A well testing and microseismic fusion fracturing fracture network parameter inversion method

Abstract

This invention relates to the field of unconventional oil and gas reservoir development and artificial intelligence inversion technology. It discloses a method for inverting fracture network parameters by integrating well testing and microseismic data, addressing the problems of high computational cost and severe non-uniqueness in traditional numerical well testing inversion. This method constructs a well testing forward model and inversion sample library, completes data preprocessing, builds a training attention-enhanced operator learning surrogate model, and efficiently solves for pressure transient responses. It inverts global reservoir seepage parameters through multi-starting-point gradient descent, and constructs a probabilistic inversion model combined with microseismic constraints to accurately determine the half-length of segmented fractures. It relies on latent variable neighborhood search to screen effective solutions, quantifies parameter uncertainty, and outputs the results. This invention integrates well testing dynamics and microseismic monitoring data, improving inversion accuracy and rationality, enabling rapid well testing interpretation and parameter probabilistic evaluation, and is suitable for field development applications in shale oil fractured horizontal wells.

Description

Technical Field

[0001] This invention belongs to the field of unconventional oil and gas reservoir development and artificial intelligence inversion technology, and relates to a method for inverting fracture network parameters by integrating well testing and microseismic analysis. Background Technology

[0002] Shale oil reservoirs typically require multi-stage fracturing of horizontal wells to create complex fracture networks, thereby expanding the stimulated volume and increasing single-well productivity. Parameters such as fracture half-length, fracture conductivity, matrix permeability, microfractures, and skin factor directly influence well test curve morphology, production dynamics interpretation, and subsequent well network optimization. Therefore, accurately, rapidly, and with uncertainty characterization capabilities, inverting fracture parameters from multi-stage fracturing horizontal wells is a key technical issue in shale oil reservoir development and evaluation. However, existing well test inversion methods for multi-stage fracturing horizontal wells in shale oil still have the following shortcomings: Multi-stage fracturing horizontal wells typically require explicit description of the multi-scale coupled flow between artificial fractures, natural fractures, and the matrix. While high-precision numerical simulations based on embedded discrete fracture models or unstructured meshes can characterize the fracture network well, a single forward modeling calculation is time-consuming and cannot support inversion, history fitting, and uncertainty assessment that require a large number of forward modeling calls.

[0003] Bottom hole pressure response is essentially a one-dimensional dynamic observation, while parameters such as fracture half-length, fracture conductivity, matrix permeability, and microfractures to be inverted have higher dimensions. Different combinations of parameters may produce similar pressure and pressure derivative curves, which makes it easy for traditional deterministic optimization methods to converge to a single local optimum, making it difficult to reveal the equivalent solution space and parameter compensation relationship.

[0004] While conventional fully connected networks or ordinary multilayer perceptron structures can approximate pressure curves, they are prone to spectral bias at the transition stages of flow, such as early wellbore storage, bilinear flow, linear flow, and boundary control flow. This leads to insufficient prediction of pressure derivative inflection points and curve details, thus affecting inversion stability.

[0005] Microseismic monitoring can provide indirect information on the extent of fracturing and the spatial distribution of fractures. However, microseismic events reflect rock fracturing or shearing activity and are not equivalent to effective conductive fractures. Existing methods often treat microseismic point clouds as static displays or modeling boundaries, failing to introduce them as soft physical constraints into the probabilistic inversion latent variable space. This results in inconsistent fracture geometric inversion with the spatial distribution of microseismic events.

[0006] Field test data inevitably contains pressure noise, derivative calculation errors, and model structure errors. If only a single optimal parameter combination is output, the interpretation results are easily too certain, making it difficult to provide a risk range for fracturing effect evaluation and reservoir development decision-making.

[0007] Therefore, there is an urgent need to propose a probabilistic inversion method for multi-stage fracturing horizontal well testing in shale oil that can integrate fast operator learning forward modeling proxy, well test dynamic response, and microseismic spatial proxy constraints, in order to reduce computational costs, alleviate inversion non-uniqueness, and output uncertainty interpretation results with physical consistency. Summary of the Invention

[0008] This invention provides a method for inverting fracture network parameters by integrating well testing and microseismic analysis, in order to solve the problems of high computational cost and serious non-uniqueness of well testing inversion in the prior art.

