of the pressure signal into IMF components may allow specific frequency behaviors to be correlated to seismic activity (e.g., “typical” vs “atypical” behavior) and disregards noise by decomposing signals based on local extrema and oscillatory modes. In this context, “typical” behavior refers to the evolution of frequency behavior over time considered to be the norm for hydraulic fracturing, such as the development of resonant behavior as fractures develop. In contrast, “atypical” behavior refers to specific patterns temporally to IS events, such as those that occur before or after IS events.

[0114] As illustrated in FIG. 7, the behavior of different IMF’s can be quantified based on features extracted from the IMFs. By way of example In FIG. 7, the IMF_1 has a distinct gain-decay cycle as illustrated by arrows 702. IMF_1 and its corresponding set of IFs have a consistent cycle length (as indicated by line 706), and stable frequency (indicated by line 704). In contrast, IMF_2 has no well-defined schedule and has an unstable frequency (indicated by line 708). The decomposition of the signal into IMFs allows for the quantified behaviors to be characterized to identify the occurrence of temporal trends or patterns in behaviors. These patterns may be transitory in nature and are interpreted as different physical phenomena including but not limited to shear or tensile fracturing, shear sliding of planes, or induced seismic events.

[0115] Combinations of these patterns or changes in these patterns over time may precede seismic events, such that identifying them provides an indicator of a possibleimminent seismic event. For example, the identification of a stable stick-slip cycle may indicate a plane in the subsurface is undergoing periodic shear sliding. The acceleration or deceleration of this cyclic behavior may indicate changes in the sliding rate of such a plane. The increase or decrease in the amplitude of this behavior may indicate changes in the surface area of such a plane. Abrupt changes in the periodicity or rapid halting of this behavior may indicate that motion of the plane has stopped, referred to as a “stick” event. In contrast to stick-slip where a plane gradually accelerates and decelerates, a stick event is when a plane abruptly halts movement. A stuck plane may result in “fault locking”, which allows a fault plane to build up stress until eventual rupture, conventionally known as an induced seismic event.

[0116] The distinct gain-decay cycle 702, having a consistent cycle length 706, and stable frequency 704 are representative of episodic tremor and slip (ETS) signature occurring at stable intervals during active fracturing. In this example, the ETS signature can be observed from IMF_1 and its corresponding set of IFs but may not always be easily discernable from the raw pressure amplitude vs time data of plot 602 or the Fourier energy-based spectrogram data of plot 604.

[0117] With reference to FIG. 9, in the second set of plots 902, 904, 906 that correspond to the second duration leading up to atypical seismic event 310, the distinct gain-decay cycle 702 is absent, as the periodic stick and slide along the fault plane 650 has been replaced by a longer “active slide” duration followed by a long “stick” duration, culminating in atypical seismic event 310. The third set of plots 912, 914, 916 correspond to active sliding that occurs after the start of atypical seismic event 310. It will be noted from the respective amplitude scales of IMF plots 606, 906 and 916 that there is substantial amplitude variation between the three different durations. IMF_1 has a peak amplitude of: approximately 0.5 MPa during the cyclic stick-slip durations (plot 606); approximately 1.5 MPa during the duration that includes an pre-event active slide and major stick (plot 906); and approximately 2.5MPa during the active sliding after the commencement of atypical event 310 (plot 916).

[0118] Accordingly, in an example embodiment, a feature extraction and processing path 406-i of feature extraction operation 606 has been configured to receive as inputs signal segments 458 that have been extracted in respect of operational signal segments 430 that have been classified as “fracturing” segments. The feature extraction and processing path 406-i is configured to then extract and classify behavioral segments 462 that each conform to respective periods of “stick-slip”; “active-slide”; “stick” and “other”, resulting in a time-series of extracted behavioral features 470, with each feature representing one of “stick-slip”; “active-slide”; “stick” and “other”. In such example embodiment, a corresponding event prediction path 408-j has been configured to receive the time-series of extracted behavioral features 470 (e.g., a sequence including occurrence of “stick-slip”; “active-slide”; “stick” and “other”) and based on such timeseries, output an event probability 474 that an induced seismic (IS) event that exceeds a defined threshold is likely to occur.

