Intelligent early warning method and system combining fracturing big data and PINN algorithm

Abstract

This invention proposes an intelligent early warning method and system combining fracturing big data and the PINN algorithm, belonging to the field of intelligent prediction technology for drilling fracturing; it includes the following steps: S1: Dynamic envelope surface modeling based on time-varying manifold learning to obtain the pressure envelope surface model of the block; S2: Constructing a spatiotemporally adaptive physical information neural network (PINN), training the PINN using a meta-learning algorithm and spatiotemporal domain decomposition to obtain the trained PINN; S3: Predicting the construction pressure after a period of time using the trained PINN to obtain the predicted construction pressure value; comparing the predicted construction pressure value with the center line of the envelope surface corresponding to the pressure envelope surface model of the block, and outputting the comparison result; S4: Outputting an alarm signal of the corresponding level based on the comparison result.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction technology for drilling fracturing, and in particular to an intelligent early warning method and system that combines fracturing big data with the PINN algorithm. Background Technology

[0002] After an oil well reaches a certain production stage, its productivity and permeability decrease. To enhance oil discharge capacity, fracturing technology is needed to create fractures in the formation using external forces, thereby increasing oil and gas production. However, during fracturing operations, the heterogeneity of the formation and the influence of complex factors such as discharge rate, sand concentration, and fracturing fluid properties may cause sand plugging, which can lead to an abnormal increase in operating pressure.

[0003] Conventional methods for preventing sand blockage involve on-site personnel monitoring fracturing instruments. This approach is not only prone to accuracy issues due to the experience of the personnel, but also requires sustained concentration over several hours. To mitigate the economic losses associated with sand blockage, this paper proposes an intelligent early warning method that combines fracturing big data with the PINN algorithm. By integrating data acquisition data and using the PINN model to predict the construction pressure after a certain period, the method assesses whether the pressure changes meet actual requirements, issues alarm signals of varying levels, and takes corresponding measures. This transforms real-time monitoring into proactive prediction, significantly improving the automation level of construction pressure prediction. Summary of the Invention

[0004] In view of this, the present invention proposes an intelligent early warning method and system that combines fracturing big data and the PINN algorithm to improve the problems of lagging fracturing pressure early warning, lack of physical constraints in existing data-driven models, and poor real-time performance of alarms.

[0005] On the one hand, this invention provides an intelligent early warning method that combines fracturing big data with the PINN algorithm, including the following steps: S1: Dynamic envelope surface modeling based on time-varying manifold learning to obtain the pressure envelope surface model of the briquette; S2: Construct the spatiotemporally adaptive physical information neural network PINN, and train the physical information neural network PINN by combining meta-learning algorithm and spatiotemporal domain decomposition to obtain the trained physical information neural network PINN. S3: The trained physical information neural network PINN is used to predict the construction pressure after a period of time, and the predicted construction pressure value is obtained; the difference between the predicted construction pressure value and the center line of the envelope surface corresponding to the pressure envelope surface model of the briquette is compared, and the comparison result is output. S4: Output alarm signals of the corresponding level based on the comparison results.

[0006] Based on the above technical solution, preferably, step S1 includes: acquiring construction pressure, discharge rate, and sand concentration parameters; supplementing missing values ​​and removing outliers; then standardizing the parameters to adjust them to values ​​within the range of [0, 1]; performing feature encoding in a fixed order to obtain several sets of feature vectors with the same format; and based on the feature vectors, unifying the dynamic spatiotemporal structure into a low-dimensional manifold through a custom time-varying distance metric, quantifying uncertainty on the low-dimensional manifold, and constructing a pressure envelope surface function. E ( x , t ), in parentheses x The input characteristic vector of the envelope surface. t For a specified time.

[0007] Preferably, step S2 includes: Based on the physical process of fracturing, the governing equations are set; In a unified spatiotemporal domain, a physical information neural network (PINN) is constructed, and the total loss function is defined based on the approximate solution of the control equation and the neural network parameters. The meta-learning algorithm is applied to the physical information neural network PINN to train the physical information neural network PINN; Spatiotemporal decomposition is performed, dividing the complete spatiotemporal domain into temporal and spatial decompositions. Temporal decomposition involves breaking down the construction phase according to time decomposition criteria. m Spatial decomposition, which involves dividing a three-dimensional spatial region into different time slices according to spatial decomposition criteria, is a process of dividing the spatial region into... n A spatial region, m Different time slices and n These spatial regions together constitute several spatiotemporal subdomains. For each spatiotemporal subdomain Construct a sub-PINN, and each sub-PINN has a corresponding local loss function; Each obtained sub-PINN is subjected to personalized federated learning until the convergence condition is met.

[0008] Preferably, the time decomposition criterion is as follows: if the difference between the rate of change of the injected pressure and the rate of change of the pressure in the current time slice continuously exceeds a set rate of change threshold, it is determined that a phase transition has occurred and a new time slice needs to be created. Subsequent moments will be assigned to the new time slice; otherwise, no new time slice is created.

