A method for evaluating the compressibility of continental shale based on conditional generative adversarial network

Abstract

The application discloses a method for evaluating the compressibility of continental shale based on a conditional generative adversarial network. The method comprises the following steps: obtaining the compressive strength, Young's modulus and fracture toughness mechanical parameter samples of cores at a plurality of discrete depth points of a target well; performing Savitzky-Golay filtering on full-well continuous logging curves to extract low-frequency trend items, and using the well depth of the discrete depth points and the low-frequency trend items as components to construct conditional variables; constructing a conditional generative adversarial network composed of a generator and a discriminator, and performing alternating adversarial training by using a combined loss function composed of an adversarial loss, a reconstruction loss and a gradient penalty; using the trained generator to predict the full-well continuous mechanical parameter profile; and after normalization, using the analytic hierarchy process and the entropy method to combine weighting to calculate the compressibility index, so as to realize the classification of the reservoir compressibility. The application combines the discrete core measured data and the continuous logging data, overcomes the excessive smoothing of the traditional regression, and improves the objectivity and precision of the compressibility evaluation.

Description

Technical Field

[0001] This invention relates to the field of unconventional oil and gas reservoir evaluation technology, specifically to a method for evaluating the compressibility of continental shale based on conditional generative adversarial networks. Background Technology

[0002] As an important unconventional oil and gas resource, the large-scale and efficient development of shale oil relies on the accurate assessment of reservoir compressibility. Continental shale is characterized by small-scale laminar flow, diverse mineral composition, and significant mechanical heterogeneity, leading to extremely complex reservoir mechanical behavior and fracture propagation mechanisms. Accurately obtaining and continuously characterizing key mechanical parameters of continental shale reservoirs, such as compressive strength, Young's modulus, and fracture toughness, is fundamental to conducting compressibility assessments.

[0003] Existing technologies for evaluating the compressibility of continental shale mainly have the following shortcomings:

[0004] First, the high-precision static mechanical parameters obtained from core experiments are discretely distributed along the well depth, and due to the constraints of core sampling conditions, there are missing samples, making it difficult to construct a continuous mechanical profile of the entire well section.

[0005] Secondly, the original logging curves have problems such as large sampling intervals, obvious local fluctuations and random noise interference. If the original logging curves are directly used as the input of the prediction model, the model is likely to learn the pseudo fluctuations caused by noise at the same time, making it difficult for the input signal to stably support the fine characterization of the mechanical parameters of heterogeneous shale.

[0006] Third, the supervised regression methods used in the past, such as multiple linear regression and artificial neural networks, are prone to oversmoothing when dealing with heterogeneous layer data. Although they can fit the overall trend, they are difficult to recover the fluctuations of local layers.

[0007] Fourth, existing compressibility assessments are mostly based on a single brittleness index or rely solely on subjective weighting methods such as the analytic hierarchy process, making it difficult to comprehensively reflect the true fracturing response under the coupled effects of multiple mechanical factors.

[0008] Therefore, there is an urgent need for a method to evaluate the compressibility of continental shale that can integrate discrete core data with continuous logging data throughout the well, suppress noise interference from the original logging curves, avoid excessive smoothing of supervised regression, and objectively integrate the contributions of multiple mechanical parameters. Summary of the Invention

[0009] To address the aforementioned shortcomings of existing technologies, the present invention aims to provide a method for evaluating the compressibility of continental shale based on conditional generative adversarial networks, thereby solving the technical problems of difficulty in making discrete core mechanical parameters continuous, large noise interference in original logging curves, excessive smoothing in traditional supervised regression, and strong subjectivity in compressibility evaluation.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A method for evaluating the compressibility of continental shale based on conditional generative adversarial networks includes the following steps:

[0012] S1. Obtain mechanical parameter samples of core samples at several discrete depth points in the target well. The mechanical parameter samples include compressive strength, Young's modulus, and fracture toughness.

[0013] S2. Perform Savitzky-Golay filtering on the continuous logging curve of the entire target well to extract the low-frequency trend term, and construct conditional variables with the well depth of the discrete depth points and the low-frequency trend term as components.

