Physically Guided Intelligent Optimization Method for Segmentation and Clustering of Strongly Heterogeneous Reservoirs

By optimizing the fracturing design using an improved temporal convolutional network and the NSGA-II genetic algorithm, the problem of uneven reservoir stimulation in highly heterogeneous reservoirs was solved, resulting in more efficient reservoir stimulation.

CN122904738BActive Publication Date: 2026-11-17SOUTHWEST PETROLEUM UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611386230.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-08
Publication Date
2026-11-17
Estimated Expiration
2046-09-08

AI Technical Summary

Technical Problem

In highly heterogeneous reservoirs, traditional fracturing designs fail to fully utilize the spatial distribution information of fracturing pressure, resulting in uneven reservoir stimulation and affecting the overall productivity of the horizontal section.

Method used

A physical-guided intelligent optimization method for segmented clustering of highly heterogeneous reservoirs is adopted. The improved temporal convolutional network is used to predict the fracturing pressure, and the NSGA-II genetic algorithm is combined to optimize the number of fracturing segments, the number of clusters, and the location of clusters. Multi-label dynamic programming is used to determine the specific cluster location, thereby reducing risk and uncertainty.

Benefits of technology

It improved the adaptability and reliability of fracture pressure spatial prediction, optimized the reservoir segmentation and clustering scheme, reduced the average risk of the whole well and the risk within the segment, and achieved more efficient reservoir stimulation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure SMS_1
    Figure SMS_1
  • Figure SMS_2
    Figure SMS_2
  • Figure SMS_22
    Figure SMS_22
Patent Text Reader

Abstract

This invention discloses a physically guided intelligent optimization method for segmented clustering of highly heterogeneous reservoirs, relating to the field of fracturing technology. This optimization method uses an improved temporal convolutional network to obtain the conditional mean and prediction standard deviation of candidate locations for the horizontal well segment to be optimized. Then, based on inner and outer layer fracturing segmented clustering optimization models, it determines the segmentation and clustering scheme for the horizontal well segment to be optimized: the outer layer uses the NSGA-II genetic algorithm to search for the number of fracturing segments, the boundaries of each fracturing segment, and the number of clusters in each fracturing segment; when evaluating each individual in the outer layer population, the inner layer calls multi-label dynamic programming and combines it with risk-corrected pressure to determine the location of fracturing clusters within the segment under given segment boundaries, cluster numbers, and mean-variance tradeoff coefficients. The method of this invention directly couples the prediction results with the joint design of fracturing segment boundaries, cluster numbers, and cluster locations, constructing an integrated technology chain from multi-source wellbore information to executable segment-cluster schemes, making the prediction results more accurate.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of fracturing technology, specifically to a physically guided intelligent optimization method for segmenting and clustering highly heterogeneous reservoirs. Background Technology

[0002] Multi-cluster fracturing is a core production enhancement method for the industrial-scale exploitation of unconventional reservoirs such as deep coalbed methane. Its basic principle is to divide the horizontal wellbore into several fracturing sections, and within each section, arrange multiple perforation clusters. Through simultaneous or staggered multi-fracture initiation and propagation, the reservoir is fully stimulated. Fracture pressure is a key engineering parameter for determining whether the rock can fracture at a specific location and whether the clusters within the same section can fracture evenly. If the fracture pressure difference between different clusters within the same fracturing section is too large, clusters with low fracture pressure will preferentially absorb fluid and form dominant channels, while clusters with high fracture pressure may experience delayed or even ineffective fracturing, ultimately leading to uneven stimulation within the section and limited overall production capacity of the horizontal section.

[0003] For highly heterogeneous reservoirs where lithology, geostress, pore pressure, and natural fracture conditions vary significantly along well depth, fracture pressure exhibits complex nonlinear and non-stationary spatial distribution characteristics along the horizontal section. Traditional fracturing design schemes with equal section length, equal cluster number, and equal cluster spacing facilitate construction organization but do not fully utilize the spatial distribution information of fracture pressure; schemes that select only a few advantageous locations based on sweet spot evaluation may sacrifice horizontal section coverage and overall stimulation scale. Therefore, how to obtain high-precision, quantifiable fracture pressure spatial distribution predictions and directly couple the prediction results with the joint design of fracturing section boundaries, cluster number, and cluster location to construct an integrated technology chain from multi-source wellbore information to executable section and cluster schemes is a core technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] To address at least one of the aforementioned problems, this invention proposes a physically guided intelligent optimization method for segmenting and clustering highly heterogeneous reservoirs.

[0005] The technical solution of this invention to solve the above problems is as follows: A physically guided intelligent optimization method for segmentation and clustering of highly heterogeneous reservoirs, comprising the following steps: S1. Obtain historical data of the target block and preprocess it. The historical data includes logging parameters, rock mechanics and geostress parameters, wellbore trajectory parameters, fracturing operation time data, and historical segment cluster structure and constraint data. S2. An improved temporal convolutional network is established and trained using preprocessed historical data to obtain a trained network model. The improved temporal convolutional network is based on a temporal convolutional network and uses the physical baseline of the fracture pressure as the prior center for gated residual fusion. Parallel receptive fields are set according to three engineering scales: local heterogeneity, inter-cluster variation, and segment-level variation. The conditional mean and prediction standard deviation are output synchronously through the heteroscedasticity probability head. S3. Generate a set of candidate locations consisting of multiple candidate locations in the horizontal well section to be optimized, and input the candidate locations into the trained network model to obtain the conditional mean and prediction standard deviation of each candidate location, and then calculate the risk correction pressure of the candidate location; each candidate location record includes candidate coordinates, improved temporal convolutional network input feature vector and fracture pressure physical baseline; S4. Determine the segmentation and clustering scheme of the horizontal well section to be optimized based on the inner and outer layer fracturing segmentation and clustering optimization model: The outer layer uses the NSGA-II genetic algorithm to search for the number of fracturing segments, the boundary of each fracturing segment, and the number of clusters in each fracturing segment; When evaluating each individual in the outer layer population, the inner layer calls multi-label dynamic programming and combines it with risk correction pressure to determine the location of fracturing clusters within the segment under the given segment boundary, number of clusters, and mean-variance tradeoff coefficient; After the outer layer completes the non-dominated sorting, it outputs Pareto candidate solutions and determines the final scheme from the Pareto candidate solutions according to the preset engineering preferences.