[0009] A further preferred spatial decomposition criterion is to obtain spatial points separately. xThe magnitude of the gradient vector of the predicted pressure at point 0 and spatial points x Predicted crack width at point 0 Establish a spatial region classification function and obtain the corresponding classification region number: when spatial point x The magnitude of the gradient vector of the predicted pressure at point 0 If the pressure gradient exceeds the threshold, return space number 1; if the space point... x Predicted crack width at point 0 If the width exceeds the effective opening threshold of the crack, return to space number 2; otherwise, return to space number 3.

[0010] Further optimization involves performing personalized federated learning on each of the obtained sub-PINNs, enabling nodes to... s The corresponding set of spacetime subdomains is S s Each node corresponds to at least one spatiotemporal subdomain; set globally shared parameters. and local personalized parameters The local loss function is defined as the sum of the local loss functions of all spatiotemporal subdomains corresponding to the local node, plus a regularization term. An alternating optimization approach is adopted, first fixing the globally shared parameters. Each node optimizes its corresponding local personalized parameters. To achieve node adaptation to local data, the local loss function is minimized, and then all local personalized parameters are fixed. Calculate all local personalized parameters Use the average value to update the globally shared parameters Repeated iterations R In each round, if the following convergence criteria are met, the sub-PINN is considered convergent and training is stopped: 1) Simultaneously satisfying: the absolute change of the globally shared parameters is less than the global convergence threshold, and the average change of the local personalized parameters is less than the personalized convergence threshold, and maintaining this... P Each round, P < R 2) Use only globally shared parameters A new global network is initialized. In the evaluation set of this new global network, the global loss is constructed and calculated by referring to the content of the local loss function corresponding to each sub-PINN. recent assessment Q Global loss during rounds of iteration The decrease was always less than the set threshold. , Q < R .

[0011] More preferably, in step S3, the use of a trained physical information neural network (PINN) to predict the construction pressure after a certain period of time, to obtain the predicted construction pressure value, defines the prediction time point. t pred Current time t 0 and the forecast period The sum of these features defines the input feature vector of the trained physical information neural network PINN as follows: Xin Output the location at the predicted time point t pred The pressure prediction set is used to obtain the maximum predicted pressure from the pressure prediction set. Predicted average pressure .

[0012] Further preferably, the step S3, which compares the predicted construction pressure value with the center line of the envelope surface corresponding to the pressure envelope surface model of the compaction block, and outputs the comparison result, is derived from the input feature vector. Xin Extract the envelope surface input feature vector x For each predicted point location x pi Query the pressure envelope surface function E ( x , t ), predicting the time point t pred and predicted point location x pi Substituting these values, we obtain the expected median pressure value at the predicted point location; from the expected median pressure values ​​at the predicted point locations, we extract the centerline value corresponding to the predicted point location with the largest median pressure value, and define it as... p center,max The centerline value of the mean pressure is defined as follows: p center,avg Based on maximum predicted pressure Centerline value corresponding to the predicted point with the largest median pressure value p center,max The difference is the maximum pressure absolute difference. Based on predicted average pressure Centerline value of average pressure p center,avg The difference is the absolute difference in mean pressure. ; Centerline value corresponding to the prediction point with the maximum absolute pressure difference and the median pressure value. p center,max The ratio of the maximum pressure relative difference Based on the absolute difference of mean pressure and the centerline value of mean pressure p center,avgThe ratio of the average pressure relative difference ; Based on the center line of the envelope during the prediction period The change yields the rate of change of the centerline at the point of maximum pressure. and the rate of change of the centerline of the mean pressure The predicted rate of change of the maximum pressure is obtained based on the maximum predicted pressure and the measured pressure value at the initial moment. The predicted rate of change of average pressure is obtained by comparing the predicted average pressure with the average pressure at the initial moment; the predicted rate of change is based on the maximum pressure. and the rate of change of the centerline of the point of maximum pressure The difference yields the predicted rate of change at the point of maximum pressure. Based on the predicted rate of change of mean pressure and the rate of change of the centerline of the mean pressure The difference yields the predicted rate of change of mean pressure. .

[0013] Furthermore, step S4 is further optimized by defining the following preset conditions: 1) Absolute difference of maximum pressure Not exceeding 5 MPa, or the absolute difference in mean pressure. 1) Not exceeding 3 MPa; 2) Relative difference in maximum pressure Relative difference between average pressure and average pressure All not exceeding 15%; 3) Centerline change rate and the rate of change of the centerline of the mean pressure All are located in the range of [-0.5, 1.0] MPa / min; 4) Predicted rate of change of the maximum pressure point Predicted rate of change of mean pressure All are located in the range of [-0.4, 0.4] MPa / min; When all the above conditions are met, a level 0 alarm signal is output. If any one condition is not met, a level 1 alarm signal is issued. If two or more conditions are not met, a level 2 alarm signal is issued.