[0119] In another example embodiment, a plurality of pressure signals and / or seismic wave signals have undergone a feature extraction and processing path 406-i configured to extract and classify behavioral segments 462 to identify respective representations. In such example embodiment, a corresponding event prediction path 408-j has been configured to receive a plurality of extracted behavioral features 470 and output an event probability 474 that an IS event that exceeds a defined threshold is likely to occur.

[0120] Multiple Well Cross-Correlation

[0121] In some example implementations, time-series measurements can be obtained simultaneously from multiple hydraulic fracturing wells that run adjacent each other in a subterranean formation, and information derived in real-time from the multiple wells can then be correlated to obtain more accurate predictions.

[0122] Measurements obtained from the offset wells may be processed in real-time in a similar way to the main well as discussed previously to obtain behavioral segments in respect of a primary well. Cross correlation techniques may be used to determinewhich wells, if any, are showing similar behavioral patterns to the main well which is being actively fractured. The existence of similar behavioral patterns, such as stick-slip behavior, which is aligned in time with the primary well, could indicate regional behaviors rather than localized behaviors. The spatial distribution and correlation strength of synchronized wells (that is, wells which show similar behavioral patterns at the same times) may give an indication as to the extent of regional motion in the subsurface. Lag times of behaviors between offset wells may give an indication of the approximate source location of movement. In contrast, the lack of synchronization or correlation between wells may indicate that a particular pattern is not regional and is instead localized around the wellbore. In one example implementation, correlations between wells may be used to augment the live data feed with features describing the level of correlation. In another instance, correlations between wells may be used to adjust the output probability of the NN to tune the level of risk. By way of example, FIG. 10A illustrates and example where raw pressure measurement data 602A, 602B and 6020 are obtained from respective pressure sensors located at the well heads of three parallel wellbores 112A, 112B, 1120 located in subterranean formation 140. The raw pressure measurement data 602A, 602B and 6020 are each respective subjected to EMD to provide respective IMF component data 606A, 606B and 606D and respective sets of IF data for the wellbores 112A, 112B, 1120.

[0123] In this regard, FIG. 10B illustrates an offset well analysis process 1000 according to example embodiments. A data collection operation 1002, similar to data collection operation 402 described above in respect of a primary well (e.g. wellbore 112A), is performed for one or more offset wells (e.g. wellbores 112B, 112C) that are located in the same geological formation as the primary well. Data collection operation 1002 generates sets of offset well operational pressure signal segments 430-0 for each of the respective offset wellbores that includes data that is similar to operational pressure signal segments 430 collected in respect of the primary wellbore. Decomposition and behavioral feature extraction operations 1004 (similar todecomposition and feature extraction operations 404, 406) are performed in respect of offset well operational pressure signal segments 430-0 for each of the respective offset wellbores. Data collection operation 1002 generates sets of offset well operational pressure signal segments 430-0 for each of the respective offset wellbores that includes data that is similar to operational pressure signal segments 430 collected in respect of the primary wellbore. In the illustrated example, the process 1000 includes a correlated feature extraction operation 1006 that is configured to receive behavioral features 470 generated in respect of the primary well and the offset well behavior features 470-0. Correlated feature extraction operation 1006 can include one or more processing algorithms and / or ML models that map the primary well behavioral features 470 and the offset well behavior features 470-0 to offset well correlation features 1008. In some examples, process 1000 can also include an offset well geospatial feature extraction operation 1016 that is configured to map the offset well correlation features 1008 together with well geospatial data embedded in geospatial features 1012 ( and in some examples, together with selected primary well behavioral features 470 and the offset well behavior features 470-0) to generate a set of offset well geospatial temporal features 1018. The mapping can be performed, for example, by one or more deterministic algorithm or ML model based processes. In some examples, geospatial features 1012 are extracted from geospatial location data 1009 of the primary well and offset well(s) using geospatial characterization operation 1010.