[0009] A method for inverting fracture network parameters by integrating well testing and microseismic data includes the following steps: S1. Establish a forward model for multi-stage fracturing horizontal well testing in shale oil and construct a sample library for probabilistic inversion of well testing results; S2. Preprocess the multi-source data in the well test probability inversion sample library to obtain the normalized parameter vector, time coordinate vector, pressure transient characteristic vector and microseismic envelope proxy vector; S3. Construct an attention-enhanced operator learning forward surrogate model, train and validate the attention-enhanced operator learning forward surrogate model based on the well test probability inversion sample library, and obtain the validated well test forward surrogate model; S4. Based on the validated well test forward modeling proxy model, the global seepage parameters of multi-stage fractured horizontal wells in shale oil are inverted using the multi-starting point gradient descent method to obtain the global parameter inversion results; S5. Construct a conditional variational autoencoder probabilistic inversion model with microseismic envelope constraints, search in the latent variable space, combine pressure matching error, microseismic correlation penalty term and latent variable prior constraints, invert the fracture half length of each fracturing segment, and obtain the segmented fracture half length inversion results. S6. Search the equivalent solution space near the optimal latent variable, screen candidate inversion solutions that meet the preset pressure matching tolerance, statistically obtain the uncertainty interval of global seepage parameters and segmented fracture half-length parameters, and output the probabilistic inversion results of shale oil multi-stage fracturing horizontal well testing.

[0010] Furthermore, S1 includes S1.1, establishing a dual-medium reservoir model for a multi-stage fractured horizontal well in shale oil, describing the matrix system, natural fracture system, and artificial fracture system as flowing media respectively, and setting the horizontal wellbore, fractured section, and half-length of each fracture section; S1.2, the matrix-crack conductivity is calculated based on the analytically modified embedded discrete crack model, and the transient correction factor, the initiation pressure gradient and the stress sensitivity coefficient are introduced to correct the early pressure transient response and the crack-matrix exchange intensity, respectively, to construct the pressure transient forward model; S1.3, using the Latin hypercube sampling method to generate multiple physically valid parameter combinations within a preset parameter range, the parameter combinations including reservoir static parameters, global flow parameters and segmented fracture half-length parameters; The reservoir static parameters include matrix porosity. Effective reservoir thickness Overall compression ratio Fluid viscosity and the number of effective fracturing stages ; The global seepage parameters include matrix permeability. , fracture conductivity Mechanical skin coefficient and microcracks ; The segmented fracture half-length parameter includes the hydraulic fracture half-length of each effective fracturing segment. And it is represented by padding using a fixed-length vector; S1.4 Perform pressure transient simulation for each parameter combination to obtain the bottom hole pressure response and pressure derivative curves. Construct a well test probability inversion sample library from multiple sets of input parameter combinations and corresponding simulation response data.

[0011] Furthermore, S2 includes S2.1, which performs maximum-minimum normalization or logarithmic normalization on the input parameters; S2.2, regarding test time Perform a logarithmic scaling transformation to map the production test time into a time coordinate vector; S2.3, Calculate the pressure change based on the pressure response. and pressure derivative ; S2.4 Obtain the measured microseismic events during fracturing operations. Based on the event density, magnitude-weighted energy, or event envelope width along the horizontal wellbore direction, generate a microseismic envelope proxy vector corresponding to the fracturing segment number. And the microseismic envelope proxy vector Normalization process.

[0012] Furthermore, the attention-enhanced operator learning forward surrogate model includes a parameter branch network, a temporal backbone network, a multi-head self-attention module, a residual connection module, and a stress response output module; The parameter branch network is used to receive a high-dimensional parameter vector composed of reservoir static parameters, global flow parameters, and segmented fracture half-length parameters. The time backbone network is used to receive production test time coordinates; The multi-head self-attention module is set in the parameter branch network and is used to dynamically assign weights to the reservoir static parameters, global seepage parameters and segmented fracture half-length parameters in the high-dimensional parameter vector. The residual connection module is used to fuse the attention-enhanced features with the original embedded features to reduce the spectral shift generated by the multilayer perceptron structure at the transition between pressure transient and multi-flow stages. The pressure response output module obtains the predicted pressure response value at the corresponding time coordinate by performing an inner product operation between the output features of the parameter branch network and the output features of the time backbone network. The output of the attention-enhancing operator learning the forward surrogate model is represented as follows:

[0013] ; In the formula, For the forward surrogate model in time Predicted pressure response at the location, For the parameter branch network, the first Each output feature For the time backbone network Each output feature For the basis function dimension, This is a bias term.

[0014] Furthermore, when training and validating the attention-enhancing operator to learn the forward surrogate model, the stress response error and the stress derivative error are both incorporated into the training objective. A weighted loss function is obtained by coupling the stress response error and the stress derivative error. The weighted loss function is as follows: ; In the formula, To learn the forward surrogate model loss function for the operator, The number of time sampling points, For the first The weight of each time sampling point This represents the actual change in pressure. To predict pressure changes, For the true pressure derivative, To predict the pressure derivative, The pressure derivative constraint weighting coefficient.