[0014] On the other hand, the present invention provides an intelligent early warning system combining fracturing big data and the PINN algorithm to implement the above-mentioned method, including: The briquetting pressure envelope surface model building unit obtains the briquetting pressure envelope surface model by dynamic envelope surface modeling based on time-varying manifold learning, and maps the obtained input characteristic vector-specified time pair to different quantiles of the pressure distribution. The Physical Information Neural Network (PINN) is constructed and trained using a unit that constructs a spatiotemporally adaptive PINN. The PINN is then trained by combining a meta-learning algorithm and spatiotemporal domain decomposition to obtain the trained PINN. The prediction and comparison unit is used to predict the construction pressure after a period of time by applying the trained physical information neural network PINN to obtain the predicted construction pressure value; the predicted construction pressure value is compared with the center line of the envelope surface corresponding to the pressure envelope surface model of the briquette, and the comparison result is output. The alarm output unit outputs an alarm signal of the corresponding level based on the comparison result output by the prediction comparison unit.

[0015] The intelligent early warning method combining fracturing big data and the PINN algorithm provided by this invention has the following advantages compared with the prior art: 1. This invention constructs and trains a physical information neural network (PINN) to ensure that the prediction conforms to the laws of fluid mechanics and rock mechanics, thereby predicting the construction pressure in the future. The envelope model is based on big data statistics to ensure the reliability of the early warning benchmark. By comparing the predicted construction pressure with the center line of the envelope, it achieves intelligent early warning in advance rather than alarm after the fact.

[0016] 2. The Physical Information Neural Network (PINN) is trained using the NAML meta-learning approach, which allows for rapid adaptation of initial parameters. This enables quick adaptation for new wells with limited training data, mitigating the need for extensive historical data for learning. Through spatiotemporal decomposition and federated learning, the complex spatiotemporal domain is divided into multiple subdomains for parallel training, improving training efficiency while maintaining accuracy. By establishing temporal and spatial decomposition criteria, precise time-slice partitioning and spatial decomposition are achieved. During the federated learning phase, global parameters can share physical laws, while local parameters better adapt to the characteristics of local regions.

[0017] 3. Combining the prediction results of the physical information neural network PINN with the dynamic adaptive early warning benchmark provided by the envelope model as a parallel design, a two-layer verification is achieved; by calculating whether the number of judgment values ​​meets the preset conditions, the corresponding level of alarm signal is output to achieve graded early warning output. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram illustrating the method steps of the intelligent early warning method and system combining fracturing big data and the PINN algorithm of the present invention.

[0020] Figure 2 This invention describes the workflow of an intelligent early warning method and system that combines fracturing big data with the PINN algorithm. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] Traditional methods rely on historical statistical thresholds, which cannot provide early warnings of pressure anomalies in future periods; they can only issue alerts after the anomalies occur. Existing data-driven models lack physical constraints and are unreliable in data-sparse regions. On the other hand, pure physical simulations are computationally inefficient and have poor real-time performance. The physical mechanisms of fracturing operations often differ significantly at different stages and in different spatial regions, making it difficult for a single model to accurately describe the entire spatiotemporal process.

[0023] In view of this, such as Figure 1 As shown, on the one hand, the present invention provides an intelligent early warning method combining fracturing big data and the PINN algorithm, including the following steps: S1: Dynamic envelope surface modeling based on time-varying manifold learning to obtain the pressure envelope surface model of the briquette.

[0024] Step S1 involves: acquiring construction pressure, discharge rate, and sand concentration parameters; supplementing missing values ​​and removing outliers; then standardizing the parameters to ensure they are within the range of [0, 1]; and finally encoding them in a fixed order to obtain several sets of feature vectors with the same format. x i Based on feature vectors, a custom time-varying distance metric is used to unify the dynamic spatiotemporal structure into a low-dimensional manifold. Uncertainty quantification is then performed on the low-dimensional manifold to construct a pressure envelope surface function. E ( x , t ), in parentheses x The input characteristic vector of the envelope surface. t For a specified time; Let the historical dataset be ,in x i for d 3D eigenvectors ti For historical time, p i These are pressure observations; the sample size of the historical dataset is... N , i and j Define a time-weighted distance metric function for the sequence numbers of different samples in the historical dataset. , To calculate the Euclidean distance operation, used to represent spatial differences, Used to describe temporal similarity γ and β These are hyperparameters, all taking values ​​greater than 0 real numbers; define the manifold embedding function. , , To be consistent with the sample i and j corresponding m Low-dimensional manifold coordinates m < d The optimization objective of real-time rheology is defined as minimizing the total error between the pairwise Euclidean distance in the low-dimensional space and the original high-dimensional distance metric function: In one embodiment, the feature vector x i The elements within are distance to the wellbore, vertical depth, local permeability, local Young's modulus, minimum horizontal principal stress, and current construction discharge rate.

[0025] Low-dimensional manifold coordinates Transpose and time t i By splicing them together, we can obtain m +1 dimension yields the joint state-time feature vector Construct the conditional quantile function: , The target quantile; For kernel functions; h For bandwidth parameters; Ii is the empirical quantile; II is the indicator function, which has a value of 1 when the condition is met and a value of 0 when the condition is not met; the envelope function is... , .

