A method for predicting fracture development intensity of a reservoir and quantifying uncertainty

Abstract

This invention discloses a method for predicting reservoir fracture development intensity and quantifying its uncertainty, comprising the following steps: S1: Establishing a fracture development intensity dataset based on well logging curves and preprocessing it; S2: Optimizing well logging response features using Automatic Correlation Determination (ARD) technology; S3: Constructing and training a Bayesian Neural Network (BNN) model based on fracture development intensity; S4: Calibrating the trained BNN model using a calibration set to perform ordinal regression uncertainty calibration for fracture development intensity, resulting in a C-BNN model for quantifying fracture development intensity uncertainty; S5: Inputting the well interval data to be predicted into the C-BNN model, outputting the expected value of fracture development intensity and a 95% confidence interval through the C-BNN model. This invention can achieve a reliable uncertainty quantification assessment of reservoir fracture development intensity with high accuracy and strict consistency with the true probability.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration and development technology, and in particular to a method for predicting reservoir fracture development intensity and quantifying uncertainty. Background Technology

[0002] In the field of oil and gas exploration and development, for complex heterogeneous reservoirs, natural fracture systems serve as crucial reservoir spaces and seepage channels, and their development directly determines the oil and gas enrichment patterns and single-well productivity. The heterogeneous distribution of natural fracture systems is a key factor contributing to differences in oil and gas enrichment levels and single-well productivity. Oil and gas in complex reservoirs are often enriched in fracture development zones, but due to the complex interference of the subsurface medium and the strong nonlinearity and ambiguity of well logging responses, the difficulty of fracture prediction and data fidelity vary significantly across different regions. Relying solely on traditional deterministic prediction values ​​often fails to assess the underlying geological risks. In production areas with concentrated complex reservoirs, if we can move beyond a single numerical prediction approach, simultaneously achieving high-precision prediction of fracture development intensity and quantifying the uncertainty of prediction results, and establishing the relationship between predicted fracture intensity and geological risk by eliminating false positives based on differences in confidence interval widths, this will undoubtedly benefit the reduction of exploration risks in complex reservoirs and the achievement of large-scale production.

[0003] Existing reservoir fracture evaluation technologies have mainly evolved from traditional empirical interpretation to modern intelligent prediction, and have the following two typical characteristics.

[0004] 1. Traditional interpretation methods are based on linear empirical models and are widely used in simple formations, but they are poorly adaptable to complex reservoirs. Traditional well logging interpretation usually relies on single or combined curves such as sonic logging, density, or resistivity, and identifies fractures by establishing empirical formulas or cross plots. For example, the difference in resistivity between shallow and deep sides is used to qualitatively determine the degree of fracture development, or sonic transit time formulas are used to calculate fracture porosity. These methods are well-established in relatively homogeneous sandstone and mudstone formations, but in highly heterogeneous reservoirs, due to the complex nonlinear relationship between well logging response and fracture parameters, linear empirical models are unable to accurately characterize the quantitative features of fracture development intensity.

[0005] 2. Intelligent prediction methods have improved nonlinear mapping capabilities, but generally lack quantitative evaluation of the reliability of prediction results. To overcome the limitations of traditional methods, researchers have begun to apply artificial intelligence technology to fracture prediction. Currently, studies have used algorithms such as depth kernel methods and K-nearest neighbors to make predictions by establishing a complex nonlinear mapping relationship between logging response and fracture intensity, significantly improving prediction accuracy. In addition, Bayesian neural networks (BNNs), as a probabilistic model, have also been introduced to attempt to capture the uncertainty of prediction. However, most existing intelligent predictions focus on improving the accuracy of point estimation, while neglecting the evaluation of the "credibility" of prediction results. Furthermore, the confidence intervals generated by standard Bayesian methods are often uncalibrated, leading to statistically unreliable prediction results, which seriously restricts the smooth progress of fracture selection and risk avoidance work in complex reservoirs.

[0006] Therefore, there is an urgent need to develop an integrated technical method for fracture intensity prediction and uncertainty quantification, in order to reveal the different intensity categories of fracture development in reservoirs, as well as the cognitive uncertainty characteristics of the distribution of various prediction results in three-dimensional space. Summary of the Invention

[0007] This invention aims to solve a technical problem in the oil and gas exploration and development stage: how to provide reliable uncertainty quantification results that have undergone rigorous statistical calibration while predicting fracture intensity in volcanic reservoirs using well logging data. The Calibrated Bayesian Neural Network (C-BNN) method proposed in this invention not only provides high-precision fracture intensity predictions but also outputs confidence intervals that are strictly consistent with the nominal confidence level. This helps geological experts identify high-risk areas and reduce the blind spots in exploration and development.

