A channeling early warning method based on multi-step prediction of adjacent well pressure

Abstract

The application provides a channeling early warning method based on adjacent well pressure multi-step prediction, relates to the field of unconventional oil and gas reservoir fracturing reconstruction and interwell channeling intelligent early warning, and specifically comprises the following steps: weighted estimation is conducted on the measured instantaneous pump-off pressure to obtain the predicted instantaneous pump-off pressure; a sliding time window is used to construct input samples; an adjacent well pressure multi-step prediction model is constructed to output the predicted future adjacent well pressure sequence; the dynamic adjacent well pressure response integral index is calculated; a clustering method is used to divide three channeling degrees, namely weak channeling, medium channeling and strong channeling; the future dynamic adjacent well pressure response integral index prediction track is calculated; and the prediction result is updated rolling. The technical scheme of the application overcomes the problems in the prior art that the fracturing construction channeling risk identification relies on the measured instantaneous pump-off pressure, resulting in lagging judgment, and the instantaneous pump-off pressure cannot be obtained when the current construction section is not completed, so that the dynamic adjacent well pressure response integral index cannot be calculated to effectively perform early warning.

Description

Technical Field

[0001] This invention relates to the field of unconventional oil and gas reservoir fracturing and intelligent early warning of inter-well pressure crosstalk, specifically to an early warning method for pressure crosstalk based on multi-step prediction of pressure in adjacent wells. Background Technology

[0002] In recent years, machine learning and deep learning have been used for pressure prediction and early warning of complex events during fracturing operations. Existing technologies can predict future operational pressure using time series data such as pump pressure, discharge rate, and sand concentration in the fracturing well, or identify conditions such as sand blockage and pump shutdown based on the predicted pressure change trend. These methods mainly use the pressure of the fracturing well itself as the prediction target, serving operational safety or single-well condition identification; however, the key to pressure crosstalk early warning is to characterize the pressure coupling relationship between the fracturing well and adjacent wells, and to answer the question of "how will the pressure of adjacent wells change in the future, and when will the pressure crosstalk risk reach the threshold?" Therefore, directly transplanting single-well operational pressure prediction methods to the pressure crosstalk scenario makes it difficult to generate quantitative results of inter-well pressure crosstalk risk with clear physical meaning.

[0003] On the other hand, on-site assessments of pressure cross-contamination often directly use the maximum pressure increase or pressure rise rate of adjacent wells. The maximum pressure increase only answers "how much it ultimately increased," making it difficult to distinguish different response timings and rise rates under the same increase. The pressure rise rate can reflect the speed of local changes, but it is easily affected by instantaneous fluctuations and cannot reflect the duration of the high-pressure state. For typical pressure response scenarios such as "same increase but different rates, same rate but different pressures, same parameters but different shapes," a single increase or rate may provide similar judgments, making it difficult to fully characterize the intensity and speed of pressure communication between wells. Therefore, a comprehensive index that can simultaneously integrate pressure amplitude, time history, and pressure background is needed to convert the dynamic pressure curves of adjacent wells into a quantity representing the degree of pressure cross-contamination with comparable physical meaning.

[0004] In addition, the dynamic adjacent well pressure response integral index ( It is necessary to utilize the instantaneous pump stop pressure of the current section of the construction well. The pressure response of adjacent wells is normalized. However, when the current construction section is not yet completed and is still in the continuous pumping stage, the actual pressure response of this section is... Not yet generated, leading to The critical denominator parameters are missing during the construction process, which most requires real-time early warning. If we wait until the current section of the pump actually stops before obtaining these parameters... Therefore, it is meaningless to make an advance judgment on the risk of fracturing in this segment. Thus, it is necessary to obtain a prediction of the current segment's fracturing risk before the actual pumping stoppage, based on data from previously fractured segments in the same well or historically similar segments. .

[0005] In summary, existing technologies have solved some problems in adjacent well pressure monitoring, single-well pressure prediction, inter-well interference identification, and fracturing accident early warning, but have not yet formed a comprehensive system. Pre-obtaining pressure from adjacent wells at multiple time steps to predict future trends. The complete early warning link for "predicting trajectory calculation and identifying threshold arrival time" is established.

[0006] Therefore, a method is needed that can obtain predictions before the pump stops in the current stage. The GCN-TCN model is used to predict future adjacent well pressures in a rolling manner, and then the future pressures are calculated. A pressure crossover early warning method based on multi-step prediction of adjacent well pressure predicts the trajectory and identifies the first sustained threshold crossing time, thereby outputting the pressure crossover risk level and early warning lead. Summary of the Invention

[0007] The main objective of this invention is to provide a method for early warning of pressure cross-flow based on multi-step prediction of adjacent well pressure, in order to solve the problems in the prior art where the identification of pressure cross-flow risk during fracturing construction relies on measured pump shutdown pressure, resulting in a lag in judgment, and the inability to obtain instantaneous pump shutdown pressure before the current construction section is completed, thus making it impossible to calculate the dynamic adjacent well pressure response integral index for effective early warning.

[0008] To achieve the above objectives, this invention provides a method for early warning of pressure crosstalk based on multi-step prediction of adjacent well pressure, specifically including the following steps: S1. Obtain the geological and engineering parameters of the current construction section and the completed fracturing sections in the same well. Calculate the comprehensive similarity between the completed fracturing sections and the current construction section. Select completed fracturing sections with a comprehensive similarity not lower than a preset threshold. Determine the weight based on the comprehensive similarity and apply it to the measured instantaneous pump stop pressure. A weighted estimate is performed to obtain the prediction for the current construction section. When the number of completed fracturing segments with a comprehensive similarity not lower than a preset threshold is less than [number missing] At any given time, a regression prediction model is used to obtain the prediction for the current construction section. .

[0009] S2 collects dynamic and static data and constructs input samples using a sliding time window.

[0010] S3. A multi-step prediction model for adjacent well pressure, GCN-TCN, is constructed using a combination of graph convolutional network (GCN) and temporal convolutional network (TCN). The input samples obtained in step S2 are input into the multi-step prediction model for adjacent well pressure, and the predicted future adjacent well pressure sequence is output.

[0011] S4, Calculate the dynamic adjacent well pressure response integral exponent. .

[0012] S5, based on historical crosstalk samples of the target block, automatically classifies crosstalk into three levels: weak crosstalk, medium crosstalk, and strong crosstalk using a clustering method, and verifies this by combining cluster boundaries and physical phenomena. The grading threshold.

[0013] S6, Calculate the future after the current moment. Predict the trajectory.

[0014] S7, for the future The predicted trajectory is scanned step-by-step over time to continuously An early warning is triggered when a predicted point continuously crosses a preset threshold. The moment of the first continuous threshold crossing is identified, and the early warning lead time is calculated. The prediction results are updated on a rolling basis as new data is obtained. After the construction section ends, the actual measured values ​​are used as the warning. Replace the prediction obtained in step S1 The model is also updated on a rolling basis by adding historical sample libraries to the multi-step prediction model of adjacent well pressure.

[0015] Furthermore, step S1 specifically includes the following steps: S1.1, for the current construction well number For each fracturing section, geological and engineering parameters are used as similarity evaluation parameters to calculate the similarity of the fracturing sections in the same well. Normalized deviation of parameters of a completed fracturing segment relative to the current fracturing segment. : ; in, and The first The completed fracturing section and the currently constructed well. The first fracturing stage Item parameters; For geological engineering parameter index; To take the absolute value.