[0124] Induced Fracture Driven Interaction (FDI)

[0125] In some example implementations, feature extraction processing operation 406 and event prediction operation 408 can include respective processing paths that are configured to enable prediction are configured to enable prediction of the probability of a Fracture Driven Interaction (FDI) event based on the behavioral patterns in sensed pressure and / or seismic wave data. Conceptually, FDIs may occur as a result of fractures within the subsurface, allowing for fracturing fluid or pressure fluctuations fromone well to reach an offset well horizontally, vertically, or otherwise at distance from the main well. FDIs may occur as a result of planar bedding features or discontinuities failing in the subsurface. These failures may be more prevalent in regions with stress regimes which apply shear forces across bedding planes, which is a cause of induced seismic events. Specific behavioral patterns may be present in formations which indicate large scale shear failure and movement which results in FDIs through a similar process to those that result in IS events. The data preprocessing steps are analogous to those presented for the prediction or risk assessment of IS events, with the distinction of the training procedure for the machine learning / artificial intelligence / big data model. Similar procedures may be used as with offset well or multiple wells when monitoring for IS events to understand the regional extent of movement in the subsurface. Additional features may be computed which include the pressure derivative or others of the offset well(s). These features may specifically be used to assess the level of pressure change resulting from the main well, and the spatiotemporal relation of these changes. This may be combined with the cross-correlation analysis of offset wells to create augmented features for input into the NN. The NN may be trained on this combination of features and may accept these features as input to predict the probability of FDIs. In other example implementations, FDI susceptibility may be determined from a plurality of historical data, using similar features to those discussed for seismic susceptibility.

[0126] Induced Casing Deformation (CD)

[0127] In some example implementations the shifting subsurface may create movement in the active well or in offset wells where the casing or integrity of the wellbores themselves can deform partially or completely, restricting the flow of fluids from the reservoir to surface. In some example implementations, feature extraction processing operation 406 and event prediction operation 408 can include respective processing paths that are configured to enable prediction are configured to enable prediction of the probability of a CD event. One or more ML models within the processing paths can be fine-tuned to predict the probability of casing deformation asthe mode of failure is similar to that associated with an atypical IS event. Conceptually, casing deformation may occur as a result of the same shear along planar surfaces as discussed for induced seismic events and fracture driven interactions. In this context, the model may take any of the aforementioned features as input, and may be trained on similar historical datasets outlining known cases of casing deformation.

[0128] Fracture Growth Optimization

[0129] In some example implementations, feature extraction processing operation 406 and event prediction operation 408 can include respective processing paths that are configured to enable optimization of fracture effectiveness during hydraulic stimulation in real time. In such implementations, the system can be configured to detect, classify, and respond to behavioral patterns in decomposed pressure signals to dynamically guide fracturing operations. These behavioral patterns may be indicative of optimal or suboptimal stimulation conditions and can be leveraged to trigger real-time operator alerts, populate dynamic dashboards, or enable closed-loop control systems to automatically adjust operational parameters such as pump rate, fluid viscosity, or proppant concentration.

[0130] In one example embodiment, decomposed signal features such as instantaneous frequency, amplitude, or gain-decay trends from one or more intrinsic mode functions (IMFs) are monitored to identify conditions such as stable resonance (suggesting effective fracture propagation), damping (indicating energy loss or complexity), or sudden frequency shifts (associated with potential fracture screen-out or height growth). These features are classified using a hybrid model incorporating both deterministic logic and ML classifiers trained on historical stimulation data.

[0131] A corresponding optimization path 408 can be configured to receive behavioral indicators derived from feature extraction path 406 and output a real-time fracture quality signal. This signal may include an operator-facing alert, a visual dashboard recommendation, or a machine-readable control signal for automated adjustment. For example, a stable oscillation pattern may trigger a “continue” signal,while signs of energy loss or overpressure may prompt a reduction in pump rate or fluid volume.