[0015] Furthermore, the multi-starting-point gradient descent process specifically includes: Fix or initialize the segmented fracture half-length parameters, and set the fracture half-length of each fracturing segment to a uniform prior value or empirical initial value. Multiple initial points are generated within the global seepage parameter range, and each initial point is input into the validated well test forward modeling proxy model. The goal is to minimize the error between the predicted and observed pressure responses, and to study the matrix permeability. , fracture conductivity Mechanical skin coefficient and microcracks Perform iterative optimization; By comparing the optimization results corresponding to multiple initial points, the parameter combination with the smallest pressure matching error is selected as the global parameter inversion result.

[0016] Furthermore, the conditional variational autoencoder probabilistic inversion model with microseismic envelope constraints includes an encoder, a latent variable sampling layer, and a decoder; The encoder is used to learn the latent variable distribution of the segmented crack half-length vector under conditional features; The latent variable sampling layer is used to obtain latent variables through reparameterized sampling. ; The decoder is used to determine the latent variables. Generate segmented crack half-length vectors using conditional features. .

[0017] Furthermore, the objective function for the latent variable space search is set as follows: ; In the formula, as latent variables The corresponding inversion objective function, For pressure matching error, For microseismic constraint weighting coefficients, For the preset correlation threshold and The value range is 0.6 to 0.9. The half-length vector of the segmented crack With microseismic envelope proxy vector The Pearson correlation coefficient between them, with a value ranging from -1 to 1. For the prior regularization coefficients of latent variables; When the correlation coefficient between the segmented crack half-length vector and the microseismic envelope surrogate vector is less than the preset correlation threshold, a penalty is applied to the latent variable candidate solution so that the generated segmented crack half-length distribution simultaneously satisfies the pressure dynamic response constraint and the microseismic spatial distribution constraint. When the correlation coefficient is not less than the preset correlation threshold, the candidate segmented crack half-length vector satisfies the microseismic spatial distribution constraint, the microseismic correlation penalty term is zero, and no longer applies microseismic penalty to the latent variable candidate solution. The objective function is calculated based on the pressure matching error and the latent variable prior regularization.

[0018] Furthermore, the pressure matching error The candidate fracture half-length vector obtained by decoding the conditional variational autoencoder and the global parameter inversion result are combined and input into the well test forward model to obtain the predicted pressure response. The mean square error or relative error between the predicted pressure response and the observed pressure response is then calculated. The microseismic correlation penalty term is used to constrain a positive correlation between the segmented crack half-length vector and the microseismic envelope surrogate vector. When the correlation coefficient is lower than a preset correlation threshold... When the correlation coefficient is not lower than the preset correlation threshold, a secondary penalty is applied to the candidate latent variables; At that time, the microseismic correlation penalty is not applied or is weakened.

[0019] Furthermore, the equivalent solution space search in the vicinity of the optimal latent variable includes generating multiple candidate latent variables in the neighborhood of the optimal latent variable using the Latin hypercube sampling method, decoding, forward prediction and pressure error calculation for each candidate latent variable, and retaining the candidate solution with pressure matching error not greater than the preset pressure matching tolerance as the equivalent inversion solution. The preset pressure matching tolerance is determined based on the accuracy of the field pressure gauge, the noise of derivative calculation, and the model error. After statistical analysis of the retained equivalent inversion solutions, the empirical quantiles of P10, P50, and P90 of the global seepage parameters and the segmented fracture half-length parameters are calculated respectively, and parameter compensation relationship diagrams, fracture half-length probability distribution diagrams, and well test probability inversion interpretation results are generated.

[0020] Compared with the prior art, the present invention has the following technical effects: This invention proposes a probabilistic inversion method for multi-stage fracturing horizontal well testing in shale oil based on operator learning and microseismic surrogate constraints. It uses an attention-enhanced operator learning forward model to replace the traditional analytical correction embedded discrete fracture numerical simulator. While maintaining the accuracy of pressure response and pressure derivative prediction, it significantly reduces the time consumption of a single forward modeling calculation, thus providing an efficient computational basis for large-scale inversion search and uncertainty evaluation.