[0016] S1.2, calculate the first based on the normalized deviation. Overall similarity between completed fracturing sections and currently constructed sections : ; in, To evaluate the number of parameters; For the first The weights of the geological engineering parameters, and satisfying ; It is a natural exponential function.

[0017] Furthermore, step S1 also includes: S1.3, when At that time, the first The completed fracturing sections were identified as similar fracturing sections, among which The similarity threshold is selected from high to low based on the overall similarity. The fracturing segments are identified as similar, and their weights are determined based on the overall similarity. The specific calculation method is as follows: ; in, For the first Weights of similar fracturing segments; For the first The overall similarity between the completed fracturing section and the current construction section.

[0018] S1.4, the current construction well number Prediction of each fracturing segment for: ; in, A set of similar fracturing segments; For the first The actual completed fracturing section ; For the current construction well number Prediction of each fracturing segment .

[0019] S1.5 When the number of fractured sections completed in the same well is less than the preset number of sections, a regression prediction model trained on historical fractured sections is used as a supplement. The input of the regression prediction model is geological engineering parameters, and the output is the prediction of the current section. Regression prediction models include: gradient boosting trees, random forests, support vector regression, and neural networks.

[0020] Furthermore, step S2 specifically includes the following steps: S2.1 Align each fracturing section of the construction well with the pressure monitoring interval of the adjacent well using a unified timestamp. Dynamic data includes: pump pressure, discharge rate, sand concentration of the construction well and wellhead pressure of the adjacent well. Static data includes: well spacing, section length, minimum horizontal principal stress, horizontal stress difference and natural fracture characteristics.

[0021] S2.2 cleans up missing values, outliers, and invalid intervals caused by pump stoppages in both dynamic and static data.

[0022] S2.3, The input sample is constructed using a sliding time window. Let the current time be... The length of the sliding window is The input window is... to It consists of dynamic data and static data of the current segment, tagged with... to The adjacent well pressure sequence; For predicting time steps.

[0023] S2.4, divide the training set, validation set and test set according to well section or platform.

[0024] Furthermore, step S3 specifically includes the following steps: S3.1, treat each type of feature variable as a graph node and calculate the normalized adjacency matrix. : ; in, for The corresponding degree matrix; This is the original feature adjacency matrix; It is an identity matrix.

[0025] S3.2, in the In the graph convolutional network (GCN), the normalized adjacency matrix is ​​used to aggregate neighboring node information and perform linear mapping and nonlinear activation to obtain the first layer. Layer Node Feature Matrix : ; in, For the first Layer node feature matrix; For the first Layer trainable weight matrix; It is a non-linear activation function.

[0026] S3.3, after using GCN to generate and embed the time series graphs for each time step within the sliding window, the time series graphs are then embedded into the temporal convolutional network TCN in chronological order. At the current time step... The one-dimensional dilated convolution is represented as: ; in, For TCN at the current moment The output; Historical inputs selected based on the expansion coefficient; The first convolution kernel Each weight; The kernel length is [length]. The coefficient of thermal expansion; This is the weight index of the convolution kernel.

[0027] S3.4, adopts a multi-output structure at the output end. The training process uses the mean squared error as the loss function to predict the future pressure sequence of adjacent wells at each predicted time step. ; in, For multi-step prediction mean square error loss; This represents the number of training samples; For the first The first sample Predicted pressure of adjacent wells at one future time step; For the first The first sample The measured adjacent well pressure corresponding to each future time step.

[0028] Furthermore, step S4 specifically includes: During real-time construction, the stable wellhead pressure before the construction of the adjacent well or before the start of the current segment is used as the baseline pressure. Previously collected pressure data from adjacent wells were based on measured values; currently... The specific calculation formula is as follows: ; in, For the current moment The dynamic adjacent well pressure response integral index; For the neighboring well Measured pressure at each sampling time; Baseline pressure of adjacent wells; This represents the cumulative construction time from the start of this segment to the current moment; the online early warning stage uses the prediction obtained in step S1. Replacement for actual measurements that have not yet been obtained ; From the start of this fracturing stage to the current moment Number of adjacent well pressure sampling points that have been collected.

[0029] Furthermore, step S5 specifically includes the following steps: S5.1, collect baseline pressure of adjacent wells before historical fracturing operations, complete pressure response curves of adjacent wells, and measured data of the operated wells. Construction duration and geological engineering data were collected and calculated segment by segment. Samples with sensor malfunctions, unclear pump stop times, or missing pressure curves are removed to form a historical pressure crosstalk evaluation sample library for the target block.

[0030] S5.2, with Using geological engineering data as an auxiliary feature, K-means, hierarchical clustering, Gaussian mixture model or deep clustering network are used to classify historical samples in an unsupervised or semi-supervised manner, automatically classifying the samples into three categories: weak pressure cross-flow, medium pressure cross-flow and strong pressure cross-flow.

[0031] S5.3, based on adjacent categories The boundary of the distribution determines the first grade threshold. Second grade threshold Alternatively, on-site pressure channeling records, whether adjacent well pressure exceeds fracture closure pressure, and whether pump shutdown and diversion measures were taken can be used as physical tags. When the proportion of pressure channeling responses in strong pressure channeling samples is not lower than a preset value, the first grading threshold is set. Second grade threshold It meets the engineering verification requirements.

[0032] S5.4 verifies the physical rationality of the classification thresholds, and statistically analyzes whether the pressure of adjacent wells in the weak pressure channeling, medium pressure channeling, and strong pressure channeling samples exceeds the fracture closure pressure, the maximum pressure increase of adjacent wells, whether well control risks occur, and whether on-site anti-channeling measures are triggered.

[0033] Furthermore, step S6 specifically includes: Without obtaining the actual situation from the previous stage At that time, the prediction will be obtained Introduced together with future adjacent well pressure sequences In the calculation formula, the current time The previous integration section used measured pressure from adjacent wells; at the current time... The subsequent integration section uses the future adjacent well pressure sequence to obtain the future... The predicted trajectory is calculated using the following formula: ; in, For the current moment After that A future time step ; This represents the number of pressure points currently observed. For the future Predicted pressure in adjacent wells at each time step; To standardize the forecast time step; For the current construction section prediction .

[0034] Furthermore, step S7 specifically includes the following steps: S7.1, based on one or more calibrated risk thresholds, for the future The predicted trajectory is scanned step-by-step from the first time step. Starting from each candidate prediction time step, continuously Each predicted point is no lower than the standard. Risk threshold Then the first Each candidate prediction time step is defined as the moment when the prediction first sustains a threshold crossing, and the specific judgment formula is as follows: ; in, Predicting the future ; For candidate prediction time steps; This refers to the check sequence number starting from the candidate prediction time step.

[0035] S7.2, Calculate the advance warning amount based on the current time. The specific formula is as follows: ; in, To satisfy the minimum threshold condition for continuous crossing .

[0036] Furthermore, step S7 also includes: S7.3, When setting multi-level thresholds, based on future... The risk level is determined by the highest threshold that the predicted trajectory can continuously cross, when the future... The predicted trajectory only crosses Outputting a medium-level warning in real time; when the future Predicted trajectory continues to cross The alert level was upgraded to a high level.