[0006] The beneficial effects of this invention are as follows: (1) By using the physical baseline of fracture pressure obtained from the elasticity of the surrounding rock of the wellbore as the prior center, and by using a multi-scale temporal convolutional network to learn the data residuals caused by local heterogeneity, inter-cluster variation and segment-level variation, the adaptability of spatial prediction of fracture pressure in strongly heterogeneous reservoirs can be improved while preserving physical interpretability. (2) The conditional mean and prediction standard deviation are output synchronously through the heteroscedasticity probability head, and the risk correction pressure is constructed accordingly. This enables subsequent cluster optimization to not only consider the prediction pressure level, but also to explicitly consider the prediction uncertainty of different candidate positions, thereby reducing the risk caused by clustering based solely on a single point prediction. (3) By jointly searching the number of fracturing sections, section boundaries and number of clusters in each section through the outer NSGA-II, and nesting the inner multi-label dynamic programming to determine the specific cluster location during the evaluation of each outer individual, it can simultaneously reduce the average risk correction pressure of the whole well and the variance of risk correction pressure within the section under the engineering constraints such as section length, cluster spacing, end margin, forbidden zone and uncertainty, thereby obtaining an executable segmented and clustered collaborative optimization scheme. Detailed Implementation

[0007] The specific embodiments of the present invention will be clearly and completely described below with reference to examples. Obviously, the described examples are only some embodiments of the present invention, and not all embodiments.

[0008] A physically guided intelligent optimization method for segmented clustering of highly heterogeneous reservoirs includes the following steps: S1. Obtain historical data of the target block and preprocess it. The historical data includes logging parameters, rock mechanics and geostress parameters, wellbore trajectory parameters, fracturing operation time data, and historical segment cluster structure and constraint data.

[0009] The model input feature data includes logging parameters, rock mechanics and geostress parameters, and wellbore trajectory parameters. Logging parameters include density, sonic transit time, natural gamma, compensated neutrons, wellbore diameter, and resistivity. Rock mechanics and geostress parameters include Young's modulus, Poisson's ratio, brittleness index, maximum horizontal principal stress, minimum horizontal principal stress, pore pressure, and tensile strength. Wellbore trajectory parameters include well depth, vertical depth, inclination angle, and azimuth angle. Label construction data includes wellhead pressure, flow rate, fluid column height or pressure, total field friction, and, if necessary, perforation number, perforation diameter, tubing string inner diameter, tubing length, fracturing fluid density, flow coefficient, thickener concentration, and proppant concentration for friction verification. Historical segment cluster structure and constraint data include segment boundaries, cluster locations, cluster number, cluster spacing, no-fracturing zones, and end margins. Actual fracturing pressure is calculated from the label construction data or verified through field records to form training labels and is not used as network input features.

[0010] The pressure curves during fracturing operations are converted from wellhead to bottomhole to identify the initiation time corresponding to pressure peaks or slope abrupt changes during the flow rate increase. Only when the single-cluster operation record, cluster measurement results, or verified field records can map the initiation event to a specific cluster location is the fracture pressure at that moment used as the point-level true fracture pressure label for that location. For records of simultaneous fracturing of multiple clusters where the first initiation cluster cannot be clearly located, only the segment-level initiation event label is retained; it is not forcibly mapped to any specific cluster location, nor is it interpolated to the point-level label of all logging sampling points. Only when there are at least two reliable point-level labels within the same historical fracturing segment, or when there are independently obtained reliable segment-level mean and variance statistics, are segment-level statistical labels for segment-level mean-variance loss constructed. The former obtains the true segment-level mean and variance from the point-level label statistics, while the latter converts the independent segment-level mean using the same affine normalization transformation as the training label, and converts the independent segment-level variance to a unified normalized scale by squared corresponding pressure scale factors.

[0011] The construction data is first used to calculate the pressure inside the bottom pipe; after deducting the local friction of the orifice from the bottom pipe pressure, the fracture opening pressure is obtained. At the moment of rock fracturing initiation, the fracture opening pressure that can be clearly located to a specific cluster is defined as the true fracturing pressure of that cluster. In cases where multiple clusters are constructed simultaneously and the first fracturing cluster cannot be located, this pressure only represents the first fracturing event of that fracturing segment and is not directly used as a point-level label for any specific cluster.

[0012] The pressure inside the wellbore and the pressure at the joint at any given time are calculated using the following formulas: In the formula, The pressure inside the wellbore casing is in MPa. Pressure at the outlet side of the orifice, MPa; Wellhead pressure, MPa; The pressure is the liquid column pressure, in MPa. Frictional resistance along the vertical and horizontal sections of the tubing, MPa; The local friction of the orifice is measured in MPa.

[0013] The formulas for calculating orifice friction and friction along the path are as follows.

[0014] For combined friction When the total friction resistance field is verified and usable, the total friction resistance field should be used directly; when the total friction resistance field is missing or needs to be verified, [the following should be specified]. ,in To reduce local friction in the orifice. The frictional resistance along the vertical and horizontal sections of the tubing.

[0015] The formula for calculating the friction of the orifice is as follows: In the formula, The flow rate at the crack inlet is m³ / min; The density of the fracturing fluid is kg / m³. This represents the number of perforations in the cluster. The diameter of the perforation hole is in meters (m). This is the flow coefficient. When the orifice is intact. When the hole is completely eroded The inlet flow rate must be based on the diversion result of the corresponding cluster; when only the total discharge rate is available on site and reliable diversion results are lacking, the total discharge rate is not used to calculate the point-level orifice friction and point-level rupture pressure of the specific cluster.

[0016] The formula for calculating frictional resistance is: In the formula, This is the drag reduction ratio; For water friction, MPa; Fluid displacement, m³ / min; The inner diameter of the tubing is in mm; The length of the pipe is in meters (m). Thickener concentration, kg / m³; The value represents the proppant concentration, in kg / m³.

[0017] Based on the on-site construction data, the fracture pressure is obtained at the moment of crack initiation according to the above relationship. Only when the crack initiation event can be clearly mapped to a specific cluster location is the fracture pressure taken as the point-level true rupture pressure of that cluster; otherwise, it is saved as a segment-level first crack initiation event label, and the label type and quality mark are recorded.

[0018] After obtaining the above parameters, these data need to be preprocessed to facilitate subsequent model calculations.

[0019] Well logging parameters, rock mechanics, and geostress parameters are the input features. The preprocessing steps are missing value completion, outlier removal, and normalization. These preprocessing steps are standard practices in the field. For example, missing value completion can be done using Lagrange interpolation, outlier removal can be done using box plot interquartile range, and normalization can be done using minimum-maximum normalization. Other similar methods can also be used by those skilled in the art.

[0020] Logging parameters and rock mechanics input characteristics can be filled with missing values, masked for outliers, and normalized according to the above rules; then, wellhead-to-bottom conversion should be performed in physical quantity units; the labels of fracture initiation time, fracture pressure and actual fracture pressure confirmed manually or by rules should not be interpolated across the fracture initiation abrupt change point, nor should they be automatically deleted simply because of statistical outliers, and the original values, quality marks and audit records should be retained.