[0021] This invention introduces a multi-head self-attention mechanism and residual connection into the forward surrogate model, enabling the model to dynamically weight reservoir parameters, global flow parameters, and segmented fracture half-length parameters in the high-dimensional parameter vector. This enhances the ability to characterize multi-flow stage transitions, early pressure derivative abrupt changes, and late boundary responses, avoiding the inadequacy of ordinary multilayer perceptron structures in complex pressure transient curves.

[0022] This invention employs a decoupled inversion strategy that proceeds from global to local. First, global parameters such as matrix permeability, fracture conductivity, mechanical skin coefficient, and microcracks are stabilized and inverted through multi-starting-point gradient descent. Then, high-dimensional piecewise fracture half-length is inverted through a conditional variational autoencoder, which reduces the search difficulty and parameter compensation risk of high-dimensional joint inversion.

[0023] The spatial distribution information of microseismic events is converted into a microseismic envelope surrogate vector and introduced as a soft physical constraint into the latent variable search objective function of the conditional variational autoencoder. This ensures that the fracture half-length distribution obtained by inversion not only satisfies the dynamic response of bottom hole pressure, but also remains consistent with the modification range reflected by the distribution of microseismic events, thereby alleviating the geometric non-uniqueness of fractures caused by relying solely on pressure data.

[0024] By searching the equivalent solution space and outputting the empirical quantiles P10, P50, and P90, the parameter compensation relationship and fracture half-length uncertainty can be expressed in a probabilistic sense. Compared with the single deterministic inversion result, it is more conducive to fracturing effect evaluation, well pattern adjustment, production prediction, and development scheme risk decision-making. Attached Figure Description

[0025] Figure 1 This is a flowchart of the fracturing network parameter inversion method that integrates well testing and microseismic analysis according to the present invention; Figure 2 This is a flowchart of the three-stage inversion process of the present invention, which includes global parameter inversion, microseismic constrained crack half-length inversion, and equivalent solution space search. Figure 3 This is a schematic diagram of the probabilistic inversion model of the conditional variational autoencoder with microseismic envelope constraint of the present invention; Figure 4 This is a schematic diagram of the equivalent solution space search and the output of uncertainties P10, P50, and P90 in this invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, those skilled in the art can make appropriate adjustments and implementations, and these also fall within the scope of protection of this invention.

[0027] This embodiment discloses a method for inverting fracture network parameters by integrating well testing and microseismic data. Figure 1 As shown, it includes the following steps: Establish a forward model for multi-stage fracturing horizontal well testing in shale oil, and construct a sample library for probabilistic inversion of well testing.

[0028] In this embodiment, the reservoir simulation area is set to 5000m × 3000m, horizontal wells are arranged along the long axis of the reservoir, the outer boundary of the reservoir is set as a closed boundary, and the bottomhole production regime is set as a fixed production test condition. The reservoir is represented as a dual-medium system, with crossflow exchange between the matrix and natural fracture systems. Artificial fracturing fractures are embedded in the matrix grid, and the matrix-fracture conductivity is calculated by analytically correcting the embedded discrete fracture model.

[0029] Spatial discretization was performed using the finite volume method, and temporal discretization was performed using the backward Euler method. The nonlinear terms caused by the initiation pressure gradient and stress sensitivity were solved using the Newton-Raphson iterative method. The well test forward model simultaneously considered the wellbore reservoir effect, mechanical skin effect, matrix pore stress sensitivity effect, and initiation pressure gradient effect of low-permeability shale reservoirs.

[0030] In this embodiment, the Latin hypercube sampling method is used to generate 100,000 sets of physically valid parameter combinations. Each parameter combination includes 5 reservoir static parameters, 4 global flow parameters, and up to 50 segmented fracture half-length parameters. Therefore, the main input dimension of the high-dimensional parameter vector is 59 dimensions. Among them, the 5 reservoir static parameters are matrix porosity, etc. Effective reservoir thickness Overall compression ratio Fluid viscosity and the number of effective fracturing stages The four global seepage parameters are matrix permeability. , fracture conductivity Mechanical skin coefficient and microcracks ; The segmented fracture half-length parameter includes the hydraulic fracture half-length of each effective fracturing segment. When the actual number of effective fracturing segments is less than 50, the ineffective fracturing segments are padded with zeros or masked.

[0031] For each set of parameter combinations, the analytical correction embedded discrete fracture well test forward model is run to obtain the corresponding bottom hole pressure response curve and logarithmic pressure derivative curve. The parameter combinations, pressure response curves, pressure derivative curves, and corresponding microseismic envelope proxy vectors are saved together to construct a well test probability inversion sample library.

[0032] The main parameter types of the well test probability inversion sample library described in this embodiment are shown in Table 1.