[0026] In this embodiment, the target quantiles are selected as 0.05, 0.5, and 0.95; the kernel function is a Gaussian kernel function; here... v and v i These correspond to the latest and historical points, respectively. The empirical quantile is the median quantile with an index τ of 0.5. Bandwidth parameter hThe value range is [0, 1]. In this embodiment, the envelope surface function includes three dynamic functions, corresponding to the subscripts τ of the conditional quantile function as the 0.05 quantile, 0.5 quantile, and 0.95 quantile, respectively. The first and third functions represent the conservative lower and upper bounds of the pressure observations, respectively.

[0027] S2: Construct the spatiotemporally adaptive physical information neural network PINN, and train the physical information neural network PINN by combining meta-learning algorithm and spatiotemporal domain decomposition to obtain the trained physical information neural network PINN.

[0028] The content of step S2 is as follows: S21: Based on the fracturing physical process, the governing equations are set; Assume the fracturing process can be represented by the following differential equation: ,in ρ Let ρ be the density of the fracturing fluid, and v be the velocity field. Let ∠ be the source and sink term, ∠ be the gradient; the momentum equation is: , For pressure field, μ For fluid viscosity, Additional force generated by the proppant, x 0 represents spatial coordinates; the crack propagation equation is: , , w v is the crack width. f The fluid velocity within the crack, For filtration rate, υ Poisson's ratio, E For Young's modulus, The minimum horizontal principal stress is used here; the governing equations of the fracturing physical process are constructed based on the content of physical constraints, which will be used for further modeling in subsequent steps.

[0029] S22: Construct a physical information neural network (PINN) in a unified spatiotemporal domain, and define the total loss function based on the approximate solution of the control equation and the neural network parameters.

[0030] Let the approximate solution of the physical information neural network PINN be: , , , represented as As input, with neural network parameters θ As weights, output the predicted stress value. Predicted speed values and crack width prediction value Physical loss is defined as , to define the governing equations Ncoll One configuration point, A For the serial number of the configuration point, The residual operator for each configuration point; the data loss is defined as... ,definition N data One observation data point, B The sequence number of the observed data points. Let be the observed pressure value corresponding to the observed data point; define the conditional loss as . Select at the boundary and initial time N bc One sampling point, Indicates the first C The residual operators for the sampling points at each boundary and initial time are used; the physical loss, data loss, and conditional loss are combined, and the total loss is obtained by weighting. , This is the loss weight. .

[0031] Physical loss aims to make the predicted solution satisfy the physical laws of the system as much as possible; data loss aims to make the neural network's prediction match the actual observed data; and conditional loss aims to force the neural network's predicted solution to satisfy the boundary conditions and initial conditions.

[0032] S23: Apply the meta-learning algorithm to the physical information neural network PINN to train the physical information neural network PINN.

[0033] Define the stratigraphic dataset as , For the first I Data from the well, including the training set and verification set , for M The sequence number of each well, and the geological parameters of different wells. φ I Not exactly the same, geological parameters φ I As input, it is used in the control equations and total loss. In this embodiment, the geological parameters include the first... I The parameters of the well include permeability, elastic modulus, and geostress. The entire training process is divided into two nested loops: an inner loop and an outer loop.

[0034] Inner loop: Define the meta-learning problem as obtaining the meta-parameters for the optimal initial weights of the neural network. For each well, from the current meta-parameters start, Replace the parameters in the governing equations with the geological parameters of the well. φI Apply the current well's training set Define the inner loop loss The gradient update rule is as follows: , a For the inner loop learning rate, Neural network parameters θ Gradient, perform K Step gradient descent, update network parameters, and calculate adapted parameters. , Let the sequence number of the inner loop step be the adjusted parameter. As a temporary variable for output; the inner loop does not actually update the meta-parameters. .

[0035] External circulation: Evaluate and update meta-parameters for the performance of all wells after simulation adaptation. In the verification set of each well Upgraded temporary variables The performance is summed, and the element loss is calculated. Calculate the meta-loss against the meta-parameters The gradient is ,in Using the outer loop learning rate Update meta parameters Outer loop learning rate Less than the inner loop learning rate.

[0036] For example, the trained meta-parameters The parameters of the new well to be tested are directly derived from the trained meta-parameters. start, Perform the operation using the reference inner loop method. L Gradient descent step by step, , Steps L < K , This is data from a new well; the parameters after the new well has been adapted are... Since only a few updates are needed to achieve good results, the data requirements and training time for new wells can be reduced.

[0037] S24: Perform spatiotemporal domain decomposition. The complete spatiotemporal domain is decomposed into temporal and spatial components. Temporal decomposition involves breaking down the construction phase according to time decomposition criteria. m Spatial decomposition, which involves dividing a three-dimensional spatial region into different time slices according to spatial decomposition criteria, is a process of dividing the spatial region into... n A spatial region, m Different time slices and n These spatial regions together constitute several spatiotemporal subdomains. For each spatiotemporal subdomain Construct a sub-PINN, and each sub-PINN has a corresponding local loss function.