[0037] In S7.4, during online operation, each time a new real-time sampling point is obtained, the sliding window is updated, the future pressure of adjacent wells is re-predicted, and the future pressure is calculated. If the predicted risk level increases, the warning will be updated; if the subsequent future... The predicted trajectory has fallen continuously to Risk threshold The warning will be lifted or reduced according to the preset hysteresis rules.

[0038] S7.5, the current section of construction is completed and actual measurements are obtained. Then, based on actual measurements Replacement Prediction The samples will be added to the historical sample database for use in subsequent segments. predict, Risk threshold verification and rolling updates of the multi-step prediction model for adjacent well pressure.

[0039] The present invention has the following beneficial effects: This invention adopts Dimensionless characterization of the pressure response process of adjacent wells, by comprehensively considering the pressure response amplitude of adjacent wells and its cumulative change characteristics over time, can more comprehensively reflect the strength and development degree of pressure crosstalk response than traditional methods that only use single indicators such as maximum pressure increase and pressure growth rate. It can also achieve a unified quantitative evaluation of the degree of pressure crosstalk between different construction sections and different adjacent wells. Establishing a risk classification criterion for cross-cutting can provide a unified and quantifiable evaluation basis for subsequent cross-cutting status identification, risk prediction, and graded early warning.

[0040] This invention establishes a multi-step prediction method for adjacent well pressure based on GCN-TCN, integrating dynamic parameters of fracturing operations, historical pressure of adjacent wells, and static geological engineering factors. GCN is used to characterize the correlation and coupling relationships between different influencing factors, while TCN is used to extract multi-scale time-series features of fracturing operations and adjacent well pressure changes, enabling prediction of adjacent well pressure at multiple future time steps. Continuous prediction of changing trends. Compared with methods that rely solely on current pressure in adjacent wells for pressure channeling identification, this invention transforms "post-event response identification" into "future trend prediction," enabling early identification of potential pressure channeling risks and their development trends, thereby improving the timeliness and reliability of pressure channeling early warning.

[0041] This invention combines the predicted pressure changes of adjacent wells with... By combining risk assessment with on-site monitoring, a method for early warning of hydraulic cross-flow during fracturing operations can be established. This method can provide information on the degree of risk and the expected time of reaching the risk state before the risk reaches a set level, providing a basis for timely adjustments to anti-cross-flow measures such as pumping volume, operating pressure, pump shutdown timing, and pump reversal. By linking hydraulic cross-flow assessment, trend prediction, and on-site control, the possibility of severe hydraulic cross-flow can be reduced, and the pertinence and timeliness of anti-cross-flow measures can be improved. Attached Figure Description

[0042] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 The flowchart of a pressure crossflow early warning method based on multi-step prediction of adjacent well pressure according to the present invention is shown.

[0043] Figure 2 The diagram illustrates a typical adjacent well pressure response scenario for the first type of invention.

[0044] Figure 3 The diagram illustrates a typical adjacent well pressure response scenario of the second type constructed according to the present invention.

[0045] Figure 4 The diagram illustrates a typical adjacent well pressure response scenario of the third type constructed according to the present invention.

[0046] Figure 5 The flowchart of the multi-source compression data and sliding time window sample construction process in this invention is shown.

[0047] Figure 6 The loss curves of the GCN-TCN model provided in this invention are shown.

[0048] Figure 7 A comparison chart of measured pressure values ​​of adjacent wells in a typical well section and predicted values ​​of GCN-TCN is shown in this invention.

[0049] Figure 8 The pressure prediction error values ​​under different prediction lead times are shown in this invention.

[0050] Figure 9 The diagram shows the relationship between the measured and predicted values ​​of the adjacent well pressure in the test set of this invention.

[0051] Figure 10 The diagram shows the relationship between wells on platform A in this invention and the on-site playback object diagram.

[0052] Figure 11 The figure shows the 5-minute rolling prediction results of the adjacent well pressure in a typical well section of Platform A in this invention.

[0053] Figure 12 The diagram shows the future dynamic adjacent well pressure response integral index trajectory and threshold crossing results for a typical well section of Platform A in this invention.

[0054] Figure 13 The pressure response curve of the adjacent well after the pump is stopped and the direction is reversed is shown in this invention.

[0055] Figure 14 The integral exponential trajectory of the dynamic adjacent well pressure response after pump shutdown and reversal is shown in this invention. Detailed Implementation

[0056] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] like Figure 1 The method for early warning of pressure crossflow based on multi-step prediction of adjacent well pressure, as shown, specifically includes the following steps: S1, Establish the instantaneous pump stop pressure of the current construction section. The prediction method involves obtaining the geological and engineering parameters of the current fracturing section and the completed fracturing sections in the same well, calculating the comprehensive similarity between the completed fracturing sections and the current fracturing section, selecting 3 to 5 completed fracturing sections with a comprehensive similarity not lower than a preset threshold, determining the weights based on the comprehensive similarity, and then applying the measured instantaneous pump shutdown pressure. A weighted estimate is performed to obtain the prediction for the current construction section. When there are fewer than three completed fracturing sections with a comprehensive similarity not lower than a preset threshold, a regression prediction model based on geological engineering parameters is used to predict the current construction section. .

[0058] S2 collects dynamic and static data and constructs input samples using a sliding time window.

[0059] S3. A multi-step prediction model for adjacent well pressure, GCN-TCN, is constructed using a combination of graph convolutional network (GCN) and temporal convolutional network (TCN). The input samples obtained in step S2 are input into the multi-step prediction model for adjacent well pressure, and the predicted future adjacent well pressure sequence is output.

[0060] S4, Calculate the dynamic adjacent well pressure response integral exponent. .

[0061] S5, based on historical crosstalk samples of the target block, automatically classifies crosstalk into three levels: weak crosstalk, medium crosstalk, and strong crosstalk using a clustering method, and verifies this by combining cluster boundaries and physical phenomena. The grading threshold.

[0062] S6, Calculate the future after the current moment. Predict the trajectory.

[0063] S7, for the future The predicted trajectory is scanned step-by-step in a continuous manner. An early warning is triggered when a predicted point continuously crosses a preset threshold. The moment of the first continuous threshold crossing is identified, and the early warning lead time is calculated. The prediction results are updated on a rolling basis as new data is obtained. After the construction section ends, the actual measured values ​​are used as the warning. Replace the prediction obtained in step S1 The model is also updated on a rolling basis by adding historical sample libraries to the multi-step prediction model of adjacent well pressure.

[0064] Specifically, step S1 includes the following steps: S1.1, for the current construction well number For each fracturing section, geological engineering parameters, including vertical depth, minimum horizontal principal stress, horizontal stress difference, section length, cluster number, design displacement, fluid strength, and proppant strength, are used as similarity evaluation parameters to calculate the similarity of the fracturing sections in the same well. Normalized deviation of parameters of a completed fracturing segment relative to the current fracturing segment. : ; in, and The first The completed fracturing section and the currently constructed well. The first fracturing stage Item parameters; For geological engineering parameter index; To take the absolute value.