[0033] Table 1. Main parameter types of the well test probability inversion sample library .

[0034] The multi-source data in the well test probabilistic inversion sample library are preprocessed. In this embodiment, the reservoir static parameters, global seepage parameters, and segmented fracture half-length parameters are first normalized; logarithmic normalization is used for parameters that vary across orders of magnitude, and maximum-minimum normalization is used for parameters with limited value ranges. Subsequently, logarithmic sampling is performed on the production test time t, and each time point is used as the input coordinates of the time backbone network.

[0035] Calculate the pressure change based on the pressure response curve. and pressure derivative The pressure derivative is used to enhance the model's ability to identify flow stage transitions, fracture-matrix crossflow characteristics, and boundary response moments. For microseismic data, the event density, event energy, or envelope width near each fracturing segment are statistically analyzed along the horizontal wellbore direction, and microseismic envelope surrogate vectors corresponding one-to-one with the fracturing segment numbers are generated. This microseismic envelope surrogate vector does not directly represent the effective conduction fracture length, but rather serves as a soft physical constraint reflecting the range of hydraulic fracturing activity.

[0036] An attention-enhanced operator learning forward surrogate model is constructed. This model comprises a parameter branch network, a temporal backbone network, a multi-head self-attention module, a residual connection module, and a pressure response output module. The parameter branch network receives a high-dimensional parameter vector composed of reservoir static parameters, global flow parameters, and segmented fracture half-length parameters. Time backbone network reception test time coordinates The pressure response output module generates the predicted pressure response value at the corresponding time point by performing an inner product operation between the output features of the branch network and the output features of the backbone network.

[0037] To improve the model's ability to identify the coupling relationships between high-dimensional input parameters, a four-head self-attention module is incorporated into the parameter branch network. This module maps the input parameter vector to embedded features, calculates the query matrix, key matrix, and value matrix, and automatically learns the contribution relationships of matrix permeability, fracture conductivity, microcracks, and the half-length of each fracture segment to the pressure transient response through attention weights. Furthermore, residual connections preserve the original embedded features, enabling the model to enhance the expression of key parameters while avoiding instability during deep network training.

[0038] Training and validation of attention-enhanced operator-learning forward surrogate models. In this embodiment, multiple candidate forward surrogate models are trained using 100,000 sets of well test forward modeling samples, including fully connected networks, residual fully connected networks, standard operator learning models, attention-enhanced operator learning models, and Fourier residual operator learning models. All models use the same training and validation sets and are trained using a weighted loss function coupling pressure response error and pressure derivative error.

[0039] Comparative validation showed that the attention-enhanced operator-learning forward surrogate model achieved the lowest validation weighted mean square error (MSE) of 0.00300 with 140,549 trainable parameters, outperforming the standard operator-learning model and other neural network surrogate models. Furthermore, the model's single forward inference time is approximately 10... -3 Compared to the explicitly analytically corrected embedded discrete crack numerical simulator, this achieves a computational speedup of more than three orders of magnitude. Therefore, this embodiment selects the trained attention-enhanced operator-learned forward surrogate model as the forward computation engine for subsequent probability inversion.

[0040] Figure 2 This is a flowchart illustrating the three-stage inversion process of the present invention: global parameter inversion, microseismic constrained crack half-length inversion, and equivalent solution space search. Figure 2 As shown, a multi-starting-point gradient descent method is used to invert global seepage parameters. In this embodiment, the half-length of each fracturing segment is temporarily set to a uniform prior value to reduce the interference of high-dimensional fracture geometric variables on the identification of global seepage parameters; then, based on the matrix permeability... , fracture conductivity Mechanical skin coefficient and microcracks Multiple initial points are generated within a preset range, and the forward surrogate model is learned using a validated attention enhancement operator to quickly calculate the pressure response corresponding to each initial point.

[0041] Using the minimum error between the observed and predicted pressure responses as the objective function, a gradient descent method is employed to iteratively update the global seepage parameters from multiple initial points. The result with the minimum pressure matching error is selected as the global parameter inversion result. This step first locks in the range of macroscopic seepage parameters controlled by the mid-to-late stage pressure responses, reducing parameter compensation in subsequent fracture half-length inversions.

[0042] In this embodiment, the matrix permeability was verified on 100 blind test samples. The mean absolute percentage error was 3.04%, and the crack conductivity was... Mechanical skin coefficient and microcracks The mean absolute percentage error is less than 0.15%. Under pressure noise perturbation of 0% to 10%, the global parameter inversion results remain stable, indicating that the multi-starting point gradient descent method coupled with the attention-enhanced operator learning forward surrogate model has good noise resistance.