[0008] This invention is achieved using the following technical solution: a method for predicting reservoir fracture development intensity and quantifying uncertainty, comprising the following steps: S1: Establish a fracture development intensity dataset based on well logging curves and perform preprocessing; S2: Optimize logging response characteristics using Automatic Correlation Determination (ARD) technology; S3: Construct a Bayesian neural network (BNN) model based on crack development intensity and train it to establish a probabilistic network that treats weights as random variables, and approximate the posterior distribution through variational inference. S4: Use the calibration set to calibrate the crack development intensity order-preserving regression uncertainty of the trained BNN model to obtain the crack development intensity uncertainty quantification C-BNN model. S5: Input the data of the well section to be predicted into the C-BNN model, and output the expected value of fracture development intensity through the C-BNN model, and output the 95% confidence interval.

[0009] Furthermore, step S1 includes the following sub-steps: S11: Construct a high-confidence sample library; S12: Taking into account the sensitivity of logging response to lithology, physical properties and fractures, the input feature space is defined, and then the initial feature pool is constructed; S13: This only involves the partitioning of the dataset, including the training set, calibration set, and test set. S14: The data within the dataset is preprocessed using the Z-Score standardization method to eliminate dimensional differences.

[0010] Furthermore, the initial characteristic pool includes one or more of the following: natural gamma, natural potential, wellbore, deep lateral resistivity, shallow lateral resistivity, flushing zone resistivity, density, neutrons, and acoustic transit time.

[0011] Furthermore, step S2 specifically involves: establishing a Bayesian linear regression model, where each feature in the input feature vector... Assign a separate hyperparameter The second type of maximum likelihood method is used to optimize the hyperparameters. When the optimization converges, if If the value is less than the preset value, the feature is retained, and the selected preferred feature subset is stored in the training sample library.

[0012] Furthermore, the hyperparameters Control the corresponding weights The prior distribution variance is: ; The maximum log-marginal likelihood function of the second type of maximum likelihood method is: .

[0013] Furthermore, step S3 includes the following sub-steps: S31: Construct a fully connected Bayesian neural network using a variational Bayesian neural network architecture, where the weights and biases in the network architecture are random variables that follow a Gaussian distribution; S32: Define variational posterior and reparameterized sampling; S33: Construct a model based on maximizing the lower bound of evidence ELBO, perform iterative training on the BNN model until ELBO converges, and save the model parameters.

[0014] Furthermore, step S32 specifically involves: employing a variational inference method, introducing a parameterized variational distribution to approximate the true posterior, and using a reparameterization method to address the gradient non-differentiability problem during random sampling, representing the random weight w as a combination of deterministic parameters and noise. ; in, Let be the variational parameters to be learned. To help noise variables, This is for element-wise multiplication.

[0015] Furthermore, the iterative training specifically involves: performing Monte Carlo sampling using the Flipout estimator, calculating the predicted values ​​and distribution of the current batch of data, calculating the loss based on the loss function, calculating the gradient of the loss function with respect to the variational parameters using the backpropagation algorithm, updating the parameters using the Adam optimizer, repeating the above process until ELBO converges or the preset number of epochs is reached, and finally saving the final model parameters.

[0016] Furthermore, step S4 includes the following sub-steps: S41: Input the calibration set data into the trained BNN model and process each sample... Sub-Monte Carlo sampling is performed to obtain the position of the true label of each calibration sample in its predicted cumulative distribution function; S42: Construct a mapping relationship between predicted probabilities and empirical probabilities, and calculate the empirical coverage probability at different nominal confidence levels by statistically analyzing the value distribution of all samples in the calibration set. S43: Train an ordinal-preserving regression model to fit the nonlinear mapping from predicted probability to empirical probability, and integrate the trained ordinal-preserving regression model into the back end of the original BNN model to form the final C-BNN model.