[0038] Time breakdown is the process of dividing the construction phases Decomposed according to the time decomposition criteria into m Different time slices, , These are the sequence numbers of the time slices, with the start time of the first time slice corresponding to the start time of the construction phase. The end time of the last time slice corresponds to the end time of the construction phase. Spatial decomposition is the process of decomposing a three-dimensional spatial region Ω into its constituent parts according to spatial decomposition criteria. n Each spatial region Ω n , , After complete spatiotemporal domain decomposition, several spatiotemporal subdomains are obtained. .

[0039] The time decomposition criterion mentioned here is: , For the current moment t Injection pressure rate of change, For the first m Representative pressure change rate for an existing time slice, The time decomposition criterion represents the decision function. If the difference between the rate of change of the injected pressure and the rate of change of pressure in the current time slice continuously exceeds a set rate of change threshold (where the threshold can be a multiple of the standard deviation of the slope change within the same stage, kt), a stage transition is determined, requiring the creation of a new time slice, and subsequent times will be assigned to this new time slice. Otherwise, no new time slice is created. The multiple kt is related to the construction stage: kt is 1 for the pre-fluidization stage, 1.2 for the sand-carrying stage, 1.5 for the high sand ratio stage, and 0.8 for the displacement stage. The formula for the time decomposition criterion monitors the slope of the construction pressure curve; a significant change in slope indicates the start of a new construction stage, potentially requiring the assignment of a new time slice.

[0040] The spatial decomposition criterion is to obtain spatial points separately. x The magnitude of the gradient vector of the predicted pressure at point 0 and spatial points x Predicted crack width at point 0 Establish a spatial region classification function and obtain the corresponding classification region number. , The threshold value for the pressure gradient is ±2 standard deviations of the average gradient. The effective opening width threshold of the crack is defined as 0.1-1.0 mm. This is a spatial region classification function that returns the spatial region number 1, 2, or 3.

[0041] Each spatiotemporal subdomain after decomposition Construct a sub-PINN, and the local loss function of each sub-PINN is: The physical loss of the sub-PINN is obtained by calculating the residuals of the sampling configuration points within the spatiotemporal subdomain. Spacetime subdomain The inner part defines the governing equations. One configuration point, These are the spatiotemporal subdomain local physics predictions and local network references calculated by the sub-PINN, respectively; the data loss of the sub-PINN is... Spacetime subdomain Internal definition N data 100 observation data points, here This corresponds to the observation pressure value of the observation data points within the spatiotemporal subdomain; the interface loss of the subPINN is... , and These represent the prediction output of the current spatiotemporal subdomain at a certain point on the interface and the prediction output of the neighboring spatiotemporal subdomains at the same point, respectively. The summation is obtained by iterating through and summing all neighboring spatiotemporal subdomains adjacent to the current spatiotemporal subdomain.

[0042] S25: Perform personalized federated learning on each obtained sub-PINN until the convergence condition is met.

[0043] Personalized federated learning of each obtained sub-PINN is what enables nodes to... s The corresponding set of spacetime subdomains is S s Each node corresponds to at least one spatiotemporal subdomain; set globally shared parameters. and local personalized parameters Introducing a local model , representing a node s The output is determined by both globally shared parameters and personalized parameters; the local loss function is defined as... ,in This corresponds to the sum of the local loss functions of all spatiotemporal subdomains corresponding to the local node. For regularization terms, λ For hyperparameters; an alternating optimization approach is adopted, first fixing the globally shared parameters. Each node optimizes its corresponding local personalized parameters. To minimize local loss This enables nodes to adapt to local data; Then fix all local personalization parameters. Calculate all local personalized parameters Use the average value to update the globally shared parameters Repeated iterations R round, number of iteration rounds The sub-PINN is considered convergent and training is stopped when the following convergence criteria are met: 1) Simultaneously satisfying: the absolute change of the globally shared parameters is less than the global convergence threshold. Furthermore, the average change in local personalized parameters is less than the personalized convergence threshold. and maintain P Each round, The first r Wheel and First r -1 round of iterations of globally shared parameters The first r Wheel and First r Local personalized parameters during the -1st iteration For the rate of change of the globally shared parameter, For the average rate of change of personalized parameters, P < R 2) Use only globally shared parameters A new global network is initialized. In the evaluation set of this new global network, the global loss is constructed and calculated by referring to the content of the local loss function corresponding to each sub-PINN. recent assessment Q Global loss during rounds of iteration The decrease was always less than the set threshold. , Q < R .

[0044] Among them, the global convergence threshold The value range is from 1e-5 to 1e-4; the convergence threshold of the local personalized parameter. The value range is from 1e-4 to 1e-3; set a threshold. The value range is from 1e-4 to 1e-3.

[0045] Personalized federated learning allows each node to retain its local characteristics while sharing global knowledge.

[0046] S3: The trained physical information neural network PINN is used to predict the construction pressure after a period of time, and the predicted construction pressure value is obtained; the difference between the predicted construction pressure value and the center line of the envelope surface corresponding to the pressure envelope surface model of the briquette is compared, and the comparison result is output.