[0043] A conditional variational autoencoder probabilistic inversion model with microseismic envelope constraints is constructed to invert the half-length of segmented fractures. For example... Figure 3As shown, the conditional variational autoencoder probabilistic inversion model includes an encoder, a latent variable sampling layer, and a decoder. The encoder maps the piecewise crack half-length vectors in the training samples together with the conditional features into the latent variable mean and variance; the latent variable sampling layer generates latent variables through a reparameterization method. The decoder is based on latent variables. Generate segmented crack half-length vectors using conditional features. .

[0044] During the model training phase, the conditional variational autoencoder learns the spatial correlation structure of the segmented fracture half-lengths in the training samples, enabling the decoder to generate geologically plausible fracture half-length profiles from low-dimensional latent variables. In the inversion phase, the global parameter inversion results serve as macroscopic constraints, incorporating latent variables... As variables to be optimized, the pressure matching error, microseismic correlation penalty term, and latent variable prior regularization term are combined for optimization.

[0045] In this embodiment, the microseismic correlation threshold The value is set to 0.8. When the Pearson correlation coefficient between the decoded segmented fracture half-length vector and the microseismic envelope surrogate vector is below 0.8, a secondary penalty is applied to the latent variable candidate solution; when the correlation coefficient reaches or exceeds 0.8, the candidate solution is considered to satisfy the microseismic spatial distribution constraint. Through this soft constraint, the model avoids geometrically unreasonable fracture half-length combinations caused by simply relying on pressure matching.

[0046] In this embodiment, the segmented fracture half-length inversion error was statistically analyzed on 100 blind test samples. The mean absolute percentage error of the fracture half-length at the sample level was 10.67%, and the median was 10.74%. In the microseismic constrained ablation verification, after introducing a microseismic constraint with a correlation threshold of 0.8, the normalized fracture half-length error was reduced by 16.6% compared to the case without microseismic constraints, indicating that the microseismic envelope proxy constraint can effectively improve the fracture geometric inversion results.

[0047] Perform an equivalent solution space search and output the well test probability inversion results. For example... Figure 4 As shown, in this embodiment, Latin hypercube sampling is used to generate 500 sets of candidate latent variables near the optimal latent variable. For each set of candidate latent variables, the conditional variational autoencoder decoder is first input to obtain the candidate segmented crack half-length vector, and then combined with the global parameter inversion result to input the attention enhancement operator to learn the forward surrogate model, and calculate the corresponding pressure response curve.

[0048] The relative error between the predicted pressure response and the observed pressure response of the candidate solution is used as the screening criterion. When the relative error is no greater than a preset pressure matching tolerance of 10%, the corresponding candidate solution is retained as the equivalent inversion solution. The 10% pressure matching tolerance is adjusted according to the field pressure noise, derivative processing error, and interpretation requirements.

[0049] Statistical analysis was performed on the retained equivalent inversion solutions, and the matrix permeability was calculated separately. , fracture conductivity Mechanical skin coefficient Microcracks and the half length of the fracture in each fracturing section The P10, P50, and P90 empirical quantiles are calculated, and the parameter compensation relationship diagram and fracture half-length confidence interval are output. P10, P50, and P90 represent the low quantile, median, and high quantile interpretation results in the equivalent solution set, respectively, and are used to describe the uncertainty range of the well test inversion results.

[0050] In this embodiment, equivalent solution space analysis shows that the matrix permeability With fracture conductivity There is a clear positive correlation and compensation relationship between them, while the mechanical skin coefficient and microcracks The feasible range is relatively narrow. This result demonstrates that the present invention can not only output a single optimal inversion result, but also reveal the compensation direction and credible interpretation range between parameters, providing a basis for fracturing effect evaluation, development adjustment, and risk decision-making.

[0051] In summary, the probabilistic inversion method for multi-stage fracturing horizontal wells in shale oil proposed in this invention, based on operator learning and microseismic proxy constraints, achieves rapid, stable, and uncertainty-interpretable inversion of fracture parameters in multi-stage fracturing horizontal wells in shale oil through efficient forward modeling proxy, global parameter decoupling inversion, microseismic constraint fracture half-length probabilistic inversion, and equivalent solution space statistical output.