[0021] S2. An improved temporal convolutional network is established and trained using preprocessed historical data to obtain a trained network model. The improved temporal convolutional network is based on a temporal convolutional network and uses the physical baseline of rupture pressure as the prior center for gated residual fusion. Parallel receptive fields are set according to three engineering scales: local heterogeneity, inter-cluster variation, and segment-level variation. The conditional mean and prediction standard deviation are output synchronously through the heteroscedasticity probability head.

[0022] All three parallel branches employ multi-layer dilated causal convolutional residual blocks. Each branch contains multiple layers of dilated causal convolutions, with each layer convolving the hidden features of the previous layer according to the dilation rate corresponding to that branch. The result of the convolution transformation is added to the hidden features of the previous layer through direct mapping. When the number of channels is inconsistent, a trainable linear projection is used to ensure dimension consistency. Each layer employs causal padding, ensuring that the dilated causal convolution operator only accesses the current position and its preceding hidden features at the current position. Local heterogeneity features are pre-calculated based on the complete candidate well section, according to the following two-sided sliding window, before network input, and are input into the network as spatial auxiliary feature channels; therefore, the causality only limits the unidirectional indexing relationship of the dilated convolution operator, and does not limit the pre-constructed two-sided spatial auxiliary features. This method is geared towards offline spatial prediction and segmented clustering optimization after obtaining logging and engineering parameters for the complete well section, and does not interpret the causality of the convolutional layer as a strictly unidirectional causality of the entire input processing flow.

[0023] Input feature sequences arranged along well depth , No. The first branch The hidden layer output of the dilated causal convolution is: Input sequence Includes discrete well depth location Pre-processed continuous logging characteristics, rock mechanics and geostress characteristics, wellbore trajectory characteristics, spatial location coding, and fracture pressure physical baseline. Local heterogeneity and missing mask channels; does not contain labels of the true rupture pressure to be predicted. .

[0024] In the formula, This is a discrete location index along the well depth; For parallel branch indexes; For network layer indexing; Indices within the convolution kernel; The width of the convolution kernel; The void ratio of the corresponding layer; and These are the convolution weights and biases, respectively. This represents the result of the convolution transformation; This indicates that the residual block ultimately hides the output; This is a linear projection operator applied to the hidden features of the previous layer, taking direct mapping when the number of input and output channels is the same. The hidden features of layer 0 are defined as equal to the input feature vector. When the candidate location grid is consistent with the network input grid, and One-to-one correspondence.

[0025] The void ratio sequence for each branch is uniformly denoted as: The void ratio sequence of the three branches is based on the spatial step size of the candidate location. and kernel width Settings. The approximate physical receptive field of the branch is And the short, medium, and long-scale branches cover 5–20m, 20–60m, and 60–150m respectively. As an example, when… , In this case, short-scale branching can be achieved using a void ratio sequence. Mesoscale branching can be adopted Long-scale branches can be adopted The three-branch outputs are concatenated and then fused using a 1×1 convolution to obtain a multi-scale hidden representation. The general branch index corresponds to the short, medium, and long scale branches, respectively, and the general hole rate sequence in the formula corresponds to the short, medium, and long scale hole rate sequences in turn; the layer number variable in the formula represents the number of holed causal convolutional layers contained in the corresponding branch.

[0026] To explicitly characterize the local heterogeneity near the candidate location, we first let ξ represent the original continuous features before splicing, and then... The candidate positions and the first Calculate the local heterogeneity index for each selected feature. . The width of the sliding window. This serves as the position index within the window. For multi-feature scenarios, the local heterogeneity index for each feature channel is calculated separately, and then concatenated with the original input features, spatial location encoding, rupture pressure physical baseline, and missing mask to form the final input feature vector. The local heterogeneity indices corresponding to each selected feature are concatenated in feature channel order to form the local heterogeneity vector for that candidate position, used in the following concat operation.

[0027] In the formula, Indicates spatial location encoding, This indicates a missing mask. When the mask is located at the ends of the sequence and the sliding window is insufficient, it is filled by boundary copying; the same boundary handling rules are used in the training and prediction phases.

[0028] The mean parameter of the probability distribution is obtained by fusing the physical baseline of the fracture pressure and the data residual branch through a gating method to avoid over-correction of sparse regions of the data by the data branch: In the formula, The non-homogeneous correction residuals for the data branch outputs; the gating coefficient is defined as follows: The value range is ; This represents the mean value of the rupture pressure condition output by the model. The smaller the gating coefficient, the closer the mean output is to the physical baseline of the rupture pressure; the larger the gating coefficient, the stronger the data-driven correction effect.

[0029] The uncertainty head first outputs the unconstrained scaling parameters. The non-negative prediction standard deviation is then obtained through softplus transformation: Therefore, the first The prediction object for each candidate location is the real rupture pressure random variable. The model outputs the parameters of the conditional probability distribution of the random variable, rather than a single deterministic point value: The explicit output vector of the model is denoted as .in, This represents the mean value of the rupture pressure condition under given candidate location input features and rupture pressure physical baseline conditions; This represents the standard deviation of conditional predictions learned and calibrated by the heteroscedasticity probability head, primarily reflecting sample noise and unexplained random fluctuations under given input conditions. A single heteroscedasticity output head does not independently identify model cognitive uncertainty.

[0030] The purpose of this step is to simultaneously provide an actionable rupture pressure point prediction and confidence level for each candidate location. Under a normal conditional distribution, the conditional mean can be used as a risk-neutral point prediction; given an upper confidence level... When the prediction interval is constructed, the corresponding standard normal quantile coefficient can be used. Compared with the model that directly outputs the single point value using only the mean square error, this model uses joint learning of probability and likelihood to make the prediction error of the same point have different weights under different noise levels, and can distinguish candidate positions with "similar prediction values ​​but different confidence levels".

[0031] The above model requires the use of the physical baseline of fracture pressure. In this step, the physical baseline of fracture pressure is calculated based on the analytical solution of the elasticity of the surrounding rock of the wellbore. The physical baseline of fracture pressure serves as the physical guiding prior for the subsequent probabilistic prediction model. The specific calculation method and process are as follows: First, the effective stress of pore pressure is corrected for the in-situ principal stress tensor. Then, the effective stress tensor in the geographic coordinate system is rotated to the wellbore column coordinate system to reflect the influence of well inclination angle and azimuth angle on the wellbore stress state. In the formula, This is the total stress tensor in the geographic coordinate system; The effective stress tensor in the geographic coordinate system; This is the Biot coefficient; if no measured value is available, it can be set to 1 in the examples. Formation pore pressure; Unit tensor; and These are the azimuth rotation matrix and the well inclination rotation matrix, respectively. Let be the effective stress tensor in the wellbore column coordinate system, and let its relevant components be denoted as . , , and In this context, the inclination angle and azimuth angle are both represented by angle variables with candidate position indices in the above rotation formula, and vary with the candidate position.