[0047] Among these methods, a trained Physical Information Neural Network (PINN) is used to predict construction pressure over a certain period of time, thus obtaining the predicted construction pressure value, which defines the prediction time point. t pred Current time t 0 and the forecast period The sum of these features defines the input feature vector of the trained physical information neural network PINN as follows: Xin Output the location at the predicted time point t pred The pressure prediction set is used to obtain the maximum predicted pressure from the pressure prediction set. Predicted average pressure Forecast period It is usually no less than 5 minutes, such as 5 minutes, 15 minutes, 30 minutes, etc.

[0048] In one embodiment, the input feature vector Xin The information includes, in order: spatial coordinates, predicted time, current construction parameters, historical pressure time series, and geomechanical parameters.

[0049] Then, the predicted construction pressure value is compared with the center line of the envelope surface corresponding to the pressure envelope surface model of the compaction block, and the comparison result is output from the input feature vector. Xin Extract the envelope surface input feature vector x The envelope surface is used as the input feature vector. x Dimensions and input feature vectors Xin Since the dimensions are different, feature extraction or mapping operations are required to obtain the desired input feature vector. x .

[0050] For a predicted point location x pi Query the pressure envelope surface function E ( x , t (), is to predict the time point t pred and predicted point location x pi Substituting the values, we obtain the expected median pressure value at the predicted point location, which is based on the envelope surface function. Obtain the expected median pressure value; extract the centerline value corresponding to the predicted point with the largest median pressure value from the expected median pressure values ​​at the predicted point locations, and define it as... p center,max The centerline value of the mean pressure is defined as follows: p center,avg Based on maximum predicted pressure Centerline value corresponding to the predicted point with the largest median pressure value p center,max The maximum absolute pressure difference is The physical meaning is the deviation of the predicted maximum pressure point from the historical normal level of that point; based on the predicted average pressure Centerline value of average pressure p center,avg The average pressure absolute difference is The physical meaning is the deviation of the overall pressure level from the historical normal level.

[0051] Centerline value corresponding to the prediction point with the maximum absolute pressure difference and the median pressure value. p center,max The maximum pressure relative difference is This describes the percentage change in pressure at the point of maximum pressure relative to its normal level; it is based on the absolute difference in mean pressure and the centerline value of mean pressure. p center,avg The average pressure relative difference is This describes the percentage change in overall pressure levels relative to historical averages.

[0052] Based on the center line of the envelope during the prediction period The change yields the rate of change of the centerline at the point of maximum pressure. and the rate of change of the centerline of the mean pressure The two parameters are used to evaluate the normal variation of the centerline of the envelope surface. , , This represents the centerline value corresponding to the initial pressure value. The centerline value corresponds to the average pressure at the initial moment; the predicted rate of change of the maximum pressure is obtained based on the maximum predicted pressure and the measured pressure value at the initial moment. , , The measured pressure at the initial measurement point is given; the predicted rate of change of the average pressure is obtained by comparing the predicted average pressure with the average pressure at the initial time. , , The average measured pressure at the measuring points at the initial moment; the predicted rate of change based on the maximum pressure. and the rate of change of the centerline of the point of maximum pressure The predicted rate of change at the point of maximum pressure is obtained. , Based on the predicted rate of change of mean pressure and the rate of change of the centerline of the mean pressure The predicted rate of change of mean pressure was obtained. , .

[0053] S4: Output alarm signals of the corresponding level based on the comparison results.

[0054] Define the following preset conditions: 1) Absolute difference of maximum pressure Not exceeding 5 MPa, or the absolute difference in mean pressure. 1) Not exceeding 3 MPa; 2) Relative difference in maximum pressure Relative difference between average pressure and average pressure All not exceeding 15%; 3) Centerline change rate and the rate of change of the centerline of the mean pressure All are located in the range of [-0.5, 1.0] MPa / min; 4) Predicted rate of change of the maximum pressure point Predicted rate of change of mean pressure All are located in the range of [-0.4, 0.4] MPa / min; When all the above conditions are met, a level 0 alarm signal is output. If any one condition is not met, a level 1 alarm signal is issued. If two or more conditions are not met, a level 2 alarm signal is issued.

[0055] The zero-level alarm signal indicates that the pressure is normal; the first-level alarm signal indicates that at least one parameter exceeds the safe threshold range. It is necessary to check whether the threshold needs to be relaxed according to different operating conditions. For example, the threshold for the predicted rate of change can be relaxed by 20% in the high-displacement stage, the threshold for the predicted rate of change can be kept unchanged or tightened by 10% in the low sand ratio stage, and the threshold for the predicted rate of change in the formation sensitive section can be tightened by 20%. If adjusting the threshold to eliminate the influence of different operating conditions still results in exceeding the standard, then the pressure is considered to be abnormal and timely intervention is required. Therefore, the corresponding first-level or second-level alarm signal is issued, and corresponding handling suggestions are proposed. In severe cases, it is necessary to consider stopping the pump or shutting down the well.