[0052] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for inverting fracture network parameters by integrating well testing and microseismic analysis, characterized in that, Includes the following steps: S1. Establish a forward model for multi-stage fracturing horizontal well testing in shale oil and construct a sample library for probabilistic inversion of well testing results; S2. Preprocess the multi-source data in the well test probability inversion sample library to obtain the normalized parameter vector, time coordinate vector, pressure transient characteristic vector and microseismic envelope proxy vector; S3. Construct an attention-enhanced operator learning forward surrogate model, train and validate the attention-enhanced operator learning forward surrogate model based on the well test probability inversion sample library, and obtain the validated well test forward surrogate model; S4. Based on the validated well test forward modeling proxy model, the global seepage parameters of multi-stage fractured horizontal wells in shale oil are inverted using the multi-starting point gradient descent method to obtain the global parameter inversion results; S5. Construct a conditional variational autoencoder probabilistic inversion model with microseismic envelope constraints, search in the latent variable space, combine pressure matching error, microseismic correlation penalty term and latent variable prior constraints, invert the fracture half length of each fracturing segment, and obtain the segmented fracture half length inversion results. S6. Search the equivalent solution space near the optimal latent variable, screen candidate inversion solutions that meet the preset pressure matching tolerance, statistically obtain the uncertainty interval of global seepage parameters and segmented fracture half-length parameters, and output the probabilistic inversion results of shale oil multi-stage fracturing horizontal well testing.

2. The method for inverting fracture network parameters by integrating well testing and microseismic analysis according to claim 1, characterized in that, S1 includes S1.1, establishing a dual-medium reservoir model for multi-stage fracturing horizontal wells of shale oil, describing the matrix system, natural fracture system and artificial fracturing fracture system as flowing media respectively, and setting the horizontal wellbore, fracturing section and half-length of each fracture section; S1.2, the matrix-crack conductivity is calculated based on the analytically modified embedded discrete crack model, and the transient correction factor, the initiation pressure gradient and the stress sensitivity coefficient are introduced to correct the early pressure transient response and the crack-matrix exchange intensity, respectively, to construct the pressure transient forward model; S1.3, using the Latin hypercube sampling method to generate multiple physically valid parameter combinations within a preset parameter range, the parameter combinations including reservoir static parameters, global flow parameters and segmented fracture half-length parameters; The reservoir static parameters include matrix porosity. Effective reservoir thickness Overall compression ratio Fluid viscosity and the number of effective fracturing stages ; The global seepage parameters include matrix permeability. , fracture conductivity Mechanical skin coefficient and microcracks ; The segmented fracture half-length parameter includes the hydraulic fracture half-length of each effective fracturing segment. And it is represented by padding using a fixed-length vector; S1.4 Perform pressure transient simulation for each parameter combination to obtain the bottom hole pressure response and pressure derivative curves. Construct a well test probability inversion sample library from multiple sets of input parameter combinations and corresponding simulation response data.

3. The method for inverting fracture network parameters by integrating well testing and microseismic analysis according to claim 1, characterized in that, S2 includes S2.1, which performs maximum-minimum normalization or logarithmic normalization on the input parameters; S2.2, regarding test time Perform a logarithmic scaling transformation to map the production test time into a time coordinate vector; S2.3, Calculate the pressure change based on the pressure response. and pressure derivative ; S2.4 Obtain the measured microseismic events during fracturing operations. Based on the event density, magnitude-weighted energy, or event envelope width along the horizontal wellbore direction, generate a microseismic envelope proxy vector corresponding to the fracturing segment number. And the microseismic envelope proxy vector Normalization process.

4. The method for inverting fracture network parameters by integrating well testing and microseismic analysis according to claim 1, characterized in that, The attention-enhanced operator learning forward surrogate model includes a parameter branch network, a temporal backbone network, a multi-head self-attention module, a residual connection module, and a stress response output module; The parameter branch network is used to receive a high-dimensional parameter vector composed of reservoir static parameters, global flow parameters, and segmented fracture half-length parameters. The time backbone network is used to receive production test time coordinates; The multi-head self-attention module is set in the parameter branch network and is used to dynamically assign weights to the reservoir static parameters, global seepage parameters and segmented fracture half-length parameters in the high-dimensional parameter vector. The residual connection module is used to fuse the attention-enhanced features with the original embedded features to reduce the spectral shift generated by the multilayer perceptron structure at the transition between pressure transient and multi-flow stages. The pressure response output module obtains the predicted pressure response value at the corresponding time coordinate by performing an inner product operation between the output features of the parameter branch network and the output features of the time backbone network. The output of the attention-enhancing operator learning the forward surrogate model is represented as follows: ； In the formula, For the forward surrogate model in time Predicted pressure response at the location, For the parameter branch network, the first Each output feature For the time backbone network Each output feature For the basis function dimension, This is a bias term.