[0032] In the wellbore column coordinate system, first calculate the effective tangential stress term at any circumferential angle of the wellbore without applied net wellbore pressure based on the effective stress components: In the formula, This represents the effective tangential stress term when there is no net pressure in the wellbore. This represents the net pressure increment of the wellbore pressure relative to the formation pore pressure. Tensile fracturing of the wellbore occurs when the effective tangential stress reaches a negative value of the rock's tensile strength. By taking the minimum required net pressure for the wellbore's circumferential angle and adding back the formation pore pressure, a physical baseline for the fracturing pressure is obtained, using the same bottom hole pressure reference as the actual fracturing pressure label in the field.

[0033] In the formula, To establish a unified physical baseline for fracture pressure under the bottom hole pressure benchmark; Formation pore pressure; Calculated from the effective stress components in the wellbore coordinate system; This represents the tensile strength of the rock. This avoids directly mixing the net pressure physical baseline with the fracture pressure label under the same bottom hole pressure benchmark.

[0034] Physical baseline Relying solely on the assumption of elasticity and approximation of a homogeneous reservoir, without considering natural fractures, local heterogeneity, and plasticity effects, it serves as the prior center for learning and correcting residuals in the probabilistic prediction model in step S2, rather than the final prediction output.

[0035] The loss function used for the above model is as follows: In the formula, This represents the total loss function value; , , and These represent point-level heteroscedasticity probability loss, adjacent annotation location pressure difference loss, historical segment-level mean-variance statistical consistency loss, and physical consistency regularization term, respectively. , , , For the corresponding weights. This indicates the total number of historically labeled locations with actual fracture pressure level tags. Index it; , and These represent the true label, predicted mean, and predicted standard deviation under standardized scales, respectively. This represents the total number of adjacent location pairs along the well depth that all have point-level labels. For position-pair indexes, and These are the indexes for the locations with smaller and larger well depths, respectively. This represents the total number of historical fracturing sections. For historical segment indexing; , , and These are the predicted mean, actual mean, predicted variance, and actual variance for that historical period, respectively. The number of training positions participating in physical consistency regularization. Index it; This is the amount of data residual correction under standardized scale. For the corresponding gating coefficient, These are empirical prior values ​​for the gating coefficients. and For regularization weights. In the above loss function, point-level pressure variables with `std` subscripts are calculated on the same dimensionless standardized scale to avoid direct addition of the mean error term and variance error term due to different dimensions; the conditional mean, prediction standard deviation, and physical baseline without `std` subscripts represent the denormalized MPa values ​​when entering S3 and subsequent optimizations. The total number of historical fracturing segments only includes fracturing segments with both reliable segment-level mean and variance statistics; single fracturing event records are not used to construct the true variance within a segment. Indicates the first A set of historical fracturing segments is used for segment-level statistical analysis. This is the location index within the set. For historical fracturing segments with at least two reliable point-level true fracture pressure labels, and Calculated using the above point-level label statistical formula; for historical fracturing segments with independently obtained reliable segment-level mean and variance statistics, but lacking sufficient point-level labels. and Independent segment-level statistical values ​​directly converted to a uniform standardized scale are used, where the mean is converted using the same affine normalization transformation as the training labels, and the variance is converted using the square of the corresponding pressure scale factor; it is not required that the values ​​be derived from the standardization process. The summation of the point-level labels within the range is obtained; the corresponding and The values ​​are still calculated based on the model predictions of the corresponding statistical locations of the historical fracturing segment.

[0036] In the above loss function, Used to measure the probabilistic statistical deviation between the predicted distribution and the true value at historical locations with point-level true labels; It is used to measure the degree of consistency between the predicted change in rupture pressure and the actual change in rupture pressure between adjacent marked locations; This is used to directly constrain the predicted mean and intra-segment variance at the historical fracturing segment scale to be consistent with the actual statistics; Constrain the data branch correction magnitude and adjust the gating coefficients towards the empirical prior. Convergence prevents excessive deviation from the physical baseline of rupture pressure in regions of sparse data.

[0037] Compared with the single minimum mean square error loss, the composite loss considers the point-level probability fitting accuracy, spatial variation consistency, historical segment-level statistical consistency, and physical prior constraints. It can reduce the problems of local oscillations, intra-segment statistical bias, and excessive correction of sparse data locations caused by only pursuing point prediction errors. It also provides the conditional mean and prediction standard deviation obtained through probability training for the risk correction pressure calculation in S3.

[0038] After establishing the above model, it is necessary to train it. In this step, the training set used is the historical data obtained in S1.

[0039] S3. Generate a set of candidate locations consisting of multiple candidate locations in the horizontal well section to be optimized, and input the candidate locations into the trained network model to obtain the conditional mean and prediction standard deviation of each candidate location, and then calculate the risk correction pressure of the candidate location; each candidate location record includes candidate coordinates, improved temporal convolutional network input feature vector and fracture pressure physical baseline.

[0040] In this step, the candidate positions are obtained as follows: Within the horizontal well section to be optimized, candidate location coordinates are generated according to a predetermined spatial step size. For each candidate location, logging parameters, rock mechanics and geostress parameters, wellbore trajectory parameters, and a missing mask are extracted or interpolated from the same data source as in the training phase. Local heterogeneity features, spatial location codes, and fracture pressure physical baselines are calculated, and input feature vectors are constructed according to the features specified in S2. All candidate locations are arranged in ascending order of coordinates to form a candidate location input sequence. The candidate coordinates, input feature vectors, and physical baselines together constitute a complete candidate location record. The total number of candidate locations is denoted as . , No. The coordinates of the candidate positions are , =1,…, .

[0041] The predetermined step size is set based on the length of the horizontal well section to be optimized, the required accuracy, and the computing power, and is usually 0.5 to 2.0 m. When generating candidate locations, only fixed restricted areas, casing coupling adjacent areas, abnormal well diameter areas, and fixed restricted areas at both ends of the horizontal well section to be optimized are removed; the end margin of each fracturing section must be determined separately for each outer layer individual after the specific section boundary is decoded in S4, and is not pre-removed in S3.

[0042] The candidate position input sequence, sorted by coordinates, is input as a whole into the network model trained by S2. When the sequence is too long and a sliding subsequence prediction is used, the overlap length of adjacent subsequences must not be less than the receptive field of the corresponding branch, and the same deterministic merging rule is applied to the overlapping positions. After the model makes one prediction, it obtains the conditional mean and prediction standard deviation of all candidate positions, and then calculates the risk correction pressure point by point. The candidate position input sequence is denoted as: In S3, the conditional mean, prediction standard deviation, and risk correction pressure for each candidate position only need to be calculated and saved once before entering the segment cluster optimization. Different individuals in the outer NSGA-II directly read the above results for the corresponding candidate positions during the inner layer evaluation, without needing to repeat training or repeat network prediction for each genetic individual.