[0017] Furthermore, the cumulative distribution function of the C-BNN model Defined as: ; in, The order-preserving regression calibration function is a piecewise constant, monotonically increasing step function; By using the cumulative distribution function, the nominal confidence interval of the original BNN output is corrected to a calibration confidence interval, ensuring that when the C-BNN outputs a 95% confidence interval, the probability of the true value falling into that interval strictly converges to 95%. The beneficial effects of this invention are: This invention addresses the bottlenecks of ambiguity in well logging responses and unknown reliability in fracture prediction of complex heterogeneous reservoirs by proposing an integrated technical solution that combines Bayesian sparse feature optimization, variational probability modeling, and non-parametric statistical calibration. The solution first utilizes ARD technology to automatically eliminate redundant features insensitive to fracture response by maximizing the marginal likelihood function, constructing a high signal-to-noise ratio feature space. Then, it trains a Bayesian neural network using variational inference and reparameterization techniques, obtaining the original prediction distribution containing cognitive uncertainty through ELBO. The core breakthrough lies in introducing a calibration mechanism based on ordinal-preserving regression, using independent calibration sets to calculate the true value quantiles and construct a mapping function to rigorously calibrate the original uncertainty interval. Finally, by simultaneously outputting the expected fracture intensity value and the calibrated confidence interval, and using NMPIW as a quantitative indicator, it achieves a leap from simple numerical prediction to "high-precision prediction plus reliable risk warning." Attached Figure Description

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

[0019] Figure 1 This is a flowchart of the present invention; Figure 2 Comparison of depth prediction maps and confidence intervals for three wells before and after calibration; Figure 3 The figure shows the validation results of the C-BNN model in blind well testing. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0022] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0023] See Figure 1 A method for predicting reservoir fracture development intensity and quantifying uncertainty includes the following steps: Establishment and Preprocessing of Fracture Development Intensity Dataset Based on Well Logging Curves Step 1: Constructing a high-confidence sample database. Based on five core wells with complete data quality in the study area, an original well logging database was constructed, extracting a total of 6,500 effective depth sampling points. Among them, the target label Y, which characterizes fracture development intensity by fine interpretation of fractures using electrical imaging logging (FMI), has a numerical distribution range of 0 to 1, covering the complete geological features from non-fractured zones to highly fractured zones.

[0024] Step 2: Define the input feature space. Taking into account the sensitivity of logging response to lithology, physical properties and fractures, nine curves are selected from the original logging series to construct the initial feature pool, which includes: natural gamma (GR), spontaneous potential (SP), caliper diameter (CALI), deep lateral resistivity (RT), shallow lateral resistivity (RI), flushed zone resistivity (RXO), density (DEN), neutron density (CNL) and sonic transit time (AC).

[0025] Step 3: Implement a strict dataset partitioning strategy. To ensure the objectivity of model evaluation and the effectiveness of subsequent uncertainty calibration, all sample data are randomly and mutually exclusively divided into three independent subsets: a training set (60%), a calibration set (20%), and a test set (20%). The calibration set is physically isolated from the training set; its data does not participate in the gradient descent and weight updates of the Bayesian neural network and is only used for training the subsequent ordinal-preserving regression calibration model to prevent overfitting during the calibration process.

[0026] Step 4: Perform leakage prevention standardization. To eliminate dimensional differences, Z-Score standardization is used to preprocess the data. Only the statistics and standard deviation of the training set are used to perform a uniform normalization operation on the training, calibration, and test sets. This ensures the authenticity and reliability of the model's evaluation results when facing unknown test data.

[0027] ARD-based feature optimization Sensitivity analysis of logging parameters was performed using Automatic Correlation Determination (ARD) technology. A Bayesian linear regression model was established, with each feature in the input feature vector as an example. Assign a separate hyperparameter (Precision parameter), this parameter controls the corresponding weight. Prior distribution variance: ; Optimizing hyperparameters using Type-II Maximum Likelihood Maximize the log-marginal likelihood function: ; When the optimization converges, if a certain feature corresponds to This indicates that its weight posterior distribution... Extremely shrinking to 0, this characteristic is irrelevant noise; if If the weight is small, the feature is retained. The selected subset of preferred features is stored in the training sample library. In this embodiment, the weight distributions of DEN, AC, RT, RI, CNL, GR, and CALI are divergent and contribute significantly to the prediction results, which is consistent with geological understanding. Conversely, the weights of SP and RXO are strongly compressed to a very small value. Therefore, this embodiment automatically removes SP and RXO, retaining only the remaining 7 features to construct the input vector.