[0065] S1.2, calculate the first based on the normalized deviation. Overall similarity between completed fracturing sections and currently constructed sections : ; in, To evaluate the number of parameters; For the first The weights of the geological engineering parameters, and satisfying ; It is a natural exponential function.

[0066] Specifically, step S1 also includes: S1.3, when At that time, the first The completed fracturing sections were identified as similar fracturing sections, among which For similarity threshold, Set the value to 0.80. Select 3 to 5 similar fracturing segments based on their overall similarity, from highest to lowest, and determine their weights according to the overall similarity. The specific calculation method is as follows: ; in, For the first Weights of similar fracturing sections; For the first The overall similarity between the completed fracturing section and the current construction section.

[0067] S1.4, the current construction well number Prediction of each fracturing segment for: ; in, A set of similar fracturing segments; For the first The actual completed fracturing section ; For the current construction well number Prediction of each fracturing segment .

[0068] S1.5 When the number of fractured sections completed in the same well is less than the preset number of sections, a regression prediction model trained on historical fractured sections is used as a supplement. The input of the regression prediction model is geological engineering parameters, and the output is the prediction of the current section. Regression prediction models include: gradient boosting trees, random forests, support vector regression, and neural networks.

[0069] Specifically, step S2 includes the following steps: S2.1 Align each fracturing section of the construction well with the pressure monitoring interval of the adjacent well using a unified timestamp. Dynamic data includes: pump pressure, discharge rate, sand concentration of the construction well, and wellhead pressure of the adjacent well. Optional dynamic data includes casing pressure, cumulative fluid volume, cumulative sand volume, and construction stage code. Static data includes: well spacing, section length, minimum horizontal principal stress, horizontal stress difference, and natural fracture characteristics. The main input and output parameters of the multi-step prediction model for adjacent well pressure are shown in Table 1. Table 1. Main input and output parameters of the multi-step pressure prediction model for adjacent wells.

[0070] S2.2 cleans up missing values, outliers, and invalid intervals caused by pump stoppages in both dynamic and static data.

[0071] S2.3, The input sample is constructed using a sliding time window. Let the current time be... The length of the sliding window is The input window is... to It consists of dynamic data and static data of the current segment, tagged with... to The adjacent well pressure sequence; For predicting time steps.

[0072] S2.4 divides the training set, validation set, and test set according to the well section or platform to avoid data leakage caused by randomly distributing data across the training set and test set at different time windows of the same well section.

[0073] Specifically, step S3 includes the following steps: A multi-step prediction model for adjacent well pressure is established by combining a graph convolutional network (GCN) and a temporal convolutional network (TCN). The GCN is used to learn the nonlinear coupling relationship between variables such as geological engineering characteristics, construction parameters, and adjacent well pressure, while the TCN is used to extract the dynamic change patterns at different time scales based on the time series graph embedded by the GCN output.

[0074] S3.1, treat each type of feature variable as a graph node, and use edges to describe the correlation, mutual information, or pre-defined coupling relationships between variables based on mechanistic understanding, and calculate the normalized adjacency matrix. : ; in, for The corresponding degree matrix; This is the original feature adjacency matrix; It is an identity matrix used to retain the information of each node.

[0075] S3.2, in the In the graph convolutional network (GCN), the normalized adjacency matrix is ​​used to aggregate neighboring node information and perform linear mapping and nonlinear activation to obtain the first layer. Layer Node Feature Matrix : ; in, For the first Layer node feature matrix; For the first Layer trainable weight matrix; It is a non-linear activation function.

[0076] S3.3, after using GCN to generate and embed time series graphs for each time step within the sliding window, the time series graphs are then embedded into the Temporal Convolutional Network (TCN) in chronological order. The TCN uses causal convolution to ensure that future information is not used during prediction and utilizes dilated convolution to expand the receptive field at the current time step. The one-dimensional dilated convolution is represented as: ; in, For TCN at the current moment The output; Historical inputs selected based on the expansion coefficient; The first convolution kernel Each weight; The kernel length is [length]. The coefficient of thermal expansion; This is the weight index of the convolution kernel.

[0077] S3.4, adopts a multi-output structure at the output end. Instead of progressively backfilling single-step prediction results as input for the next step, the system predicts the future pressure sequence of adjacent wells at each time step, thereby reducing error accumulation in recursive multi-step predictions. The mean squared error is used as the loss function during training. ; in, For multi-step prediction mean square error loss; This represents the number of training samples; For the first The first sample Predicted pressure of adjacent wells at one future time step; For the first The first sample The measured adjacent well pressure corresponding to each future time step.

[0078] Specifically, step S4 is as follows: During fracturing operations, the pressure response of adjacent wells is one of the most direct dynamic pieces of information regarding inter-well fracture communication and pressure channeling. Traditional methods often use the maximum pressure increase ΔP of adjacent wells to evaluate the degree of pressure channeling, or the pressure rise rate. (The magnitude of pressure increase in adjacent wells within 1 minute) Evaluate how fast the pressure rises, among which, The pressure rise rate represents the magnitude of pressure increase within one minute. However, the maximum increase ignores when the pressure begins to rise, the time required to reach the same increase, and the duration of the high-pressure state. A single rise rate is easily affected by short-term fluctuations such as instantaneous noise, temporary blockage, and adjustments to the construction schedule, and it cannot reflect the cumulative degree of pressure response over the entire construction period. Therefore, using only a single increase or a single rate may lead to biases in the evaluation of pressure channeling. To simultaneously consider the pressure response magnitude and time effect, this invention uses a dynamic adjacent well pressure response integral index. As a quantitative indicator of the degree of compression crosstalk.

[0079] By integrating the pressure increment of the adjacent well relative to the baseline pressure over time, and normalizing it using the difference between the instantaneous pump shutdown pressure of the drilled well and the baseline pressure of the adjacent well, as well as the cumulative drilling time, the pressure response under different well sections, different initial pressures, and different drilling durations can be compared on the same dimensionless scale. Its core function is not to replace the pressure of the adjacent well itself, but to convert the dynamic response of the entire adjacent well into a state quantity that can simultaneously describe the degree of pressure cross-flow, which is both "high" and "rapid".

[0080] During real-time construction, the stable wellhead pressure before the construction of the adjacent well or before the start of the current segment is used as the baseline pressure. Previously collected pressure data from adjacent wells were based on measured values; currently... The specific calculation formula is as follows: ; in, For the current moment The dynamic adjacent well pressure response integral index; For the neighboring well Measured pressure at each sampling time; Baseline pressure of adjacent wells; This represents the cumulative construction time from the start of this segment to the current moment; the online early warning stage uses the prediction obtained in step S1. Replacement for actual measurements that have not yet been obtained ; From the start of this fracturing stage to the current moment Number of adjacent well pressure sampling points that have been collected; The integral of the pressure change in adjacent wells, i.e., the "pressure-time area" of the adjacent well's pressure curve relative to the adjacent well's baseline pressure, shows that, given the same final pressure increase, the earlier and faster the pressure rises, the larger this area. Furthermore, when the rate of increase and the timing of the response are similar, the greater the pressure increase, the larger this area. Therefore, the numerator incorporates both the pressure amplitude and the response speed into the same cumulative quantity. Used to construct pressure scales This is used for time normalization, thereby reducing the impact of differences in initial pressure levels and construction durations across different well sections on the evaluation results. The larger the value, the stronger the normalized cumulative pressure disturbance experienced by the adjacent well during construction, which usually corresponds to more severe inter-well pressure crosstalk; The smaller the value, the weaker or later the pressure response of the adjacent well.