[0043] After obtaining the candidate location set, each candidate location in the set is substituted into the network model trained by S2 for calculation and solution, to obtain the conditional mean and prediction standard deviation of the candidate location, and then combined with the following formula to obtain the risk correction pressure: In the formula, Indicates the first Risk correction pressure for each candidate position; , κ ≥ 0 represents the conditional mean and the predicted standard deviation after inverse normalization and calibration, respectively; κ ≥ 0 represents the risk coefficient.

[0044] Regarding the risk coefficient, under the normality assumption, let ,make Corresponding to the preset upper confidence level The quantile value can also be determined jointly by the validation set and engineering risk preference. This represents the cumulative distribution function of the standard normal distribution. Describing its inverse function, it usually satisfies .

[0045] S4. Determine the segmentation and clustering scheme of the horizontal well section to be optimized based on the inner and outer layer fracturing segmentation and clustering optimization model: The outer layer uses the NSGA-II genetic algorithm to search for the number of fracturing segments, the boundary of each fracturing segment, and the number of clusters in each fracturing segment; When evaluating each individual in the outer layer population, the inner layer calls multi-label dynamic programming and combines it with risk correction pressure to determine the location of fracturing clusters within the segment under the given segment boundary, number of clusters, and mean-variance tradeoff coefficient; After the outer layer completes the non-dominated sorting, it outputs Pareto candidate solutions and determines the final scheme from the Pareto candidate solutions according to the preset engineering preferences.

[0046] In this step, the outer NSGA-II genetic algorithm is based on the number of fracturing stages. The proportion of each segment length and the number of clusters in each segment Let be the decision variable. To ensure compatibility with variable segment number search, a fixed-length activation gene encoding is used, with a maximum allowed segment number of . , No. The outermost individual is coded as follows: in, and These represent the coordinates of the start and end points of the proposed horizontal fracturing segment, respectively. This is the total length of the horizontal segment. Only the first... The segment length ratio gene and cluster number gene are in an activated state; the first The length of each fracturing segment is denoted as . The corresponding segment length ratio is ,and Simultaneously, it must also satisfy the maximum number of segments allowed by the fixed-length activation code. The final feasible number of segments must simultaneously satisfy this coding constraint and the geometric constraints given by the segment length range below. The right boundary of the segment is ,in , Let the allowable length range of a single segment be... ,but The number of fracturing stages satisfies . No. The number of segment clusters satisfies and the total number of clusters constraint, where and These represent rounding up and rounding down, respectively. The fracturing segment boundary adopts a half-open interval rule, that is, except for the last segment, each segment includes the starting boundary but does not include the ending boundary; the last segment includes both the starting boundary and the ending boundary to avoid candidate positions located at the common boundary being reused by two fracturing segments.

[0047] The algorithm execution steps are as follows: 1. Parameter initialization: Set the population size Maximum number of generations Crowding comparison parameters, diversity threshold Response speed coefficient and and the upper and lower bounds of the crossover probability , and mutation probability upper and lower limits , .

[0048] 2. Initialize the population: within the range of segment length Cluster number range Randomly generated under equal constraints An initial population of individuals is obtained by performing feasibility repairs on individuals that do not meet the constraints.

[0049] 3. The process of determining the boundary ratio of each segment is as follows: the genetic individual directly generates the gene corresponding to the segment length ratio of the activated segment. After initialization, crossover, or mutation, press Normalization is performed; then, constraint projection repair is performed based on the upper and lower limits of the segment length, satisfying the upper and lower bounds of the proportion and a sum of 1; finally, the cumulative formula is used to decode the boundaries of each segment. Therefore, segment boundaries are not obtained by dividing the total length by the number of segments, but rather by the evolution of genes within the feasible region according to the proportion of segment length, combined with the conditionally optimal cluster positions returned by the inner layer. ~ The evaluation results are confirmed. Segment length ratio repair employs a constraint projection that simultaneously satisfies the upper and lower bounds of the ratio and a sum of 1, or an equivalent method of repeatedly allocating the remaining ratios, until all activation ratios simultaneously satisfy the upper and lower bounds and normalization constraints; it is not permissible to simply truncate each ratio individually and then normalize it once, to avoid exceeding the bounds again after normalization. Crossover and mutation always act on the above segment length ratio genes, maintaining consistency with the outer layer individual coding.

[0050] 4. Fitness assessment (two objectives): For the first objective in a population... outer layer individual Decoding Each fracturing section boundary and the number of clusters in each section For each fracturing section The system filters candidate locations whose coordinates fall within the boundary of the segment and meet the requirements of end margin and no-catch zone, and calls the inner multi-label dynamic programming to pre-calculate the coordinates in S3. and Based on this, the optimal cluster position for this condition is determined, and the cumulative statistics are returned to the terminal. , , and If any fracturing segment lacks a terminal label that satisfies all specified constraints, or if the outer layer individual violates hard constraints such as segment length, cluster spacing, end margin, forbidden zone, and total number of clusters, then the outer layer individual is marked as infeasible. The outer layer employs a feasibility-first constraint dominance rule: feasible individuals are always prioritized over infeasible individuals; between two infeasible individuals, only the total constraint violation is compared, with the one having the smaller total constraint violation being retained; only when both individuals are feasible is a Pareto dominance comparison performed based on the two objective functions.

[0051] F1 represents the average risk correction pressure for the entire well (the lower the better). F2 represents the mean variance of the risk-corrected pressure across the entire well section (the smaller the better). In the formula, This represents the total number of clusters for the current outermost individual. Let... Indicates the first Within the first fracturing segment The global candidate location index selected by each fracturing cluster, then , , , The total number of constraint violations is denoted as... The cumulative statistics at the segment level are derived from the degree of violation of hard constraints such as segment length, cluster spacing, end margin, forbidden zone, total number of clusters, and the absence of feasible paths in the inner layer. These four segment-level cumulative statistics are returned by the outer layer individual after calling the inner layer's dynamic programming function; they are not parameters pre-defined by the outer layer.

[0052] The following modules have also been added to the outer NSGA-II genetic algorithm: 1. Non-dominated ranking: The population is stratified according to dominance relationships. If individual X is not inferior to individual Y in all components (F1, F2) and has at least one strictly superior component, then feasible individual X is said to Pareto dominate feasible individual Y in the target space. Ranking is performed first based on feasibility-priority constraint dominance: feasible individuals are superior to infeasible individuals; two infeasible individuals are compared only based on their total constraint violations; only feasible individuals are non-dominated stratified according to the above Pareto dominance relationships. Individuals not dominated by other feasible individuals form the first Pareto front, and stratification is completed sequentially.