5. The method for inverting fracture network parameters by integrating well testing and microseismic analysis according to claim 1, characterized in that, When training and validating the attention-enhanced operator to learn the forward surrogate model, the stress response error and stress derivative error are both incorporated into the training objective. A weighted loss function is obtained by coupling the stress response error and stress derivative error. The weighted loss function is as follows: ； In the formula, To learn the forward surrogate model loss function for the operator, The number of time sampling points, For the first The weight of each time sampling point This represents the actual change in pressure. To predict pressure changes, For the true pressure derivative, To predict the pressure derivative, The pressure derivative constraint weighting coefficient.

6. The method for inverting fracture network parameters by integrating well testing and microseismic analysis according to claim 1, characterized in that, The multi-starting-point gradient descent process is specifically as follows: Fix or initialize the segmented fracture half-length parameters, and set the fracture half-length of each fracturing segment to a uniform prior value or empirical initial value. Multiple initial points are generated within the global seepage parameter range, and each initial point is input into the validated well test forward modeling proxy model. The goal is to minimize the error between the predicted and observed pressure responses, and to study the matrix permeability. , fracture conductivity Mechanical skin coefficient and microcracks Perform iterative optimization; By comparing the optimization results corresponding to multiple initial points, the parameter combination with the smallest pressure matching error is selected as the global parameter inversion result.

7. The method for inverting fracture network parameters by integrating well testing and microseismic analysis according to claim 1, characterized in that, The conditional variational autoencoder probabilistic inversion model with microseismic envelope constraints includes an encoder, a latent variable sampling layer, and a decoder. The encoder is used to learn the latent variable distribution of the segmented crack half-length vector under conditional features; The latent variable sampling layer is used to obtain latent variables through reparameterized sampling. ; The decoder is used to determine the latent variables. Generate segmented crack half-length vectors using conditional features. .

8. The method for inverting fracture network parameters by integrating well testing and microseismic analysis according to claim 1, characterized in that, The objective function for the latent variable space search is set as follows: ； In the formula, as latent variables The corresponding inversion objective function, For pressure matching error, For microseismic constraint weighting coefficients, For the preset correlation threshold and The value range is 0.6 to 0.

9. The half-length vector of the segmented crack With microseismic envelope proxy vector The Pearson correlation coefficient between them, with a value ranging from -1 to 1. For the prior regularization coefficients of latent variables; When the correlation coefficient between the segmented crack half-length vector and the microseismic envelope surrogate vector is less than the preset correlation threshold, a penalty is applied to the latent variable candidate solution so that the generated segmented crack half-length distribution simultaneously satisfies the pressure dynamic response constraint and the microseismic spatial distribution constraint. When the correlation coefficient is not less than the preset correlation threshold, the candidate segmented crack half-length vector satisfies the microseismic spatial distribution constraint, the microseismic correlation penalty term is zero, and no longer applies microseismic penalty to the latent variable candidate solution. The objective function is calculated based on the pressure matching error and the latent variable prior regularization.

9. The method for inverting fracture network parameters by integrating well testing and microseismic analysis according to claim 1, characterized in that, The pressure matching error The candidate fracture half-length vector obtained by decoding the conditional variational autoencoder and the global parameter inversion result are combined and input into the well test forward model to obtain the predicted pressure response. The mean square error or relative error between the predicted pressure response and the observed pressure response is then calculated. The microseismic correlation penalty term is used to constrain a positive correlation between the segmented crack half-length vector and the microseismic envelope surrogate vector. When the correlation coefficient is lower than a preset correlation threshold... When the correlation coefficient is not lower than the preset correlation threshold, a secondary penalty is applied to the candidate latent variables; At that time, the microseismic correlation penalty is not applied or is weakened.

10. The method for inverting fracture network parameters by integrating well testing and microseismic analysis according to claim 1, characterized in that, The equivalent solution space search in the vicinity of the optimal latent variable includes generating multiple candidate latent variables in the neighborhood of the optimal latent variable using the Latin hypercube sampling method, decoding, forward prediction and pressure error calculation for each candidate latent variable, and retaining the candidate solution with pressure matching error not greater than the preset pressure matching tolerance as the equivalent inversion solution. The preset pressure matching tolerance is determined based on the accuracy of the field pressure gauge, the noise of derivative calculation, and the model error. After statistical analysis of the retained equivalent inversion solutions, the empirical quantiles of P10, P50, and P90 of the global seepage parameters and the segmented fracture half-length parameters are calculated respectively, and parameter compensation relationship diagrams, fracture half-length probability distribution diagrams, and well test probability inversion interpretation results are generated.

Images