[0028] Building a Bayesian Neural Network (BNN) Model The design comprises an input layer, three hidden layers (each with 30 nodes), ReLU activation, and an output layer for crack intensity. Unlike traditional neural networks with fixed weights, this invention assigns all weights in the network to... and bias Defined as a random variable following a Gaussian distribution. Define the prior distribution. It follows a standard normal distribution. Variational approximation setting: due to the true posterior Non-integrable, introducing parameterized variational distribution To approximate the true posterior. Table 1 shows the hyperparameter configuration of the neural network model: Table 1 Model Hyperparameter Configuration

[0029] Specifically, a fully connected Bayesian neural network is constructed using a variational Bayesian neural network architecture, with the following network structure parameters: Input layer: Receives 7 sensitive logging feature vectors selected through step S2. ; Hidden layers: Three hidden layers are set, each containing 30 neurons. All hidden layers use the ReLU activation function to introduce non-linear feature mapping. Output layer: Set as a probability output layer, outputting the predicted distribution parameters of crack development intensity, i.e., the mean. and variance ; Parameter definition: The weights w and biases b in the network are defined as random variables following a Gaussian distribution, rather than fixed values, and a prior distribution is set. It follows a standard normal distribution. .

[0030] Define variational posterior and reparameterized sampling Due to the true posterior distribution in Bayes' formula For non-integrable systems, a variational inference method is employed, introducing a parameterized variational distribution. To approximate the true posterior, a reparameterization method is used to address the gradient non-differentiability issue during random sampling. This method represents the random weights w as a combination of deterministic parameters and noise. ; in Let be the variational parameters to be learned; The auxiliary noise variable is sampled from a standard normal distribution. This indicates element-wise multiplication.

[0031] Model training loop based on maximizing the lower bound of evidence (ELBO) Training is performed using the training set data. Reparameterized sampling is employed, and auxiliary noise is introduced to address the non-differentiability issue of random sampling. The weights are represented as deterministic transformations. A Flipout estimator is used to reduce gradient variance. Training samples are input into the network, and the sampled weights are then used... Calculate the predicted value The negative evidence lower bound (-ELBO) is calculated as the loss function. It includes the fitting term: negative log-likelihood. Regularization term: KL divergence KL .

[0032] ; Calculate Loss with respect to variational parameters The gradient is calculated, and the parameters are updated using the Adam optimizer. This process is repeated until ELBO converges, and the model parameters are saved. ARD correlation analysis is performed in this step. After model training converges, the hyperparameters of each feature are extracted.

[0033] Specifically, the goal of training the ELBO loss function model is to make the variational distribution... As close as possible to the true posterior This is equivalent to maximizing the lower bound of the evidence, i.e., minimizing the following loss function. : ; Among them, the first item The divergence (Kullback-Leibler Divergence), as a regularization term, measures the difference between the variational distribution and the prior distribution, preventing the model from overfitting; the second term... The negative log-likelihood expectation, as a data fitting term, is used to measure the model's ability to interpret observed data through predicted distributions; Y is the true crack development intensity label in the training set; and X is the input feature in the training set.

[0034] Model iterative training and parameter updates utilize the training set data to iteratively train the model: Monte Carlo sampling is performed using the Flipout estimator to calculate the predicted values ​​and distribution of the current batch of data. According to the formula... Calculate the loss and use the backpropagation algorithm to compute the loss function with respect to variational parameters. The gradient is calculated. The Adam optimizer is used to update the parameters, with an initial learning rate of 0.001. This process is repeated until ELBO converges or the preset number of epochs is reached, at which point the final model parameters are saved. .

[0035] Calibrate Bayesian Neural Networks This step uses an independent calibration set to correct the confidence interval bias of the original BNN output, ensuring that the predicted probability is consistent with the true probability. Specifically, it includes the following sub-steps: The probability integral transform value of the calibration samples is calculated by inputting the calibration set data into the trained BNN model, and then performing the calculation on each sample. 50 Monte Carlo samplings are performed to obtain the predicted distribution. mean and variance Calculate the true label for each calibration sample. Position in its predicted cumulative distribution function : ; in, Let be the cumulative distribution function predicted by the model for the i-th calibration sample; This represents the actual crack strength observation value of the i-th calibration sample; This indicates the quantile where the true value falls within the predicted distribution; Constructing a mapping relationship between predicted probabilities and empirical probabilities for all samples in the statistical calibration set. Value distribution, calculating the empirical coverage probability at different nominal confidence levels. If the model is not calibrated, there will be a deviation between the predicted probability and the empirical probability. Constructing the dataset. ,in To predict probability quantiles, This corresponds to the actual observation frequency.

[0036] Training an ordinal-preserving regression model: Train an ordinal-preserving regression model R to fit the predicted probability p to the empirical probability. The nonlinear mapping. Ordination-preserving regression is a constrained optimization problem that aims to find a monotonically non-decreasing function R that minimizes the mean squared error: ; in: The order-preserving regression calibration function is a piecewise constant, monotonically increasing step function.