[0053] 2. Crowding Distance Calculation: Within the same non-dominated layer, calculate the crowding degree of each individual in the target space. In the formula, Represents an individual The degree of congestion; Index for the objective function; and They represent the first time. The first and last adjacent individuals located on either side of the current individual after sorting under each objective function; and These represent the current non-dominated layer. The maximum and minimum values ​​of each objective function are determined. For boundary individuals sorted by any objective dimension, their crowding distance is set to infinity. When the maximum and minimum values ​​of a certain objective dimension are equal, the contribution of that objective dimension to the crowding distance of non-boundary individuals is set to 0 to avoid a denominator of 0.

[0054] 3. Population diversity measurement and adaptive parameter adjustment: After normalizing each objective, the first objective is calculated. Generational population diversity The crossover and mutation probabilities are dynamically adjusted within preset upper and lower limits using the following formula: Each target is first normalized by the difference between its maximum and minimum values ​​in the current population; when the maximum and minimum values ​​are equal, the normalization difference for that target is set to 0. Population diversity is calculated using the average of the pairwise distances between normalized target vectors, further normalized according to the maximum possible distance between the two target vectors. In the formula, the target vector represents a vector composed of two normalized targets; the summation subscripts represent the indices of individuals from two different populations. The resulting population diversity ranges from 0 to 1.

[0055] In the formula, For the first Generational population diversity, For diversity threshold, , This refers to the response speed coefficient; , and , These are the upper and lower limits of the crossover and mutation probabilities, respectively. The crossover probability increases when... The probability of mutation increases over time.

[0056] 4. Genetic manipulation: Tournament selection was used; for continuous variables With probability Perform simulated binary crossover on discrete variables Perform discrete recombination; with probability Polynomial mutation and random reset mutation are performed on the offspring; feasibility repair is performed on the offspring after mutation and crossover. The fracturing segment number gene is randomly reset by integer or mutated by adjacent integers; when the number of segments increases, the segment length ratio gene and cluster number gene corresponding to the newly added segment are activated; when the number of segments decreases, only the gene corresponding to the current segment number is used, and the other genes are not included in the calculation. Subsequently, segment length ratio, cluster number range, total cluster number and feasible segment number repair are performed uniformly.

[0057] 5. Elite Retention: [This refers to a specific policy or measure, likely related to the preservation of elites.] Parental generation population (scale ) and offspring population (scale ) merged into scale For mixed populations, the non-dominated ranking and crowding calculation are re-executed, and the top populations are selected according to the rule of "hierarchical priority, and same-level crowding priority". Individuals constitute the parent population of the next generation.

[0058] 6. Termination criterion: Reaching the maximum number of generations. The process terminates, outputting all individuals in the first non-dominated layer of the current population, forming the Pareto candidate solution set for the outer segment structure, for subsequent comprehensive comparison. After normalizing the target values ​​in the first non-dominated layer, a comprehensive minimum distance rule or a pre-determined engineering preference weight can be used for comprehensive comparison to output a final executable segmentation and clustering scheme; the engineering preference rule is determined before optimization begins and is not adjusted afterward based on the results.

[0059] The aforementioned non-dominated sorting, crowding distance, tournament selection, simulated binary crossover, and elite retention are standard operations of NSGA-II; the key technical implementation of this method for the segmented clustering problem lies in: using the number of fracturing stages. Hybrid variable encoding combined with fixed-length activation genes enables joint feasibility repair based on segment length and cluster number; nested calls to inner-layer cluster position optimization are made during the evaluation of each outer-layer individual, instead of selecting an "optimal individual" from the outer population before running the inner layer; and population diversity is utilized. The bounded crossover / mutation probability is driven by the outer layer, which determines "how many segments to divide the structure, where the boundaries of each segment are located, and how many clusters to place in each segment." The inner layer, after fixing the outer layer structure, determines "which candidate location each cluster should be placed in." Therefore, the resulting Pareto candidate solutions belong to the inner-outer layer conditional Pareto solutions given the inner layer mean-variance tradeoff coefficient and constraint threshold.

[0060] The mean-variance tradeoff coefficient is uniformly denoted as... To ensure that this coefficient has a clear and repeatable calculation caliber, it is preferable to use the arithmetic mean of the risk correction pressures at all candidate locations in the well section to be optimized, obtained from S3, as the reference pressure scale. : In the formula, All well sections to be optimized The arithmetic mean of the risk correction pressure for each candidate location. Because The unit is MPa, therefore The unit is MPa - ¹, make · , It has the same dimensions as the mean of risk-corrected pressure. The calculation is performed once before entering the outer NSGA-II optimization and remains unchanged during a segmented cluster optimization process, thus avoiding arbitrary adjustments to the trade-offs during the evolution process.

[0061] The inner-layer mean-variance multi-label dynamic programming is a specialization of the resource-constrained shortest path labeling method. For any given fracturing segment in the outer layer, it solves for the optimal cluster position combination that satisfies the cluster spacing constraint on the candidate point sequence within that segment. The fracturing pressure variance contains a cumulative squared term, which is path-dependent and non-additive statistical objective. Conventional dynamic programming that accumulates point-by-point cannot directly handle this; therefore, the following multi-label state dynamic programming is adopted.

[0062] The inner algorithm is a multi-label dynamic programming / resource-constrained shortest path labeling algorithm. For the path-related objective of mean-variance risk correction pressure within a segment (containing a cumulative squared term), the cumulative risk correction pressure sum, the cumulative risk correction pressure sum of squares, the sum of prediction standard deviations, and the maximum risk correction pressure are used as sufficient statistical labels, and cluster spacing, forbidden zones, end margins, and uncertainty threshold constraints are embedded. The four statistics corresponding to the cumulative risk correction pressure sum, the cumulative risk correction pressure sum of squares, the sum of prediction standard deviations, and the maximum risk correction pressure are mapped to... , , and Return to the outer layer.

[0063] State definition: For the current fracturing section Index of the last selected candidate position Number of clusters already placed "This represents a stage state, where several feasible path labels are retained within the same state:" in, This indicates the number of fracturing clusters already placed along the current path; Label numbers representing different feasible paths under the same condition; Indicates the sequential index of the selected fracturing clusters in the path; Indicates the first The first path The global candidate location index corresponding to each selected fracturing cluster. This indicates the cumulative risk correction pressure up to the current state. This represents the sum of squared cumulative risk correction pressures. This represents the sum of the cumulative forecast standard deviations. This represents the maximum risk-corrected pressure among the selected fracturing clusters as of the current state. The four path statistics are calculated using the following formula: Initialization: For each feasible candidate location of the first cluster within the fracturing section that simultaneously satisfies the fixed forbidden zone and the end-end margin constraints, an initial label with a cluster count of 1 is established: Specifically, in the initial label above, the first and fourth parameters are the same. This is because when the number of clusters is 1, = = .