[0056] On the other hand, the present invention provides an intelligent early warning system combining fracturing big data and the PINN algorithm to implement the above-mentioned method, including: The briquetting pressure envelope surface model building unit obtains the briquetting pressure envelope surface model by dynamic envelope surface modeling based on time-varying manifold learning, and maps the obtained input characteristic vector-specified time pair to different quantiles of the pressure distribution. The Physical Information Neural Network (PINN) is constructed and trained using a unit that constructs a spatiotemporally adaptive PINN. The PINN is then trained by combining a meta-learning algorithm and spatiotemporal domain decomposition to obtain the trained PINN. The prediction and comparison unit is used to predict the construction pressure after a period of time by applying the trained physical information neural network PINN to obtain the predicted construction pressure value; the predicted construction pressure value is compared with the center line of the envelope surface corresponding to the pressure envelope surface model of the briquette, and the comparison result is output. The alarm output unit outputs an alarm signal of the corresponding level based on the comparison result output by the prediction comparison unit.

[0057] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent early warning method combining fracturing big data and the PINN algorithm, characterized in that, Includes the following steps: S1: Dynamic envelope surface modeling based on time-varying manifold learning to obtain the pressure envelope surface model of the briquette; S2: Construct the spatiotemporally adaptive physical information neural network PINN, and train the physical information neural network PINN by combining meta-learning algorithm and spatiotemporal domain decomposition to obtain the trained physical information neural network PINN. S3: The trained physical information neural network PINN is used to predict the construction pressure after a period of time, and the predicted construction pressure value is obtained; the difference between the predicted construction pressure value and the center line of the envelope surface corresponding to the pressure envelope surface model of the briquette is compared, and the comparison result is output. S4: Output alarm signals of the corresponding level based on the comparison results.

2. The intelligent early warning method combining fracturing big data and the PINN algorithm according to claim 1, characterized in that, Step S1 involves acquiring construction pressure, discharge rate, and sand concentration parameters, supplementing missing values, removing outliers, standardizing the parameters to ensure they are within the range of [0, 1], encoding features in a fixed order to obtain several sets of feature vectors with the same format, and then unifying the dynamic spatiotemporal structure into a low-dimensional manifold using a custom time-varying distance metric. Uncertainty quantification is then performed on this low-dimensional manifold to construct the pressure envelope surface function. E ( x , t ), in parentheses x The input characteristic vector of the envelope surface. t For a specified time.

3. The intelligent early warning method combining fracturing big data and the PINN algorithm according to claim 2, characterized in that, The content of step S2 is as follows: Based on the physical process of fracturing, the governing equations are set; In a unified spatiotemporal domain, a physical information neural network (PINN) is constructed, and the total loss function is defined based on the approximate solution of the control equation and the neural network parameters. The meta-learning algorithm is applied to the physical information neural network PINN to train the physical information neural network PINN; Spatiotemporal decomposition is performed, dividing the complete spatiotemporal domain into temporal and spatial decompositions. Temporal decomposition involves breaking down the construction phase according to time decomposition criteria. m Spatial decomposition, which involves dividing a three-dimensional spatial region into different time slices according to spatial decomposition criteria, is a process of dividing the spatial region into... n A spatial region, m Different time slices and n Several spatial regions together constitute several spatiotemporal subdomains. For each spatiotemporal subdomain, a sub-PINN is constructed, and each sub-PINN has a corresponding local loss function. Each obtained sub-PINN is subjected to personalized federated learning until the convergence condition is met.

4. The intelligent early warning method combining fracturing big data and the PINN algorithm according to claim 3, characterized in that, The time decomposition criterion is as follows: if the difference between the rate of change of the injected pressure and the rate of change of the pressure in the current time slice continuously exceeds a set rate of change threshold, it is determined that a phase transition has occurred and a new time slice needs to be created. Subsequent time slices will be assigned to the new time slice; otherwise, no new time slice is created.

5. The intelligent early warning method combining fracturing big data and the PINN algorithm according to claim 3, characterized in that, The spatial decomposition criterion is to obtain spatial points separately. x The magnitude of the gradient vector of the predicted pressure at point 0 and spatial points x Predicted crack width at point 0 Establish a spatial region classification function and obtain the corresponding classification region number: when spatial point x The magnitude of the gradient vector of the predicted pressure at point 0 If the pressure gradient exceeds the threshold, return space number 1; if the space point... x Predicted crack width at point 0 If the width exceeds the effective opening threshold of the crack, return to space number 2; otherwise, return to space number 3.