[0037] Finally, the C-BNN model output is constructed, and the trained ordinal-preserving regression model is integrated into the backend of the original BNN model to form the final C-BNN model. For any sample to be predicted, its calibrated cumulative distribution function... Defined as: ; This mapping corrects the nominal confidence interval of the original BNN output to a calibration confidence interval, ensuring that when the C-BNN outputs a 95% confidence interval, the probability of the true value falling into that interval strictly converges to 95%.

[0038] For a comparison of model training and calibration results, please refer to Figure 2 Specifically, a variational Bayesian neural network with three hidden layers and 64, 32, and 16 nodes respectively was constructed. The first 60% of the data was used as the training set, employing the Adam optimizer with a learning rate of 0.001 for 2000 epochs. During the calibration phase, statistical indicators were compared before and after calibration. Before calibration, on the test set, the model's 95% confidence interval coverage probability was only 82.4%, indicating a state of "severe overconfidence," meaning the given intervals were too narrow, missing nearly 18% of the true values. This could lead to missed oil and gas layers or misjudgments in geological engineering. Simultaneously, its normalized average prediction interval width was 0.115. After calibration, through the ordinal-preserving regression calibration module, the PICP on the test set improved and stabilized at 95.6%, extremely close to the nominal confidence level, achieving statistically "perfect calibration." More importantly, while coverage improved, NMPIW decreased to 0.091. The optimization results, which show that the interval narrows but the coverage increases, demonstrate that the calibration process effectively corrects the model's erroneous tendency to blindly expand the interval in low uncertainty regions and blindly shrink the interval in high uncertainty regions, thus achieving a re-optimal allocation of uncertainty resources.

[0039] Finally, the data of the well section to be predicted is input into the C-BNN model after step-by-step calibration. The expected value of the calibrated distribution is output. The upper and lower bounds of the calibrated 95% confidence interval are also output. The normalized prediction interval width (NMPIW) is calculated for each depth point. If NMPIW is less than the threshold, the prediction result is reliable. If NMPIW is greater than the threshold, although a prediction value is given, the model indicates extremely high cognitive uncertainty, and the system automatically marks this section as requiring "supplementary data or cautious decision-making," thus achieving intelligent geological risk early warning. The trained C-BNN model is applied to a blind well that did not participate in any training process. The fracture strength curve predicted by C-BNN shows extremely high agreement with the measured FMI curve, with a root mean square error of only 0.12 and a coefficient of determination of 0.92, significantly outperforming the prediction results of support vector machines and traditional BP neural networks under the same conditions. For details, please refer to [link to relevant documentation]. Figure 3 In the 3520m-3560m well section, the model predicted moderate fracture development, but the confidence interval was extremely narrow (NMPIW < 0.05), which the system classified as a "low-risk confidence zone." However, in the 3610m-3630m well section, although the model's point predictions indicated potential fracture development, the confidence interval expanded explosively, triggering a "high-risk warning" automatically. Blindly performing fracturing based solely on point predictions using traditional methods could easily lead to project failure. The warning system of this invention successfully avoided this risk, fully demonstrating its practical engineering value in complex geological environments.

[0040] In summary, the overall technical solution of this invention is as follows: First, a high-dimensional feature space for leakage prevention is constructed. Multi-source logging data and FMI imaging interpretation labels are strictly divided into training, calibration, and test sets. Normalization is performed on all data using only the statistics of the training set to physically prevent leakage of test information. An Automatic Relevance Determination (ARD) technique is used to construct a Bayesian linear regression model. The accuracy hyperparameter is optimized by maximizing the marginal likelihood function, automatically eliminating noisy features where the weight posterior distribution shrinks to zero, and selecting an effective feature subset highly sensitive to fracture response. Next, a variational Bayesian neural network is constructed, treating the weights as random variables. Variational inference theory is introduced, and forward propagation sampling is performed using reparameterization techniques and a Flipout estimator. The lower bound of evidence (EL) containing the negative log-likelihood fitting term and the KL divergence regularization term is maximized. The model parameters are iteratively updated using a BO (British Optimization) method to obtain the original predicted distribution of fracture intensity that can characterize cognitive uncertainty. Subsequently, non-parametric uncertainty calibration is performed by performing multiple Monte Carlo samplings using independent calibration set data, calculating the quantiles of the true label in the predicted distribution and constructing an empirical cumulative distribution function. An ordinal-preserving regression model is trained to establish a monotonically non-decreasing mapping relationship between the original predicted probability and the true empirical probability, and statistical correction is performed on the uncertainty interval. Finally, the data of the well section to be measured is input into the calibrated C-BNN model, which simultaneously outputs the point estimate of fracture intensity and the rigorously calibrated confidence interval. The normalized prediction interval width (NMPIW) is calculated as a risk proxy indicator, and the layers with NMPIW exceeding the preset threshold are automatically identified as high geological risk areas, thereby achieving collaborative decision support for high-precision numerical prediction and intelligent risk early warning.