[0064] State transition: from candidate position (Already placed) (each cluster) is moved to the next candidate position. Cluster spacing constraints must be met. Furthermore, the next candidate position does not fall into the restricted area and satisfies the end margin constraint of the fracturing section; the label is updated according to the following formula: where the index of the next candidate position indicates the next optional candidate position whose coordinates are greater than the coordinates of the current candidate position; the new label number is reassigned from the same state after the transfer.

[0065] in, The apostrophe in the text is only used to indicate the index of the next possible candidate position whose coordinates are greater than the current candidate position, and does not indicate an exponentiation; This indicates the label number that is reassigned in the new state after the state transition is completed; The number of clusters placed is indicated by Increase to .

[0066] Intra-segment objectives and end-game costs: When the current fracturing segment requires... After all clusters are placed, the mean, variance, and inner cost of the risk correction pressure within the corresponding path segment can be directly obtained from the terminal's cumulative statistics. In the formula, , and These represent the risk-corrected mean pressure, overall variance, and inner layer cost for the fracturing section, respectively. This is the mean-variance tradeoff coefficient, and its dimension is MPa. - ¹; The upper limit of the allowed intra-segment variance; The upper limit of the allowable cluster mean prediction standard deviation; The maximum allowable risk correction pressure for a single cluster. The cumulative amount of the terminal tag is mapped as follows. , , and Simultaneously, it is returned to the outer layer for evaluation of the objective function and constraints. The sum of the predicted standard deviations of this segment divided by the number of clusters in this segment represents the arithmetic mean of the predicted standard deviations of each selected cluster, and is not interpreted as the joint standard deviation of multiple random variables.

[0067] The inner-layer solution objective is to first select feasible labels from the terminal labels that satisfy the constraints of cluster spacing, forbidden zones, and end margins, and then further filter those that simultaneously satisfy the following constraints: average prediction standard deviation, maximum risk correction pressure per cluster, and intra-segment variance. In the above finality constraints, For the first The sum of the prediction standard deviations of each cluster selected in the segment. This is the number of clusters in that segment, therefore This indicates the standard deviation of the cluster average prediction for this segment; Set its maximum allowable limit. This indicates the maximum value of the risk correction pressure for the selected single cluster in this segment. This is the maximum allowed limit; This indicates the overall variance of the risk-corrected stress in this segment. This represents the upper limit of allowed intra-segment variance. Only terminal labels that satisfy all three constraints can proceed to the inner cost comparison.

[0068] Compare the inner costs among all feasible terminal labels and select the label with the lowest inner cost.

[0069] Label safety pruning rules: Terminal variance is determined by both the risk correction pressure and its sum of squares, and does not exhibit general component monotonicity with respect to the cumulative risk correction pressure. Therefore, when performing an exact solution under the same conditions, the comparison condition for deleting the second label is satisfied only if both labels simultaneously satisfy all of the following relationships: In the above formula, the superscripts 1 and 2 only represent the numbers of the two labels to be compared in the same state, and do not represent squares or other powers. Label comparison adopts a one-by-one, full-scale comparison method, rather than randomly selecting one from the retained labels for comparison: each new label is generated and compared sequentially with all the labels already retained in the current state. If the cumulative risk correction pressure of a retained label is equal to that of the new label, and its cumulative risk correction pressure square sum, cumulative prediction standard deviation sum, and maximum risk correction pressure are all no greater than the new label, and at least one of the latter three is strictly smaller, then the new label is directly deleted; otherwise, if the new label satisfies the above-mentioned dominance relationship with a retained label, then the retained label is deleted; if the two labels do not satisfy all relationships, both are retained. After all new labels in the same state have been compared, the subsequent state transition begins. Only when the number of clusters placed reaches the required number of clusters for the segment and passes the final constraints such as average prediction standard deviation, maximum risk correction pressure per cluster, and intra-segment variance, is the inner-layer cost compared among all feasible terminal labels, and the specific label with the smallest inner-layer cost is returned to the outer layer. If the cumulative risk correction pressure is different, then no label will be deleted under precise safety pruning; if ε approximate merging or a label number limit is used to control the computational load, then the inner layer result is the conditionally optimal solution under the corresponding approximate precision.

[0070] Backtracking: When the number of clusters placed reaches the required number for the fracturing segment, select the label with the minimum inner-layer cost from the retained labels that satisfy all final constraints, and backtrack along the transfer path recorded by that label to obtain the conditionally optimal cluster position sequence for the fracturing segment. Given the segment boundary, number of clusters, candidate point set, inner-layer tradeoff coefficient, and numerical precision, and under the premise of precise safety pruning and no upper limit on the number of labels, the result is the globally optimal solution within the segment under the outer structure conditions; when approximate merging or label truncation is used, the result is the conditionally optimal solution under the corresponding approximate precision. When multiple terminal labels have the same minimum inner-layer cost, select the label with the smaller maximum risk correction pressure, the smaller sum of prediction standard deviations, and the earlier candidate position index sequence to ensure that the output is deterministic and repeatable.