[0081] To explain Compared to the rationality of a single pressure increase and a single rate of increase, three typical adjacent well pressure response scenarios are constructed, such as... Figure 2 , Figure 3 and Figure 4 As shown. Figure 2 In the middle, the final pressure increase ΔP of each adjacent well and the baseline pressure of the adjacent well. The increases are the same, but the rates of increase differ; the faster the increase, the greater the accumulated pressure-time area within the same construction period. The larger. Figure 3 In the middle, the pressure increase and rate of increase of each adjacent well are the same, but the baseline pressure of the adjacent wells is the same. Different; through After normalization, it can reflect the relative degree of disturbance represented by the same increase under different pressure backgrounds. Figure 4 In the middle section, the pressure increase, the baseline pressure of adjacent wells, and the average rate of increase are all the same, but the patterns of increase are different; the curve that rises rapidly in the early stage and remains at a high level for a long time has a larger pressure-time integral, therefore It can further distinguish "isomorphic" responses that are difficult to identify with traditional single-point indicators.

[0082] Figure 2 , Figure 3 and Figure 4 The pressure curves of different adjacent wells are distinguished by solid lines, dashed lines, dotted lines, and different markings, with three scenarios explained. It can not only reflect how severe the pressure increase is in adjacent wells, but also how fast the pressure increases and whether the pressure response continues to accumulate in the early stages of construction. Therefore, it can more completely characterize the degree of pressure cross-flow than a single pressure increase or pressure increase rate.

[0083] Specifically, step S5 includes the following steps: S5.1, The grading threshold should not be mechanically applied across different blocks. Instead, it should be statistically calibrated based on historical pressure crossover events that have occurred in the target block. This involves collecting baseline pressure data from adjacent wells before fracturing operations in historical fracturing sections, complete pressure response curves from adjacent wells, and measured data from the operated wells. The construction duration and geological engineering data that can help determine the degree of pressure channeling were calculated segment by segment. Samples with sensor malfunctions, unclear pump stop times, or missing pressure curves are removed to form a historical pressure crosstalk evaluation sample library for the target block. Taking the existing evaluation data of a deep shale gas block as an example, a historical sample set can be established based on the pressure channeling platform, fractured well and fractured section from 2024 to 2026, while retaining information such as the pressure increase of adjacent wells, the initial pressure of adjacent wells, the pressure of the well where pumps were shut down, and the rate of pressure increase, so as to ensure that the threshold calibration is representative.

[0084] S5.2, with As a grading indicator, geological engineering data is used as an auxiliary feature. The geological engineering data includes well spacing, fluid intensity, construction discharge rate and horizontal principal stress difference. K-means, hierarchical clustering, Gaussian mixture model or deep clustering network are used to classify historical samples in an unsupervised or semi-supervised manner, and the samples are automatically divided into three categories: weak pressure channeling, medium pressure channeling and strong pressure channeling.

[0085] S5.3, based on adjacent categories The boundary of the distribution determines the first grade threshold. Second grade threshold Alternatively, on-site pressure channeling records, whether adjacent well pressure exceeds fracture closure pressure, and whether pump shutdown and diversion measures were taken can be used as physical tags. When the proportion of pressure channeling responses in strong pressure channeling samples is not lower than a preset value (e.g., the preset value is 70%), then the first grading threshold is set. Second grade threshold To meet engineering verification requirements, the threshold obtained in this way comes from the statistical regularity and physical response of events that have occurred in the target block, rather than being artificially fixed; examples of the threshold for the degree of crosstalk are shown in Table 2. Table 2 Example of a threshold for classifying the degree of compression

[0086] In an experiment involving blocks, historical sample clustering yielded... The classification boundary is acceptable. , It should be noted that this value is only used to illustrate the threshold calibration method; when the conditions of other blocks or reservoirs change significantly, the corresponding historical samples should be reused for calibration.

[0087] S5.4 verifies the physical rationality of the classification thresholds and statistically analyzes whether the pressure of adjacent wells in the weak pressure channeling, medium pressure channeling, and strong pressure channeling samples exceeds the fracture closure pressure, the maximum pressure increase of adjacent wells, whether well control risks occur, and whether on-site anti-channeling measures are triggered. In existing block statistics, approximately 82% of the adjacent wells in the strong pressure channeling category experienced pressures exceeding the fracture closure pressure, while this proportion was approximately 25% in the medium-to-weak pressure channeling category, indicating a higher risk. The interval has a good correspondence with stronger inter-well pressure communication. Through two-layer verification of "historical data classification + physical phenomenon verification", the engineering credibility of the early warning threshold can be improved.

[0088] Specifically, step S6 is as follows: Without obtaining the actual situation from the previous stage At that time, the prediction will be obtained Introduced together with future adjacent well pressure sequences In the calculation formula, the current time The previous integration section used measured pressure from adjacent wells; at the current time... The subsequent integration section uses the future adjacent well pressure sequence to obtain the future... The predicted trajectory is calculated using the following formula: ; in, For the current moment After that, the A future time step ; This represents the number of pressure points currently observed. For the future Predicted pressure in adjacent wells at each time step; To standardize the forecast time step; For the current construction section prediction ; After processing The evaluation metrics, originally designed for responses to existing pressure events, have been transformed into risk trajectories that can extend in tandem with future pressure forecasts. This allows for not only assessing the current level of pressure surge but also predicting... Whether it approaches or crosses the block classification threshold in the next few minutes.

[0089] Specifically, step S7 includes the following steps: S7.1, based on one or more calibrated risk thresholds, for the future The predicted trajectory is scanned step-by-step over time. To avoid individual prediction points being affected by sensor noise or model spikes, this invention uses "continuous threshold crossing" instead of "single-point threshold crossing" as the early warning trigger condition, starting from the first... Starting from each candidate prediction time step, continuously Each predicted point is no lower than the standard. Risk threshold Then the first Each candidate prediction time step is defined as the moment when the prediction first sustains a threshold crossing, and the specific judgment formula is as follows: ; in, Predicting the future ; For candidate prediction time steps; This refers to the check sequence number starting from the candidate prediction time step.

[0090] S7.2, Calculate the advance warning amount based on the current time. The specific formula is as follows: ; in, To satisfy the minimum threshold condition for continuous crossing .

[0091] Specifically, step S7 also includes: S7.3, When setting multi-level thresholds, based on future... The risk level is determined by the highest threshold that the predicted trajectory can continuously cross, when the future... The predicted trajectory only crosses Outputting a medium-level warning in real time; when the future Predicted trajectory continues to cross The alert level was upgraded to a high level.

[0092] In S7.4, during online operation, each time a new real-time sampling point is obtained, the sliding window is updated, the future pressure of adjacent wells is re-predicted, and the future pressure is calculated. If the predicted risk level increases, the warning will be updated; if the subsequent future... The predicted trajectory has fallen continuously to Risk threshold The warning will be lifted or reduced according to the preset hysteresis rules.