[0014] S3. Construct a conditional generative adversarial network including a generator and a discriminator; under the constraints of the conditional variables, the generator generates a mechanical parameter vector based on a random noise vector, and the discriminator outputs the probability that the input sample is a real sample;

[0015] S4. Using a joint loss function consisting of adversarial loss, reconstruction loss, and gradient penalty, the generator and the discriminator are trained adversarially to obtain a trained generator.

[0016] S5. Input the conditional variable of the depth point to be predicted and the random noise vector into the trained generator. The resulting mechanical parameter prediction values ​​and the mechanical parameter samples in S1 are merged by depth to form a continuous mechanical parameter profile of the entire well section.

[0017] S6. Normalize the continuous mechanical parameter profile of the entire well section, and use the combined weighting method of analytic hierarchy process and entropy method to obtain the corrected weights of each mechanical parameter. The compressibility index is obtained by weighted summation according to the corrected weights, and the reservoir is classified according to the compressibility index.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] 1. This invention generates mechanical parameters of missing well sections under logging conditions through a conditional generative adversarial network, and merges them with discrete core measured samples according to depth, thereby realizing the effective construction of a continuous mechanical profile of the entire well section;

[0020] 2. This invention uses Savitzky-Golay filtering to extract the low-frequency trend term of the logging curve as a conditional variable. While suppressing high-frequency noise and local anomalies in the original curve, it retains the main variation trend along the well depth direction, thereby improving the ability of the input signal to characterize the macroscopic variation features of the formation.

[0021] 3. This invention uses a joint loss function that combines adversarial loss, reconstruction loss and gradient penalty to learn the variation law of mechanical parameters from the distribution level rather than the point-to-point regression level. While ensuring the overall trend fit, it recovers the local segment fluctuations and improves the restoration accuracy of the mechanical parameters of heterogeneous shale.

[0022] 4. This invention adopts a weighting strategy combining the analytic hierarchy process (AHP) and the entropy method. The AHP introduces the prior knowledge of experts in fracturing engineering practice, while the entropy method provides objective correction based on the dispersion of the whole well section profile data. The correction weight obtained by multiplying and normalizing the two reflects both prior experience and data characteristics, thus achieving an objective synthesis of the contributions of multiple mechanical parameters. Attached Figure Description

[0023] Figure 1 This is a flowchart of the steps of the method described in this invention;

[0024] Figure 2 This is a schematic diagram of the training architecture of the conditional generative adversarial network described in this embodiment of the invention;

[0025] Figure 3 This is a schematic diagram of the network structure of the generator described in an embodiment of the present invention. Detailed Implementation

[0026] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that this embodiment is only used to explain the present invention and is not intended to limit the scope of protection of the present invention. This embodiment can be implemented based on any mainstream deep learning framework such as PyTorch or TensorFlow. To ensure reproducibility, it is recommended to fix all random number seeds during implementation.

[0027] like Figure 1 As shown, the method for evaluating the compressibility of continental shale based on conditional generative adversarial networks provided in this embodiment includes six steps, from S1 to S6, with data flow passed step by step between each step, as detailed below:

[0028] Step S1: Obtaining core mechanical parameter samples.

[0029] This embodiment obtains several core samples from well X1, a continental shale well in the Songliao Basin. The number of core samples is denoted as [number missing]. , It should be a positive integer; to ensure the convergence of subsequent conditional generative adversarial network training, this embodiment suggests... No less than 30. High-precision scratch measurements were performed on the surface of each core sample, and the horizontal and vertical forces acting on the cutter head during the scratching process were collected in real time. Based on conventional scratch mechanics back-calculation methods in this field, the results were obtained. Mechanical parameters at each sampling depth point, among which Number the core sampling depth points. The sample preparation specifications, blade specifications, and test parameters for the scratch test can be performed in accordance with industry standards in this field. In this embodiment, parameters are indexed... Indicates the category of mechanical parameters, Corresponding compressive strength, Corresponding to Young's modulus, Corresponding fracture toughness. Let the first... At the sampling depth point, the first The values ​​of the mechanical parameters are The mechanical parameter samples at this depth point are organized as column vectors.