[0132] In some example implementations, pressure and / or seismic wave signals from offset wells can also be decomposed and cross-correlated to assess regional fracture extension, interference, or adverse subsurface responses. Behavioral patterns associated with induced seismicity (IS) or casing deformation (CD), such as pre-event tremor signatures, sustained high-frequency oscillations, or sudden amplitude shifts, may be detected in one or more offset wells and used to augment the optimization model. In such implementations, the system can incorporate these cross-well indicators to inform spatially aware adjustments in real time, enabling proactive modifications to stimulation parameters to reduce the likelihood of unintended communication, asymmetrical fracture growth, or adverse geomechanical responses.

[0133] Stress Control

[0134] In some example implementations, feature extraction processing operation 406 can include a processing path configured to monitor stress-related indicators derived from decomposed signal features during hydraulic stimulation. Behavioral patterns extracted from intrinsic mode functions (IMFs), such as sustained amplitude growth, shifting frequency modes, or changes in gain-decay trends, may be indicative of evolving subsurface stress states. These extracted features can be used to infer increases in localized stress or the onset of stress redistribution, which may precede adverse outcomes such as fault reactivation, fracture containment loss, or casing deformation. The behavioral patterns may be monitored into the fall-off stage after shut in of the well.

[0135] In such implementations, the system may classify the observed stress conditions and determine whether mitigation is warranted. For example, elevated or unstable stress indicators may trigger a reduction in injection rate, a modification in fluid viscosity, or a temporary stage hold to allow pressure equilibration. In other embodiments, the classified stress state may be presented to the operator via a real-time dashboard or used to augment the prediction logic of event models such as those associated with induced seismicity or fracture-driven interactions. By identifying and responding to stress-related conditions early, the system may reduce operational risk and improve fracture control during treatment.

[0136] Real-Time Display

[0137] In example implementations, the control system 116 of hydraulic fracturing system 100 is configured to provide generate a graphical user interface (GUI) that includes a real-time display that indicates the predicted likelihoods of the one or more future events. An example of a GUI 1150 is shown in FIG. 11 . GUI 1150 can include multiple display sections, including for example: a live IMF display section 1152; induced atypical event display section 1154; well visualization section 1156; offset FDI Pressure Derivative section 1157 and stage summary section 1158.

[0138] Live IMF display section 1152 displays on-going spectrogram, IMF and IF plots that are obtained in respect of real-time pressure and / or seismic wave measurements.

[0139] Induced atypical event display section 1154 can for example visually display the real-time prediction results that are computed by control system 116 in respect of the real-time IMF and IF data that is shown in Live IMF display section 1152. In the illustrated examples, atypical IS event predictions are selected from a candidate set of classifications of “Low”; Medium”; and “High”. In FIG. 11 , atypical event display section 1154 includes a respective prediction display element 1160 for each wellbore that is subjected to pressure monitoring. Each display element 1160 is labelled to associate the element with a respective wellbore (e.g., “A” for wellbore 112A, “B” for wellbore 112B and so on). The color of each display element 1100 is elected based on the current prediction of an atypical IS event prediction being induced by the hydraulic fracturing that is being performed in respect of that wellbore. For example, a green colored display element 1160 in respect of “A” indicates that the present predicted risk of the fracing operations of wellbore 112A causing an atypical IS event is “LOW”; ayellow colored display element 1160 in respect of “B” indicates that the present predicted risk of the fracing operations of wellbore 112B causing an atypical IS event is “MEDIUM”; and a red colored display element 1160 in respect of “C” indicates that the present predicted risk of the fracing operations of wellbore 112C causing an atypical IS event is “HIGH”.

[0140] These predictions can be used to adjust hydraulic fracturing operations to reduce the medium and high risks. For example hydraulic fracturing fluid rates can be adjusted in the wellbores having “MEDIUM” or “HIGH” predicted risks. In at least some examples, control system 116 is configured to automatically cause risk-reducing actions to be taken to lower predicted risks.