[0093] S7.5, the current section of construction is completed and actual measurements are obtained. Then, based on actual measurements Replacement Prediction The samples will be added to the historical sample database for use in subsequent segments. predict, Risk threshold verification and rolling updates of the multi-step prediction model for adjacent well pressure.

[0094] Taking historical fracturing data from a deep shale gas block in Southwest China and a typical multi-well platform as an example, this paper describes the data preparation, model training, multi-step pressure prediction, instantaneous pump shutdown pressure prediction of the drilling well, and future... The process of early warning will be explained.

[0095] This experiment collected fracturing crosstalk data from 40 deep shale gas multi-well platforms, obtaining 1089 valid crosstalk well sections with fracturing operation records and adjacent well pressure response records. Data sources included fracturing and geological design data, real-time fracturing operation curves, adjacent wellhead pressure monitoring curves, and natural fracture interpretation maps. For data from different sources, the following procedures were performed sequentially: time reference unification, outlier removal, missing value handling, time sequence alignment between the operated well and adjacent wells, and section-level static parameter matching, forming a structured crosstalk dataset suitable for model training.

[0096] Dynamic time-series parameters include the well's construction pressure, flow rate, sand concentration, cumulative fluid volume, and historical pressure of adjacent wells; static parameters include well spacing, fracture section length, minimum horizontal principal stress, horizontal stress difference, and natural fracture parameters including natural fracture approach angle and natural fracture zone length. Assuming a data sampling interval of 5 seconds, data from 10 minutes prior to the current moment is used as the input window. One sampling point; predict the pressure of adjacent wells in the next 5 minutes, i.e. 43,560 time series training samples were obtained by using a sliding time window to convert continuous time series into supervised learning samples. Each time the time window was moved, a time series sample of "past 10 minutes multivariate input - future 5 minutes adjacent well pressure label" was formed; an average of 40 samples were constructed for each well segment within the effective pressure response interval, resulting in a total of 43,560 time series training samples.

[0097] To avoid information leakage between different well sections on the same platform, a platform-level grouping was implemented. This involved dividing the platform into training, validation, and test sets. 32 platforms were used for model training, 4 for model parameter validation, and 4 as independent test sets; the corresponding time-series sample sizes were 34,840, 4,360, and 4,360, respectively. This partitioning ensured that the test platforms were completely invisible during model training, more closely mimicking the conditions of cross-platform applications in the field. The data construction and partitioning process is as follows: Figure 5 As shown, the sample statistics are shown in Table 3; Table 3. Composition of Adjacent Well Pressure Prediction Dataset

[0098] Multi-source data is input into the adjacent well pressure multi-step prediction model according to the "feature relationship - time evolution". First, an adjacency matrix is ​​constructed based on the Pearson correlation coefficient between variables and engineering prior relationships. The engineering prior relationships mainly consider the influence of well spacing, geostress differences, natural fracture characteristics, and construction parameters on inter-well pressure transmission. Then, the correlation features between different variables are aggregated using GCN, and the time series graph output by GCN is embedded into TCN. Multi-scale time variation features during fracturing operation are extracted through causal convolution, dilatational convolution, and residual connection. Finally, the adjacent well pressure for the next 60 sampling times is output at once.

[0099] This experiment used a two-layer GCN with hidden dimensions of 64 and 32 respectively; the TCN had four residual blocks, a kernel size of 3, and dilation coefficients of 1, 2, 4, and 8 respectively, with Dropout set to 0.20; the Adam optimizer was used with an initial learning rate of 0.001, a batch size of 64, and a maximum training epoch of 200; the loss function was multi-timestep mean squared error. After approximately 150 epochs of training, the losses on the training and validation sets stabilized, with a final training loss of approximately 0.09 and a validation loss of approximately 0.12, indicating no significant overfitting. Figure 6 As shown.

[0100] Using the coefficient of determination The model performance was evaluated using mean absolute error (MAE) and root mean square error (RMSE). Test results show that the multi-step prediction model for adjacent well pressure can effectively track the overall upward trend and local fluctuations in adjacent well pressure. The training set... The value is 0.971, for the test set. The MAE was 0.943, the MAE on the test set was 0.21 MPa, and the RMSE was 0.33 MPa. Compared with the baseline models LSTM, GRU, and TCN, the multi-step prediction model for adjacent well pressure showed [significance missing]. Higher values, with lower MAE and RMSE, results are shown in Table 4; Table 4. Performance Comparison of Different Adjacent Well Pressure Prediction Models

[0101] The measured adjacent well pressure and the GCN-TCN predicted pressure in a typical test well section are as follows: Figure 7 As shown, the original time-series data sampling interval is 5 seconds. Therefore, the multi-step pressure prediction model for adjacent wells does not only predict pressure in whole minutes, but outputs 60 future pressure points at once within a 5-minute prediction window, maintaining the same time resolution as the field sampling interval. As the prediction lead time increases, the error generally increases slowly, but it still maintains good stability within the 5-minute range. Figure 8 Error variations were statistically analyzed for different lead times, using a 5-second prediction step size. The measured and predicted pressures in the test set were generally distributed around the 45° reference line. Figure 9As shown, this demonstrates that the model can meet the subsequent requirements. The requirement for multi-step prediction of adjacent well pressures for continuous trajectory calculation.

[0102] A field application was conducted on Platform A, a deep shale gas formation in Southwest China. This platform is a multi-well chain-fracturing platform with a well spacing of approximately 350m, and significant inter-well pressure responses were observed during historical fracturing operations. The main horizontal wells are designated as A1, A2, A3, and A4. This experiment selected Well A4, which had previously experienced severe pressure channeling, as the research object, and the 9th segment currently undergoing fracturing was used as the target for this early warning method. Well A3 was used as the main monitoring neighboring well, and the inter-well relationships on the platform are as follows: Figure 10 As shown.

[0103] While the pumps in the current construction section are still running, it is impossible to directly obtain the actual measurements for this section. Therefore, this experiment uses a weighted estimation based on the already completed fracturing sections of the same well. Four fracturing sections with similar geological and engineering conditions were selected from the previously completed sections of well A4, as shown in Table 5. We take 64.80 MPa, 65.10 MPa, 64.40 MPa, and 64.70 MPa respectively; and assign weights of 0.40, 0.30, 0.20, and 0.10 according to the weighting formula to obtain the prediction for the current segment. It is 64.8 MPa. After construction is completed, the current section's... The value is 64.85 MPa, with an absolute error of 0.05 MPa and a relative error of approximately 0.08%. This small error verifies the accuracy and practicality of the method. Table 5 Current Construction Section The prediction uses similar completed segments

[0104] Before well A3 construction, the wellhead pressure was stabilized at 48.00 MPa. During construction, the construction pressure, discharge rate, sand concentration, and well A3 pressure were continuously updated at 5-second sampling intervals. When the construction reached the predicted starting point... At min, multivariate time series data from the most recent 10 min, totaling 120 sampling points, are input into the GCN-TCN model, which outputs the predicted pressure values ​​of adjacent wells for the next 5 min, totaling 60 5-second intervals. Figure 11 A continuous time axis display is used: before the prediction starting point, the observed pressures of adjacent wells are displayed; after the prediction starting point, 60 future predicted values ​​are connected by dashed lines, and subsequent measured pressures are simultaneously superimposed for effect verification. The predicted sequence and the measured sequence in the figure are shown in... The curves are continuously connected and their changes are consistent with the pressure response patterns of adjacent wells in the field.