[0030]

[0031] Record No. The well depth corresponding to each sampling depth point is: It should be noted that the mechanical parameter samples can also be obtained by other indoor mechanical testing methods, and this embodiment does not limit the acquisition method. This step obtains high-precision measured samples of mechanical parameters discretely distributed along the well depth, providing real sample constraints and distribution benchmarks for subsequent conditional generative adversarial network training.

[0032] Since the core samples are discretely distributed along the well depth, they cannot independently support the construction of a continuous mechanical parameter profile for the entire well section. Therefore, continuous condition signals are extracted by logging curves.

[0033] Step S2: Well logging curve filtering and condition variable construction.

[0034] Continuous logging data for the entire target well section is acquired, including P-wave transit time, S-wave transit time, and density logging data. These are then converted using conventional dynamic and static elastic parameter conversion methods to obtain a low-precision continuous logging curve reflecting the reservoir's mechanical characteristics. This low-precision continuous logging curve consists of a compressive strength logging curve, a Young's modulus logging curve, and a fracture toughness logging curve, each corresponding to a parameter index. Given that the low-precision continuous logging curves have problems such as large sampling intervals, obvious local fluctuations, and random noise interference, if they are directly used as conditional inputs to the generation model, the model is prone to simultaneously learning pseudo-fluctuations caused by noise. Therefore, this embodiment uses Savitzky-Golay filtering to extract the low-frequency trend term separately.

[0035] The following is indexed by parameters. The implementation details of Savitzky-Golay filtering are illustrated using a low-precision continuous logging curve as an example. Let the discrete sequence obtained by sampling the curve along the depth be denoted as... , Number the depth sampling points for the well logging curve. , This represents the total number of logging sampling points throughout the entire well section; here With core sampling number in S1 They belong to different indexing systems. Usually much larger In the first Select a length of [length] near each depth sampling point. Sliding window, Half the width of the window The value is a positive integer, and the window length is an odd number to ensure that the window center uniquely corresponds to the sampling point. Construct the offset within the sliding window. of Local polynomial of order

[0036]

[0037] Indicates the point inside the window relative to the center of the window. The offset, ; For polynomial coefficients, Number the terms of the polynomial. ; Let be the order of the polynomial. It must be a positive integer and satisfy the following conditions: To ensure the solvability of the least squares problem, based on the least squares criterion, polynomial coefficients are selected that minimize the following expression.

[0038]

[0039] And with the center position of the sliding window as Fitted values ​​at As the smoothed output of this sampling point, denoted as ,Right now In this embodiment Set the value to 25, and the sliding window length to 51. A value of 3 is chosen to balance noise suppression capability and local trend fidelity; the sliding step size is set to 1, i.e., sliding point by point; boundary sampling points at both ends of the logging curve that are insufficient for the window length are processed using a boundary symmetric extension strategy to ensure that filtering is continuously computable throughout the entire well section. (Parameter index) The above filtering was applied to the three corresponding well logging curves to obtain the first... At the first well logging sampling point Low-frequency trend components corresponding to mechanical parameters The Savitzky-Golay filter suppresses high-frequency noise and local anomalies in the original logging curves while preserving the main trend along the well depth, providing a stable trend input for subsequent conditional variables.

[0040] S1 core sampling depth point ( Mapped to the nearest neighbor logging sampling point, the low-frequency trend component at that logging sampling point is extracted and denoted as... ,Right now ,in for The corresponding well logging sampling point number. The well depth of the discrete depth point is jointly encoded with the low-frequency trend component to construct a condition variable.

[0041]

[0042] The condition variable is a 4-dimensional column vector, with its first and second components... The code encodes spatial location information, while the other three components encode trend information of mechanical parameters derived from well logging data. Considering... The magnitudes of the numerical values ​​differ significantly from those of the normalized trend components. This difference is significant before inputting the data into the generative adversarial network. Each component is standardized by removing the mean and standard deviation of the entire well section to ensure consistent magnitudes and thus stable network training. This completes the training data preparation; the next step is to construct the conditional generative adversarial network.