[0141] Real-Time Operational Adjustment

[0142] In some example implementations, feature extraction processing operation 406 and event prediction operation 408 can include respective processing paths that are configured to enable real-time adjustment of hydraulic fracturing parameters based on predicted risks of one or more future events. The system may be configured to receive decomposed signal data and behavioral features in real time, and based on corresponding prediction outputs, dynamically adjust operational parameters such as pump rate, fluid viscosity, proppant concentration, or treatment pressure. In this manner, the system enables a proactive mitigation response that is tailored to evolving subsurface conditions detected during stimulation.

[0143] In example embodiments, the specific parameter adjustments may vary depending on the nature of the predicted event. For instance, when an elevated likelihood of induced seismicity (IS) is predicted, the system may initiate a reduction in pump rate or temporarily halt injection. For predicted fracture-driven interaction (FDI), an appropriate mitigation may involve a staged decrease in rate, modification of proppant concentration, or utilization of a diverter. In the case of predicted casing deformation (CD), the system may reduce treatment pressure or increase fluid viscosity to reduce mechanical loading. These risk-specific adjustments may be predefined, rule-based, or determined through a trained model that maps event likelihoods to appropriate operational responses.

[0144] In some implementations, the system may concurrently evaluate multiple risk predictions and apply an adjustment strategy that balances or prioritizes mitigation actions across different event types. For example, a high predicted risk of IS may take precedence over a moderate risk of FDI, resulting in a more aggressive down-regulation of stimulation parameters. The dynamic adjustment of operating conditions during treatment can reduce the likelihood or severity of adverse events, improve treatment continuity, and enhance overall wellbore integrity and reservoir access.

[0145] Alternatives to HHT

[0146] As noted above, in example implementations decomposition of the measured time-series signal into components that collectively represent nonstationary and nonlinear properties of the measurements and subsequent extraction of features that are representative of events occurring in the target subterranean formation can be performed using an HHT based process. However, these operations can be performed using other processing methods that enable nonstationary and nonlinear properties to be identified, including for example one or more of the following: 1. Wavelet Transform (WT): This is a powerful tool for analyzing signals whose frequency content changes over time. Wavelet transforms decompose a signal into components that occur at different times and scales (frequencies). 2. Empirical Mode Decomposition (EMD): Although EMD is a part of HHT, in some examples it can be used independently with features being extracted directly from the IMFs without computing IFs. As noted above, alternative embodiments may use Variational Mode Decomposition (VMD), an extension of EMD. Additional extensions or variations of mode decomposition techniques considered include but are not limited to: Ensemble Empirical Mode Decomposition (EEMD), 3D Empirical Mode Decomposition (3D EMD), Multivariate Variational Mode Decomposition (MVMD), Recursive Variational Mode Decomposition (RVMD), and others. 3. Wigner-Ville Distribution (WVD): This is a time-frequencyrepresentation that provides detailed information about the time-varying frequency content of a signal. 4. Time-Frequency Distributions (TFDs): Beyond WVD, there are other TFDs like the Cohen’s class and the affine class distributions, which aim to reduce the cross-term issues and provide adaptable resolutions. 5. Choi-Williams Distribution: This method, part of the Cohen class distributions, uses an exponential kernel to suppress cross terms in the time-frequency distribution, enhancing clarity in analyzing signals. 6. Matching Pursuit (MP): This algorithm decomposes a signal into a linear expansion of waveforms that are selected from a predefined dictionary of functions. 7. Adaptive Filtering: Techniques like the Kalman filter or adaptive noise canceling are used to track and analyze signals that vary in time. 8. Nonlinear Dynamical Systems Analysis: Techniques such as phase space reconstruction, Lyapunov exponents, and fractal dimensions are used primarily in physiological signal analysis and other areas dealing with complex systems behavior.