[0071] The present invention has been disclosed above with preferred embodiments. However, those skilled in the art should understand that these embodiments are only for describing the present invention and should not be construed as limiting the scope of the present invention. Further improvements can be made without departing from the principles of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A physically guided, strong heterogeneous reservoir sub-segmentation clustering intelligent optimization method, characterized in that, Includes the following steps: S1. Obtain historical data of the target block and preprocess it. The historical data includes logging parameters, rock mechanics and geostress parameters, wellbore trajectory parameters, fracturing operation time data, and historical segment cluster structure and constraint data. S2. An improved temporal convolutional network is established and trained using preprocessed historical data to obtain a trained network model. The improved temporal convolutional network is based on a temporal convolutional network and uses the physical baseline of the fracture pressure as the prior center for gated residual fusion. Parallel receptive fields are set according to three engineering scales: local heterogeneity, inter-cluster variation, and segment-level variation. The conditional mean and prediction standard deviation are output synchronously through the heteroscedasticity probability head. The physical baseline for the rupture pressure is: wherein, represents a physical baseline for fracture pressure; represents the formation pore pressure at the candidate location; represents the wellbore circumferential angle, taking values in the range [0, 2π); represents the effective tangential stress term at the candidate location at circumferential angle without the application of wellbore net pressure; represents the rock tensile strength at the candidate location; represents the angle within the wellbore circumferential angle range that minimizes the required wellbore net pressure; S3. Generate a set of candidate locations consisting of multiple candidate locations in the horizontal well section to be optimized, and input the candidate locations into the trained network model to obtain the conditional mean and prediction standard deviation of each candidate location, and then calculate the risk correction pressure of the candidate location; each candidate location record includes candidate coordinates, improved temporal convolutional network input feature vector and fracture pressure physical baseline; S4. Determine the segmentation and clustering scheme of the horizontal well section to be optimized based on the inner and outer layer fracturing segmentation and clustering optimization model: The outer layer uses the NSGA-II genetic algorithm to search for the number of fracturing segments, the boundary of each fracturing segment, and the number of clusters in each fracturing segment; When evaluating each individual in the outer layer population, the inner layer calls multi-label dynamic programming and combines it with risk correction pressure to determine the location of fracturing clusters within the segment under the given segment boundary, number of clusters, and mean-variance tradeoff coefficient; After the outer layer completes the non-dominated sorting, it outputs Pareto candidate solutions and determines the final scheme from the Pareto candidate solutions according to the preset engineering preferences; The NSGA-II genetic algorithm uses the number of fracturing segments, the proportion of each activated fracturing segment length, and the number of clusters in each activated fracturing segment as outer-layer decision variables, and employs fixed-length activation gene encoding. When evaluating the fitness of each outer-layer individual, candidate positions that meet the end margin and no-displacement zone constraints are screened within each fracturing segment. The inner-layer multi-label dynamic programming is then invoked to construct the mean and variance evaluation of the risk correction pressure within the segment based on the pre-calculated risk correction pressure, and the predicted standard deviation is used as an uncertainty constraint. Under a given mean-variance tradeoff coefficient, the optimal cluster position is determined. The outer layer calculates the overall well average risk correction pressure and the mean variance of the risk correction pressure within the entire well segment based on the risk correction pressure and the sum of squares of the risk correction pressure returned by the inner layer, thereby performing a dual-objective fitness evaluation. The optimal solution is output based on the fitness evaluation value. The optimal cluster is determined in the inner layer by the following formula: wherein represents the inner layer cost; represents the mean of the risk-corrected pressure; represents the overall variance; represents the mean-variance trade-off coefficient predetermined according to the risk-corrected pressure reference scale of the well section to be optimized; and the solution with the minimum inner layer cost is returned to the outer layer; The objective function in the outer layer is: , wherein, represents the full well average risk-corrected pressure; represents the full well average risk-corrected pressure variance; represents the total cluster number of the current outer individual; represents the number of fracture stages of the current outer individual, represents the fracture stage number; represents the cumulative risk-corrected pressure sum of represents the cumulative risk-corrected pressure sum of represents the cumulative risk-corrected pressure square sum of represents the cluster number within​​ 2. The physically guided intelligent optimization method for segmented clustering of highly heterogeneous reservoirs according to claim 1, characterized in that, The logging parameters include density, sonic transit time, natural gamma, compensated neutrons, wellbore diameter, and resistivity. The rock mechanics and geostress parameters include Young's modulus, Poisson's ratio, brittleness index, maximum horizontal principal stress, minimum horizontal principal stress, pore pressure, and tensile strength. The wellbore trajectory parameters include well depth, vertical depth, inclination angle, and azimuth angle. The fracturing operation data includes wellhead pressure, flow rate, fluid column height, or fluid column pressure and friction data. The historical cluster structure and constraint data include segment boundaries, cluster locations, cluster numbers, cluster spacing, forbidden zones, and end margins.

3. The physically guided intelligent optimization method for segmented clustering of highly heterogeneous reservoirs according to claim 1, characterized in that, The input feature vector of the improved temporal convolutional network includes: preprocessed logging parameters, preprocessed rock mechanics and geostress parameters, preprocessed wellbore trajectory parameters, and fracture pressure physical baseline.

4. The physically guided intelligent optimization method for segmented clustering of highly heterogeneous reservoirs according to claim 1, characterized in that, The improved temporal convolutional network is configured with three parallel dilated causal convolutional branches: short-scale, medium-scale, and long-scale. Each branch consists of stacked multi-layer dilated causal convolutional residual blocks, and each branch employs a dilation rate sequence. , and Based on the spatial step size of the candidate location kernel width The void ratios of each layer are used to ensure that the equivalent receptive fields of the three branches cover engineering scales of 5–20 m, 20–60 m, and 60–150 m, respectively. The outputs of the three branches are spliced ​​and then fused through 1×1 convolution to obtain a multi-scale hidden representation. The conditional mean is obtained by fusing the physical baseline of the rupture pressure and the data residual branch through a gating method, and the prediction standard deviation is obtained by softplus transformation of the independent scale parameters.

5. The physically guided intelligent optimization method for segmented clustering of highly heterogeneous reservoirs according to claim 1, characterized in that, In S2, the loss function of the improved temporal convolutional network is as follows: In the formula, This represents the total loss function value; , , and These represent point-level heteroscedasticity probability loss, adjacent annotation location pressure difference loss, historical segment-level mean-variance statistical consistency loss, and physical consistency regularization term, respectively. , , , They are respectively , , and The weights; This indicates the total number of historically labeled locations with actual fracture pressure level tags. Index it; , and These represent the true fracture pressure label, predicted mean, and predicted standard deviation at a standardized scale; locations adjacent along well depth and both having point-level labels are... Number, among which and These represent the indexes of the marked locations for smaller and larger well depths, respectively; This represents the total number of historical fracturing sections. For historical segment indexing, , , and These are the predicted mean, actual mean, predicted variance, and actual variance for that historical period, respectively. The number of training positions participating in physical consistency regularization. Index it; This represents the correction amount of the data residual branch relative to the physical baseline under standardized scaling. To correspond to the adaptive gating coefficient, These are empirical prior values ​​for the gating coefficients. and For regularization weights; point-level pressure variables with std subscripts all represent quantities under the same dimensionless standardized scale; Indicates the first A set of historical fracturing segments is used for segment-level statistical analysis. This is the position index within the set.

6. The physically guided intelligent optimization method for segmented clustering of highly heterogeneous reservoirs according to claim 1, characterized in that, In S3, within the horizontal well section to be optimized, coordinates of multiple candidate locations are generated according to a preset spatial step size, and logging parameters, rock mechanics and geostress parameters, and wellbore trajectory parameters of all candidate locations are statistically analyzed.

7. The physically guided intelligent optimization method for segmented clustering of highly heterogeneous reservoirs according to claim 1, characterized in that, No. The risk correction pressure for each candidate position is: In the formula, Indicates the first Risk correction pressure for each candidate position; and These are the conditional mean and predicted standard deviation after inverse normalization and calibration, respectively. This represents the risk coefficient.

Citation Information

Patent Citations

  • Deep learning assisted acceleration fracturing construction parameter intelligent real-time optimization method

    CN121091692A

  • Horizontal well multi-cluster fracturing crack propagation prediction and parameter optimization method and device

    CN121881450A