[0043] Step S3: Construction of the conditional generative adversarial network.

[0044] like Figure 2 As shown, the conditional generative adversarial network consists of a generator. and discriminator Composition. The generator With random noise vector With the condition variable The concatenated vector is the input, and the concatenation method is dimensional concatenation, meaning the input vector has a dimension of 100 + 4 = 104. After parallel processing by the backbone network and skip branches, the output is a mechanical parameter vector, which is denoted as... is a 3-dimensional column vector. The random noise vector... The discriminator is a 100-dimensional column vector that follows a standard normal distribution. During the training phase, it is resampled at each forward propagation to promote the diversity of generated samples. The input is a vector of mechanical parameters and the condition variable. The concatenation method is also dimensional concatenation, meaning the input vector has a dimension of 3 + 4 = 7; the mechanical parameter vector is the real sample. or the generator Output samples The discriminator The output specifies the probability that an input sample is classified as a real sample; this probability value is within an open interval. Inside, recorded as .

[0045] like Figure 3 As shown, the generator The network structure is as follows. The generator... The generator consists of a backbone network and a jump-connected branch, which process the same input vector in parallel and sum the outputs. The backbone network comprises three cascaded hidden layers. Each hidden layer sequentially performs a fully connected operation, a layer normalization operation, a ReLU activation function, and a Dropout layer. The backbone network is then mapped to a 3D output space via a fully connected layer. The jump-connected branch consists of a single fully connected layer that connects the generator... The input vector is directly linearly mapped to the 3D output space. The introduction of the skip connection branch aims to alleviate the gradient vanishing problem in deep network training and provide an approximately linear basis mapping, which helps the generator stably learn the complex nonlinear relationship between mechanical parameters and condition variables. The generator The final output is the element-wise sum of the backbone network output and the jump branch output. In this embodiment, the widths of the three hidden layers of the backbone network are 256, 512, and 256 respectively, and the dropout rate of the Dropout layer is 0.2.

[0046] The discriminator The network structure is as follows. The discriminator... It contains two hidden layers, each of which sequentially performs a fully connected operation, a LeakyReLU activation function, and a Dropout layer. The output layer is a fully connected layer followed by a Sigmoid function, which non-linearly maps the hidden layer output features to an open interval. The discrimination probability is obtained. In this embodiment, the discriminator... The widths of the two hidden layers are 128 and 64, respectively. The negative slope of the LeakyReLU activation function is set to 0.2, and the dropout rate of the Dropout layer is set to 0.3. The conditional generative adversarial network architecture establishes a nonlinear conditional mapping from the random noise space to the mechanical parameter space, providing the network foundation for generating mechanical parameters for missing well sections. Adversarial training is then performed.

[0047] Step S4: Alternate adversarial training using the joint loss function.

[0048] Constructing adversarial loss Reconstruction loss and gradient penalty The joint loss function is composed of subscripts. , , These represent adversarial, reconstruction, and penalty, respectively. Adversarial loss is defined as...

[0049]

[0050] Represents the random noise vector With the condition variable The expectation of the joint distribution is approximated during training by averaging the samples within a batch; the adversarial loss is used to measure the generator. The output samples are discriminated by the discriminator. The degree to which a sample is judged as a real sample; the smaller the value, the more the generator can deceive the discriminator.

[0051] Reconstruction loss is defined as

[0052]

[0053] L1 norm represents the sum of the absolute values ​​of the vector components; Representation and condition variable The corresponding mechanical parameter samples in S1; the reconstruction loss provides sample-by-sample numerical supervision in addition to the adversarial loss, avoiding the training process from deviating from the true sample values ​​by only trying to deceive the discriminator; this embodiment uses the L1 norm, which can also be replaced by the L2 norm in other embodiments.

[0054] Gradient penalty is defined as

[0055]

[0056] For interpolated samples, defined as

[0057]

[0058] To conform to a uniform distribution The random variables are sampled independently on each training sample; Indicates the discriminator output right The gradient is calculated using automatic differentiation; Represents the L2 norm; The gradient penalty coefficient is... For positive real numbers, in this embodiment The value is set to 10. The gradient penalty constraint ensures that the discriminator maintains a unit L2 norm gradient on the connection between real and generated samples, so that the discriminator satisfies the 1-Lipschitz condition to stabilize the training process and effectively alleviate the gradient vanishing and mode collapse problems in traditional generative adversarial network training.