[0147] FIG. 12 is a computing system 1100 having a processor 1102 and memory 1104 storing code 1106 (e.g., logic, instructions, etc.) executed by the processor 1102. The computing system 1100 may be single computing device or a computer, a server, a desktop, a laptop, multiple computing devices or nodes, a distributed computing system, control system, and the like. The computing system 1100 may be local (at the wellbore or remote from the wellbore. Indeed, the computing system may represent multiple computing systems or devices across separate geographical locations. The computing system may be a component of a control system (e.g., 116 in FIG. 1 ). The processor 102 may be one or more processors, and may have one or more cores. The hardware processor(s) 102 may include a microprocessor, a central processing unit (CPU), graphic processing unit (GPU), or other circuitry. The memory 1104 may include volatile memory (e.g., cache, random access memory or RAM, etc.), nonvolatile memory (e.g., hard drive, solid-state drive, read-only memory or ROM, etc.), and firmware, and the like.

[0148] In operation, the computing system 1100 may receive measured pressure and / or seismic wave data originating from a respective sensors and also receive data from other sensors and controllers. The code 1106 may include an analyzer or analysis logic and a neural network when executed that directs the processor 1102 to generate predictions for the likelihoods of induced atypical events. The code 1106 may include an adjuster or controller which may give instructions when executed that direct the processor 1102 to specify a set point or adjust an operating parameter of the hydraulic fracturing system. The computing system 1100 is unconventional, for example, in that the computer can decompose measurement data to obtain non-linear and non- stationary properties and use that information to predict a likelihood of atypical induced events caused by the hydraulic fracturing. In this context, the computer is innovative with respective to accuracy and speed (real time). In addition, the technology of hydraulic fracturing is improved as the risk of inducing unwanted events can be diminished. Further, this innovative computing system may result in increased production of hydrocarbon (e.g., crude oil and natural gas) from a well, derisking the concern of triggering an unwanted event.

[0149] FIG. 12 is a block diagram depicting a tangible, non-transitory, computer (machine) readable medium 1200 to facilitate analysis and control of hydraulic fracturing. The computer-readable medium 1200 may be accessed by a processor 1112 over a computer interconnect 1204. The processor 1112 may be a controller, a control system processor, a controller processor, a computing system processor, a server processor, a compute-node processor, a workstation processor, a distributed- computing system processor, a remote computing device processor, or other processor. The tangible, non-transitory computer-readable medium 1200 may include executable instructions or code to direct the processor 1112 to perform the operations of the techniques described herein, such as to receive, determine, calculate, or utilize an indicator of the amount of shear or tensile fracturing, a guidance pressure, and a rock stress prediction of an atypical subterranean event and in some examples, adjust acontroller or specify a set point for operation of a hydraulic fracturing system. The various executed code components discussed herein may be stored on the tangible, non-transitory computer-readable medium 1200, as indicated in FIG. 12. For example, an analysis code 1206 may include executable instructions to direct the processor 1112 to as to receive, determine, calculate, or utilize an indicator of predictions of one or more types of atypical events that may be induced by hydraulic fracturing activities. Adjust code 1208 may include executable instructions to direct the processor to specify a set point or adjust an operating parameter of the hydraulic fracturing system. It should be understood that any number of additional executable code components not shown in FIG. 12 may be included within the tangible non-transitory computer-readable medium 1200 depending on the application.

[0150] A number of implementations have been described. Nevertheless, it will be understood that various modifications may be made without departing from the spirit and scope of the disclosure.

Claims

WHAT IS CLAIMED IS:1 . A method of predicting events during hydraulic fracturing of a target subterranean formation, comprising: obtaining a time-series of measurements associated with the hydraulic fracturing; decomposing the measurements into a plurality of respective components that collectively represent nonstationary and nonlinear properties of the measurements; extracting respective features from the respective components; and predicting, based on the extracted features, a likelihood of one or more future events being induced by the hydraulic fracturing.

2. The method of claim 1 wherein decomposing the measurements into a plurality of respective components comprises performing a decomposition in respect of the measurements, wherein each component comprises a respective intrinsic mode function (IMF).

3. The method of claim 4 wherein the decomposition comprises an empirical mode decomposition (EMD).

4. The method of claim 4 wherein the decomposition comprises a variational mode decomposition (VMD) and each component comprises a respective band limited IMF.