6. The intelligent early warning method combining fracturing big data and the PINN algorithm according to claim 3, characterized in that, Personalized federated learning of each obtained sub-PINN is what enables nodes to... s The corresponding set of spacetime subdomains is S s Each node corresponds to at least one spatiotemporal subdomain; Set globally shared parameters and local personalized parameters The local loss function is defined as the sum of the local loss functions of all spatiotemporal subdomains corresponding to the local node, plus a regularization term. An alternating optimization approach is adopted, first fixing the globally shared parameters. Each node optimizes its corresponding local personalized parameters. To achieve node adaptation to local data, the local loss function is minimized, and then all local personalized parameters are fixed. Calculate all local personalized parameters Use the average value to update the globally shared parameters Repeated iterations R In each round, the sub-PINN is considered convergent and training is stopped when the following convergence criteria are met: 1) Simultaneously satisfying: the absolute change of the globally shared parameters is less than the global convergence threshold, and the average change of the local personalized parameters is less than the personalized convergence threshold, and maintaining this... P Each round, P < R 2) Use only globally shared parameters A new global network is initialized. In the evaluation set of this new global network, the global loss is constructed and calculated by referring to the content of the local loss function corresponding to each sub-PINN. recent assessment Q Global loss during rounds of iteration The decrease was consistently less than the set threshold. Q < R .

7. The intelligent early warning method combining fracturing big data and the PINN algorithm according to claim 3, characterized in that, Step S3, which involves using a trained physical information neural network (PINN) to predict construction pressure over a period of time to obtain a predicted construction pressure value, defines the prediction time point. t pred Current time t 0 and the forecast period The sum of these features defines the input feature vector of the trained physical information neural network PINN as follows: Xin Output the location at the predicted time point t pred The pressure prediction set is used to obtain the maximum predicted pressure from the pressure prediction set. Predicted mean pressure .

8. The intelligent early warning method combining fracturing big data and the PINN algorithm according to claim 7, characterized in that, Step S3, which compares the predicted construction pressure value with the center line of the envelope surface corresponding to the pressure envelope surface model of the compaction block, and outputs the comparison result, is derived from the input feature vector. Xin Extract the envelope surface input feature vector x For each predicted point location x pi Query the pressure envelope surface function E ( x , t ), predicting the time point t pred and predicted point location x pi Substituting these values, we obtain the expected median pressure value at the predicted point location; from the expected median pressure values ​​at the predicted point locations, we extract the centerline value corresponding to the predicted point location with the largest median pressure value, and define it as... p center,max The centerline value of the mean pressure is defined as follows: p center,avg Based on maximum predicted pressure Centerline value corresponding to the predicted point with the largest median pressure value p center,max The difference is the maximum pressure absolute difference. Based on predicted average pressure Centerline value of average pressure p center,avg The difference is the absolute difference in mean pressure. ; Centerline value corresponding to the prediction point with the maximum absolute pressure difference and the median pressure value. p center,max The ratio of the maximum pressure relative difference Based on the absolute difference of mean pressure and the centerline value of mean pressure p center,avg The ratio of the average pressure relative difference ; Based on the center line of the envelope during the prediction period The change yields the rate of change of the centerline at the point of maximum pressure. and the rate of change of the centerline of the mean pressure The predicted rate of change of the maximum pressure is obtained based on the maximum predicted pressure and the measured pressure value at the initial moment. The predicted rate of change of average pressure is obtained by comparing the predicted average pressure with the average pressure at the initial moment; the predicted rate of change is based on the maximum pressure. and the rate of change of the centerline of the point of maximum pressure The difference yields the predicted rate of change at the point of maximum pressure. Based on the predicted rate of change of mean pressure and the rate of change of the centerline of the mean pressure The difference yields the predicted rate of change of mean pressure. .

9. The intelligent early warning method combining fracturing big data and the PINN algorithm according to claim 8, characterized in that, Step S4 defines the following preset conditions: 1) Absolute difference of maximum pressure Not exceeding 5 MPa, or the absolute difference in mean pressure. 1) Not exceeding 3 MPa; 2) Relative difference in maximum pressure Relative difference between average pressure and average pressure All not exceeding 15%; 3) Centerline change rate and the rate of change of the centerline of the mean pressure All are located in the range of [-0.5, 1.0] MPa / min; 4) Predicted rate of change of the maximum pressure point Predicted rate of change of mean pressure All are located in the range of [-0.4, 0.4] MPa / min; When all the above conditions are met, a level 0 alarm signal is output. If any one condition is not met, a level 1 alarm signal is issued. If two or more conditions are not met, a level 2 alarm signal is issued.

10. An intelligent early warning system combining fracturing big data and the PINN algorithm, used to implement the method described in any one of claims 1-9, characterized in that, include: The briquetting pressure envelope surface model building unit obtains the briquetting pressure envelope surface model by dynamic envelope surface modeling based on time-varying manifold learning, and maps the obtained input characteristic vector-specified time pair to different quantiles of the pressure distribution. The Physical Information Neural Network (PINN) is constructed and trained using a unit that constructs a spatiotemporally adaptive PINN. The PINN is then trained by combining a meta-learning algorithm and spatiotemporal domain decomposition to obtain the trained PINN. The prediction and comparison unit is used to predict the construction pressure after a period of time by applying the trained physical information neural network PINN to obtain the predicted construction pressure value; the predicted construction pressure value is compared with the center line of the envelope surface corresponding to the pressure envelope surface model of the briquette, and the comparison result is output. The alarm output unit outputs an alarm signal of the corresponding level based on the comparison result output by the prediction comparison unit.

Images