[0059] The generator The total loss function is defined as

[0060]

[0061] To reconstruct the loss weight coefficients, For positive real numbers, in this embodiment Setting it to 100 makes the reconstruction loss and adversarial loss comparable in magnitude. The discriminator... The total loss function is defined as

[0062]

[0063] The classification loss based on binary cross-entropy is defined as follows after introducing a label smoothing strategy:

[0064]

[0065] The target probability label corresponding to the real sample is replaced by 0.9, and the target probability label corresponding to the generated sample is replaced by 0.1, in order to prevent the discriminator from becoming overconfident and causing training instability.

[0066] The generator is trained using an alternating adversarial training strategy. and the discriminator Perform parameter updates. In each iteration, first fix the generator. The parameters are obtained by randomly drawing a batch with replacement from the mechanical parameter sample described in S1 and sampling a new random noise vector for each sample. By the generator Generate corresponding generated samples, and calculate using real samples and generated samples. And then propagate in reverse to update the discriminator. The parameters are then fixed; then the discriminator is fixed. The parameters, resample Generate new generated samples and calculate And then propagate in reverse to update the generator. The parameters.

[0067] The training hyperparameters in this embodiment are explicitly disclosed as follows to facilitate reproduction by those skilled in the art. The optimizer used is the Adam optimizer, whose first-order moment decay coefficient... Take 0.5 as the second-order moment attenuation coefficient. Take 0.999; the generator With the discriminator initial learning rate Set all to A cosine annealing learning rate scheduling strategy is adopted, and the learning rate is adjusted during 5000 iterations. Smooth decay to The training sample batch size is set to 64; the total number of training rounds is fixed at 5000; the generator With the discriminator All weights of the fully connected layers are initialized using Xavier uniform initialization, and the bias terms are initialized to zero. After training, the generator described in the last round of training is saved. The network parameters are used to obtain a trained generator. The joint loss function constrains the generator to learn the variation law of mechanical parameters from the distribution level rather than the point-to-point regression level, which restores the local segment fluctuations while ensuring the overall trend fit, thus improving the accuracy of the restoration of mechanical parameters of heterogeneous shale. The trained generator is then used for prediction and profile fusion.

[0068] Step S5: Mechanical parameter prediction and profile fusion.

[0069] Select the entire well section of the target well, excluding S1. All logging sampling points other than the discrete depth points are used as the depth points to be predicted, meaning the spacing of the set of depth points to be predicted is consistent with the sampling interval of the continuous logging curves for the entire well section. For the first... One depth point to be predicted , Number the depth points to be predicted, and construct corresponding condition variables from the logging data at those depth points according to the method in step S2. Input to the trained generator With the random noise vector The predicted values ​​of the mechanical parameters at that depth point are obtained. , It is a 3-dimensional column vector. To obtain deterministic prediction output, during the evaluation and inference phase, The vector is fixed at zero; if it is necessary to assess the prediction uncertainty, multiple samples can be taken from the same depth point. The mean or variance of the prediction results is then calculated.

[0070] Predicted values ​​of all points at the desired depth The mechanical parameter sample described in S1 According to their respective well depths , After sorting in ascending order and merging, a continuous mechanical parameter profile of the entire well section of the target well is obtained. If the well depth of the point to be predicted coincides with the well depth of the core sampling point in S1, the measured core samples are retained first during merging, and the generated predicted values ​​at that depth are discarded to ensure the accuracy of the profile at the measured depth. This merging process achieves the organic integration of discrete measured core samples and continuous logging prediction values, solving the problem of profile discontinuity caused by discrete and missing core data. The compressibility evaluation and classification are then performed.

[0071] Step S6: Normalization, combined weighting, and compressibility classification.