[0105] Predict The following values ​​were substituted together: (64.80 MPa), initial pressure of the adjacent well (48.00 MPa), measured pressure of the adjacent well before the prediction starting point, and predicted pressure at points every 5 minutes and 5 seconds. Calculation formula. Figure 12 Chinese history ,predict and subsequent tests All are derived from the same pressure sequence and the same prediction. The pressure prediction chart is calculated using a unified 5-second time step and is therefore consistent with... The graphs correspond continuously on the time axis. Online early warning results indicate that the model can predict events as early as time t0. The threshold for malignant cross-flow was expected to be reached in about 250 seconds. Field measurements showed that the well section entered the malignant cross-flow zone about 255 seconds after the prediction, with an error of only about 5 seconds in the time the threshold was reached. The prediction and the actual measurement were in good agreement.

[0106] The above experiments are used to verify " Value selection — Adjacent well pressure prediction — The complete calculation chain is "real-time calculation - early warning threshold triggering". In actual online operation, the input window moves forward once for each new 5-second sampling point, and the next 60 pressure points and corresponding values ​​are re-outputted. Trajectory; if multiple consecutive sampling points of the predicted trajectory cross 0.45, a high-level warning for severe pressure surge will be triggered, and minutes of processing time will be reserved for measures such as pump shutdown, diversion, reduction of discharge, or temporary blockage.

[0107] To verify the positive effect of the early warning results on on-site anti-channeling measures, further application was conducted in the actual fracturing operation of the 10th section of Well A4. During the construction of this section, the method of this invention was run online on-site, continuously receiving construction parameters and pressure monitoring data from adjacent wells. When the model predicts future... When the pressure continues to rise and approaches the threshold for severe pressure surge, pump shutdown and diversion measures will be implemented on-site accordingly; the actual on-site construction and pressure response of adjacent wells are shown in [reference needed]. Figure 13 and Figure 14 .

[0108] During this construction phase, the invention continuously outputs the risk trajectory for the next 5 minutes before the pump shutdown and rerouting. Online prediction results indicate that if the original construction schedule is maintained without adjustment, the future... The pressure will continue to rise and cross the 0.45 malignant pressure surge threshold; however, after the pump was stopped and redirected based on the early warning results of this invention, the pressure increase in the adjacent well was controlled, as measured in actual measurements. The value never exceeded the threshold for severe pressure surge, eventually stabilizing at 0.39. This example demonstrates that the present invention can not only identify the risk of severe pressure surge in advance, but also directly provide a basis for selecting the timing of pump shutdown and diversion, transforming the early warning results into on-site construction control measures, thereby improving the timeliness and pertinence of on-site anti-surge control.

[0109] This indicates that the present invention can effectively solve the problems of delayed identification of pressure channeling risk and insufficient on-site control basis during fracturing operations, and its prediction of adjacent well pressure change trends, The risk trajectory and threshold arrival times showed good consistency with the field measurement results. Furthermore, implementing anti-channeling measures such as pump shutdown and diversion based on the early warning results effectively controlled the continuous rise in pressure in adjacent wells and reduced the risk of pressure channeling. These results demonstrate that multi-step pressure prediction of adjacent wells based on GCN-TCN and… The dynamic evaluation method for early warning of hydraulic fracturing has good predictive reliability and engineering applicability. It can provide auxiliary decision-making basis for early identification of hydraulic fracturing risks in the field, selection of the timing of implementation of anti-fracturing measures, and dynamic control of fracturing construction parameters.

[0110] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for early warning of pressure channeling based on multi-step prediction of adjacent well pressure, characterized in that, Specifically, the steps include the following: S1. Obtain the geological and engineering parameters of the current construction section and the completed fracturing sections in the same well. Calculate the comprehensive similarity between the completed fracturing sections and the current construction section. Select completed fracturing sections with a comprehensive similarity not lower than a preset threshold. Determine the weight based on the comprehensive similarity and apply it to the measured instantaneous pump stop pressure. A weighted estimate is performed to obtain the prediction for the current construction section. When the number of completed fracturing segments with a comprehensive similarity not lower than a preset threshold is less than [number missing] At any given time, a regression prediction model is used to obtain the prediction for the current construction section. ; S2, collect dynamic and static data, and construct input samples using a sliding time window; S3. A multi-step prediction model for adjacent well pressure, GCN-TCN, is constructed by combining graph convolutional network (GCN) and temporal convolutional network (TCN). The input samples obtained in step S2 are input into the multi-step prediction model for adjacent well pressure, and the predicted future adjacent well pressure sequence is output. S4, Calculate the dynamic adjacent well pressure response integral exponent. ; S5, based on historical crosstalk samples of the target block, automatically classifies crosstalk into three levels: weak crosstalk, medium crosstalk, and strong crosstalk using a clustering method, and verifies this by combining cluster boundaries and physical phenomena. The grading threshold; S6, Calculate the future after the current moment. Predicted trajectory; S7, for the future The predicted trajectory is scanned step-by-step in a continuous manner. An early warning is triggered when a predicted point continuously crosses a preset threshold. The moment of the first continuous threshold crossing is identified, and the early warning lead time is calculated. The prediction results are updated on a rolling basis as new data is obtained. After the construction section ends, the actual measured values ​​are used as the warning. Replace the prediction obtained in step S1 The model is also updated on a rolling basis by adding historical sample libraries to the multi-step prediction model of adjacent well pressure.

2. The method for early warning of pressure channeling based on multi-step prediction of adjacent well pressure according to claim 1, characterized in that, Step S1 specifically includes the following steps: S1.1, for the current construction well number For each fracturing section, geological and engineering parameters are used as similarity evaluation parameters to calculate the similarity of the fracturing sections in the same well. Normalized deviation of parameters of a completed fracturing segment relative to the current fracturing segment. : ； in, and The first The completed fracturing section and the currently constructed well. The first fracturing stage Item parameters; For geological engineering parameter index; To take the absolute value; S1.2, calculate the first based on the normalized deviation. Overall similarity between completed fracturing sections and currently constructed sections : ； in, To evaluate the number of parameters; For the first The weights of the geological engineering parameters, and satisfying ; It is a natural exponential function.

3. The method for early warning of pressure channeling based on multi-step prediction of adjacent well pressure according to claim 2, characterized in that, Step S1 also includes: S1.3, when At that time, the first The completed fracturing sections were identified as similar fracturing sections, among which The similarity threshold is selected from high to low based on the overall similarity. The fracturing segments are identified as similar, and their weights are determined based on the overall similarity. The specific calculation method is as follows: ； in, For the first Weights of similar fracturing sections; For the first The overall similarity between the completed fracturing section and the currently constructed section; S1.4, the current construction well number Prediction of each fracturing segment for: ； in, A set of similar fracturing segments; For the first The actual completed fracturing section ; For the current construction well number Prediction of each fracturing segment ; S1.5 When the number of fractured sections completed in the same well is less than the preset number of sections, a regression prediction model trained on historical fractured sections is used as a supplement. The input of the regression prediction model is geological engineering parameters, and the output is the prediction of the current section. Regression prediction models include: gradient boosting trees, random forests, support vector regression, and neural networks.