5. The method of claim 2 wherein extracting the respective features from the respective components comprises computing an instantaneous frequency (IF) for each IMF.

6. The method of claim 5 wherein computing the instantaneous frequency (IF) for each IMF comprises applying a Hilbert transform to each IMF.

7. The method of claim 1 wherein decomposing the measurements into a plurality of respective components and extracting the respective features collectively comprise performing a Hilbert-Huang Transform (HHT).

8. The method of claim 1 wherein extracting respective features comprises obtaining time-frequency-energy distribution data based on the respective components.

9. The method of claim 5 wherein extracting respective features comprises using one or more machine learning models to extract behavioral features based on one or both of the IF or IMF.

10. The method of claim 1 wherein predicting the likelihood of one or more future events being induced by the hydraulic fracturing comprises using one or more machine learning models to map the extracted features to respective predictions of the one or more future events.11 . The method of claim 10 comprising training the one or more machine learning models based on historic data obtained in respect of historic hydraulic fracturing of one or more subterranean formations.

12. The method of claim 11 wherein the one or more machine learning models comprises deep neural networks.

13. The method of claim 11 wherein training the one or more machine learning models comprises pre-training a first machine learning model of the one or more machine learning models based on historic data obtained in respect of the historic hydraulic fracturing of the one or more subterranean formations other than the targetsubterranean formation, and fine-tuning the first machine learning model based on historic data obtained in respect of historic hydraulic fracturing of target subterranean formation.

14. The method of claim 1 wherein the one or more future events is selected from a group comprising: events that are expected to cause environmental or structural damage to or in the vicinity of the target subterranean formation; and events that can adversely impact extraction of a hydrocarbon through the hydraulic fracturing.

15. The method of claim 1 wherein the one or more future events are selected from a group comprising: an induced seismicity (IS) event that exceeds threshold event criteria; a fracture driven interaction (FDI) between different wells; a casing deformation; and a plug slip.

16. The method of claim 1 wherein the hydraulic fracturing comprises: injecting a frac slurry comprising frac fluid and proppant through a wellbore into the target subterranean formation; adjusting operating parameters of the hydraulic fracturing in real time based on the predicting to reduce the likelihood of the one or more future events.

17. The method of claim 1 comprising generating a real-time display that indicates the predicted likelihoods of the one or more future events.

18. The method claim 1 wherein: obtaining the time-series of measurements associated with the hydraulic fracturing is performed in respect of a first wellbore in the target subterranean formation;the method further comprises obtaining further time-series of measurements in respect of one or more further wellbores in the target subterranean formation; wherein: decomposing the measurements into a plurality of respective components is performed in respect of each of the measurements from the first wellbore and the further time-series of measurements from the one or more further wellbores; and the predicting is based on extracted features from the respective components from both the first wellbore and the one or more further wellbores, a likelihood of one or more future events being induced by the hydraulic fracturing.

19. The method of claim 1 , further comprising: analyzing the extracted features to assess fracture effectiveness during the hydraulic fracturing; and based on the assessed fracture effectiveness, generating an optimization signal to guide stimulation operations in real time, wherein the optimization signal comprises one or more of:(i) a visual indicator displayed on an operator interface,(ii) a real-time alert recommending an adjustment to one or more operational parameters, and(iii) a machine-readable control signal for automatic adjustment of one or more of a pump rate, fluid viscosity, or proppant concentration.

20. The method of any one of claims 1 to 19 wherein the time-series of measurements comprises a time-series of pressure measurements.21 . The method of claim 20 wherein the time-series of pressure measurements are obtained at a wellbore head associated with the hydraulic fracturing.

22. The method of any one of claims 1 to 19 wherein the time-series of measurements comprises a time-series of seismic wave measurements.

23. A computer implemented system configured for performing the method of any one of claims 1 to 19.

24. The system of claim 23 wherein the time-series of measurements comprises a time-series of pressure measurements.

25. The system of claim 23 wherein the time-series of measurements comprises a time-series of seismic wave measurements.

Images