[0072] Record the first continuous mechanical parameter profile of the entire well section. At the depth point, the first The values ​​of the mechanical parameters are , Number the depth points of the profile. , This represents the total number of points at the profile depth. Using the mechanical parameter index from step S1, .remember , The first The maximum and minimum values ​​of the mechanical parameters in the continuous mechanical parameter profile of the entire well section are listed. Considering that a larger Young's modulus and lower compressive strength and fracture toughness are more conducive to fracturing, the continuous mechanical parameter profile of the entire well section is normalized so that a larger normalized value indicates greater fracturing potential. The Young's modulus is normalized using a positive index.

[0073]

[0074] Normalization of compressive strength using negative index

[0075]

[0076] Fracture toughness normalized using a negative index

[0077]

[0078] Indicates the first Mechanical parameters in the first Normalized values ​​at each profile depth point. It should be noted that if a certain type of mechanical parameter is constant along the entire well section, i.e. If the parameter does not vary along the well depth, the normalized value is... All values ​​are set to 1 to avoid division by zero and to ensure that the parameter is treated equally in subsequent compressibility indices.

[0079] After normalization, the subjective weights of each mechanical parameter are calculated using the analytic hierarchy process (AHP) to incorporate prior knowledge from experts in fracturing engineering practice. Pairwise importance comparisons are performed on compressive strength, Young's modulus, and fracture toughness, and a judgment matrix is ​​constructed using a 1-9 scaling method. , It is a 3×3 positive reciprocal matrix, that is , And diagonal elements , Indicates the first Class parameters relative to the first The importance of class parameters, This embodiment uses a judgment matrix based on practical experience in continental shale fracturing engineering. for

[0080]

[0081] Solve the judgment matrix The eigenvector corresponding to the largest eigenvalue is normalized according to the L1 norm to obtain the eigenvalue. Subjective weights of mechanical parameters , The subjective weight obtained in this embodiment is... , , These correspond to compressive strength, Young's modulus, and fracture toughness, respectively.

[0082] Then, the entropy method is used to calculate the initial weights of each mechanical parameter, in order to introduce an objective correction for subjective biases based on the inherent dispersion of the full-well section profile data. This is based on the normalized values ​​of the continuous mechanical parameter profile of the entire well section. , for the Normalized probability of mechanical parameters by depth

[0083]

[0084] Indicates the first Mechanical parameters in the first The proportion at each depth point in the profile satisfies The first probability is calculated from the normalized probability. Entropy value of mechanical parameters

[0085]

[0086] Agreement The entropy value is used to obtain the first... Initial weights of mechanical parameters

[0087]

[0088] The greater the data dispersion, the smaller the entropy value, and the larger the initial weight. In this embodiment, the initial weight calculated based on the actual data from well X1 is: , , The specific value of the initial weight depends on the actual data distribution of the target well; different initial weights can be obtained for different well sections or different wells.

[0089] After multiplying the subjective weights and the initial weights according to the corresponding mechanical parameters and normalizing, we obtain the first... Correction weights for mechanical parameters

[0090]

[0091] The modified weights reflect both expert prior knowledge and the inherent dispersion of the data, achieving an organic combination of subjective and objective perspectives. The modified weights obtained in this embodiment are: , , In the first At each profile depth point, the compressibility index is calculated based on the above normalized values ​​and corrected weights.

[0092]

[0093] The compressibility index ranges from [value range missing]. The larger the value, the easier it is for the reservoir at that depth to be fractured and stimulated.

[0094] The compressibility of the reservoir at each profile depth point is classified according to the following criteria: It was determined to be a Class I reservoir, with optimal compressibility. It was initially determined to be a Class II reservoir with good compressibility; It was initially determined to be a Class III reservoir with moderate compressibility. The reservoir was initially classified as Class IV, indicating poor compressibility. Based on this, the compressibility classification results for the entire reservoir section of the target well along the well depth are output, providing a quantitative basis for the optimal selection of fracturing stimulation zones. This step combines the experience of fracturing engineering experts with the discrete characteristics of the data to obtain objective weights, and then calculates the compressibility index through weighted summation, achieving compressibility classification and providing a quantitative decision-making basis for the optimal selection of fracturing stimulation zones.