4. The method for early warning of pressure channeling based on multi-step prediction of adjacent well pressure according to claim 1, characterized in that, Step S2 specifically includes the following steps: S2.1 Align each fracturing section of the construction well with the pressure monitoring interval of the adjacent well according to a unified timestamp. Dynamic data includes: pump pressure, discharge rate, sand concentration of the construction well and wellhead pressure of the adjacent well. Static data includes: well spacing, section length, minimum horizontal principal stress, horizontal stress difference and natural fracture characteristics. S2.2 Cleans up missing values, outliers, and invalid intervals caused by pump stoppage in both dynamic and static data; S2.3, The input sample is constructed using a sliding time window. Let the current time be... The length of the sliding window is The input window is... to It consists of dynamic data and static data of the current segment, tagged with... to The adjacent well pressure sequence; For predicting time steps; S2.4, divide the training set, validation set and test set according to well section or platform.

5. The method for early warning of pressure channeling based on multi-step prediction of adjacent well pressure according to claim 1, characterized in that, Step S3 specifically includes the following steps: S3.1, treat each type of feature variable as a graph node and calculate the normalized adjacency matrix. : ； in, for The corresponding degree matrix; This is the original feature adjacency matrix; It is the identity matrix; S3.2, in the In the graph convolutional network (GCN), the normalized adjacency matrix is ​​used to aggregate neighboring node information and perform linear mapping and nonlinear activation to obtain the first layer. Layer Node Feature Matrix : ； in, For the first Layer node feature matrix; For the first Layer trainable weight matrix; It is a non-linear activation function; S3.3, after using GCN to generate and embed the time series graphs for each time step within the sliding window, the time series graphs are then embedded into the temporal convolutional network TCN in chronological order. At the current time step... The one-dimensional dilated convolution is represented as: ； in, For TCN at the current moment The output; Historical inputs selected based on the expansion coefficient; The first convolution kernel Each weight; The kernel length is [length]. The coefficient of thermal expansion; The weight index of the convolution kernel; S3.4, adopts a multi-output structure at the output end. The training process uses the mean squared error as the loss function to predict the future pressure sequence of adjacent wells at each predicted time step. ； in, For multi-step prediction mean square error loss; This represents the number of training samples; For the first The first sample Predicted pressure of adjacent wells at one future time step; For the first The first sample The measured adjacent well pressure corresponding to each future time step.

6. The method for early warning of pressure channeling based on multi-step prediction of adjacent well pressure according to claim 1, characterized in that, Step S4 is as follows: During real-time construction, the stable wellhead pressure before the construction of the adjacent well or before the start of the current segment is used as the baseline pressure. Previously collected pressure data from adjacent wells were based on measured values; currently... The specific calculation formula is as follows: ； in, For the current moment The dynamic adjacent well pressure response integral index; For the neighboring well Measured pressure at each sampling time; Baseline pressure of adjacent wells; This represents the cumulative construction time from the start of this segment to the current moment; the online early warning stage uses the prediction obtained in step S1. Replacement for actual measurements that have not yet been obtained ; From the start of this fracturing stage to the current moment Number of adjacent well pressure sampling points that have been collected.

7. The method for early warning of pressure channeling based on multi-step prediction of adjacent well pressure according to claim 1, characterized in that, Step S5 specifically includes the following steps: S5.1, collect baseline pressure of adjacent wells before historical fracturing operations, complete pressure response curves of adjacent wells, and measured data of the operated wells. Construction duration and geological engineering data were collected and calculated segment by segment. Samples with sensor malfunctions, unclear pump stop times, or missing pressure curves are removed to form a historical pressure crosstalk evaluation sample library for the target block. S5.2, with As a grading indicator, and with geological engineering data as an auxiliary feature, K-means, hierarchical clustering, Gaussian mixture model or deep clustering network are used to classify historical samples in an unsupervised or semi-supervised manner, and the samples are automatically divided into three categories: weak pressure cross-flow, medium pressure cross-flow and strong pressure cross-flow. S5.3, based on adjacent categories The boundary of the distribution determines the first grade threshold. Second grade threshold Alternatively, on-site pressure channeling records, whether adjacent well pressure exceeds fracture closure pressure, and whether pump shutdown and diversion measures were taken can be used as physical tags. When the proportion of pressure channeling responses in strong pressure channeling samples is not lower than a preset value, the first grading threshold is set. Second grade threshold Meets engineering verification requirements; S5.4 verifies the physical rationality of the classification thresholds, and statistically analyzes whether the pressure of adjacent wells in the weak pressure channeling, medium pressure channeling, and strong pressure channeling samples exceeds the fracture closure pressure, the maximum pressure increase of adjacent wells, whether well control risks occur, and whether on-site anti-channeling measures are triggered.

8. The method for early warning of pressure channeling based on multi-step prediction of adjacent well pressure according to claim 1, characterized in that, Step S6 is as follows: Without obtaining the actual situation from the previous stage At that time, the prediction will be obtained Introduced together with future adjacent well pressure sequences In the calculation formula, the current time The previous integration section used measured pressure from adjacent wells; at the current time... The subsequent integration section uses the future adjacent well pressure sequence to obtain the future... The predicted trajectory is calculated using the following formula: ； in, For the current moment After that, the A future time step ; This represents the number of pressure points currently observed. For the future Predicted pressure in adjacent wells at each time step; To standardize the forecast time step; For the current construction section prediction .

9. The method for early warning of pressure channeling based on multi-step prediction of adjacent well pressure according to claim 1, characterized in that, Step S7 specifically includes the following steps: S7.1, based on one or more calibrated risk thresholds, for the future The predicted trajectory is scanned step-by-step from the first time step. Starting from each candidate prediction time step, continuously Each predicted point is no lower than the standard. Risk threshold Then the first Each candidate prediction time step is defined as the moment when the prediction first sustains a threshold crossing, and the specific judgment formula is as follows: ； in, Predicting the future ; For candidate prediction time steps; The check sequence number is the number of checks performed starting from the candidate prediction time step. S7.2, Calculate the advance warning amount based on the current time. The specific formula is as follows: ； in, To satisfy the minimum threshold condition for continuous crossing .

10. A method for early warning of pressure channeling based on multi-step prediction of adjacent well pressure according to claim 9, characterized in that, Step S7 also includes: S7.3, When setting multi-level thresholds, based on future... The risk level is determined by the highest threshold that the predicted trajectory can continuously cross, when the future... The predicted trajectory only crosses Outputting a medium-level warning in real time; when the future Predicted trajectory continues to cross The alert level was upgraded to a high-level warning. In S7.4, during online operation, each time a new real-time sampling point is obtained, the sliding window is updated, the future pressure of adjacent wells is re-predicted, and the future pressure is calculated. If the predicted risk level increases, the warning will be updated; if the subsequent future... The predicted trajectory has fallen continuously to Risk threshold The warning will be lifted or reduced according to the preset hysteresis rules. S7.5, the current section of construction is completed and actual measurements are obtained. Then, based on actual measurements Replacement Prediction The samples will be added to the historical sample database for use in subsequent segments. predict, Risk threshold verification and rolling updates of the multi-step prediction model for adjacent well pressure.

Images