[0095] The above description is merely a preferred embodiment of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the compressibility of continental shale based on conditional generative adversarial networks, characterized in that... Including steps: S1. Obtain mechanical parameter samples of core samples at several discrete depth points in the target well. The mechanical parameter samples include compressive strength, Young's modulus, and fracture toughness. S2. Perform Savitzky-Golay filtering on the continuous logging curves of the entire target well to extract the low-frequency trend term, and construct conditional variables using the well depth at the discrete depth points and the low-frequency trend term as components; S3. Construct a conditional generative adversarial network including a generator and a discriminator; under the constraints of the conditional variables, the generator generates a mechanical parameter vector based on a random noise vector, and the discriminator outputs the probability that the input sample is a real sample; S4. Using a joint loss function consisting of adversarial loss, reconstruction loss, and gradient penalty, the generator and the discriminator are trained adversarially to obtain a trained generator; S5. Input the conditional variable of the depth point to be predicted and the random noise vector into the trained generator, and merge the obtained mechanical parameter prediction values ​​with the mechanical parameter samples in S1 according to depth to form a continuous mechanical parameter profile of the entire well section; S6. Normalize the continuous mechanical parameter profile of the entire well section, and use the combined weighting method of analytic hierarchy process and entropy method to obtain the corrected weights of each mechanical parameter. The compressibility index is obtained by weighted summation according to the corrected weights, and the reservoir is classified according to the compressibility index.

2. The method according to claim 1, characterized in that... In step S2, the Savitzky-Golay filter selects an odd-length sliding window near each sampling point of the continuous logging curve throughout the well section, constructs a local polynomial within the sliding window, solves the coefficients of the local polynomial based on the least squares criterion, and uses the fitted value of the local polynomial at the center of the sliding window as the smoothed output of that sampling point; the components of the condition variable include the well depth of the discrete depth point and the low-frequency trend term components corresponding to the compressive strength, the Young's modulus, and the fracture toughness.

3. The method according to claim 1, characterized in that... In step S3, the generator is a multilayer fully connected neural network. The generator takes the concatenated vector of the random noise vector and the condition variable as input. The generator also includes a skip connection branch, which directly and linearly maps the concatenated vector to the output of the generator. The final output of the generator is obtained by adding the output of the main branch of the multilayer fully connected neural network to the output of the skip connection branch. The discriminator is a multilayer fully connected neural network. The discriminator takes the concatenated vector of the mechanical parameter vector and the condition variable as input. The output layer of the discriminator nonlinearly maps the output features of its hidden layer to the open interval (0,1).

4. The method according to claim 1, characterized in that... In step S4, the adversarial loss is used to measure the degree to which the sample output by the generator is judged as a real sample by the discriminator; the reconstruction loss is the norm between the sample output by the generator and the mechanical parameter sample in S1; the gradient penalty is based on the interpolated sample between the mechanical parameter sample in S1 and the sample output by the generator, and is obtained by multiplying the expected square of the difference between the L2 norm of the gradient output by the discriminator on the interpolated sample and 1 by the gradient penalty coefficient; the alternating adversarial training first fixes the parameters of the generator and updates the parameters of the discriminator in each iteration, and then fixes the parameters of the discriminator and updates the parameters of the generator, and a label smoothing strategy is adopted in the alternating adversarial training process.

5. The method according to claim 1, characterized in that... In step S6, the Young's modulus is normalized using a positive index, and the compressive strength and fracture toughness are normalized using a negative index. The combined weighting of the analytic hierarchy process (AHP) and entropy method includes: constructing a judgment matrix by comparing the pairwise importance of the compressive strength, Young's modulus, and fracture toughness; obtaining the subjective weights of each mechanical parameter from the judgment matrix using the AHP; obtaining the initial weights of each mechanical parameter from the normalized values ​​of the continuous mechanical parameter profile of the entire well section using the entropy method; and normalizing the subjective weights and the initial weights according to the corresponding mechanical parameters to obtain the corrected weights. The compressibility index is the weighted sum of the normalized mechanical parameters and the corrected weights.

Images