[0041] For the foregoing embodiments, in order to simplify the description, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to this application.

[0042] The above embodiments describe the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Modifications and variations made by those skilled in the art without departing from the spirit and scope of the invention should be within the protection scope of the appended claims.

Claims

1. A method for predicting reservoir fracture development intensity and quantifying uncertainty, characterized in that, Includes the following steps: S1: Establish a fracture development intensity dataset based on well logging curves and perform preprocessing; S2: Optimize logging response characteristics using Automatic Correlation Determination (ARD) technology; S3: Construct a Bayesian neural network (BNN) model based on crack development intensity and train it to establish a probabilistic network that treats weights as random variables, and approximate the posterior distribution through variational inference. S4: Use the calibration set to calibrate the crack development intensity order-preserving regression uncertainty of the trained BNN model to obtain the crack development intensity uncertainty quantification C-BNN model. S5: Input the data of the well section to be predicted into the C-BNN model, and output the expected value of fracture development intensity through the C-BNN model, and output the 95% confidence interval.

2. The method for predicting reservoir fracture development intensity and quantifying uncertainty as described in claim 1, characterized in that, Step S1 includes the following sub-steps: S11: Construct a high-confidence sample library; S12: Taking into account the sensitivity of logging response to lithology, physical properties and fractures, the input feature space is defined, and then the initial feature pool is constructed; S13: This only involves the partitioning of the dataset, including the training set, calibration set, and test set. S14: The data within the dataset is preprocessed using the Z-Score standardization method to eliminate dimensional differences.

3. The method for predicting reservoir fracture development intensity and quantifying uncertainty as described in claim 2, characterized in that, The initial characteristic pool includes one or more of the following: natural gamma, natural potential, wellbore, deep lateral resistivity, shallow lateral resistivity, flush zone resistivity, density, neutrons, and acoustic transit time.

4. The method for predicting reservoir fracture development intensity and quantifying uncertainty as described in claim 1, characterized in that, Step S2 specifically involves: establishing a Bayesian linear regression model, where each feature in the input feature vector is... Assign a separate hyperparameter The second type of maximum likelihood method is used to optimize the hyperparameters. When the optimization converges, if If the value is less than the preset value, the feature is retained, and the selected preferred feature subset is stored in the training sample library.

5. The method for predicting reservoir fracture development intensity and quantifying uncertainty as described in claim 4, characterized in that, The hyperparameters Control the corresponding weights The prior distribution variance is: ； The maximum log-marginal likelihood function of the second type of maximum likelihood method is: 。 6. The method for predicting reservoir fracture development intensity and quantifying uncertainty as described in claim 1, characterized in that, Step S3 includes the following sub-steps: S31: Construct a fully connected Bayesian neural network using a variational Bayesian neural network architecture, where the weights and biases in the network architecture are random variables that follow a Gaussian distribution; S32: Define variational posterior and reparameterized sampling; S33: Construct a model based on maximizing the lower bound of evidence ELBO, perform iterative training on the BNN model until ELBO converges, and save the model parameters.

7. The method for predicting reservoir fracture development intensity and quantifying uncertainty as described in claim 6, characterized in that, Step S32 specifically involves: employing a variational inference method, introducing a parameterized variational distribution to approximate the true posterior, and using a reparameterization method to address the gradient non-differentiability problem during random sampling, representing the random weight w as a combination of deterministic parameters and noise. ； in, Let be the variational parameters to be learned. To help noise variables, This is for element-wise multiplication.

8. The method for predicting reservoir fracture development intensity and quantifying uncertainty as described in claim 6, characterized in that, The cyclic training specifically involves: using the Flipout estimator to perform Monte Carlo sampling, calculating the predicted value and distribution of the current batch of data, calculating the loss based on the loss function, calculating the gradient of the loss function with respect to the variational parameters using the backpropagation algorithm, updating the parameters using the Adam optimizer, repeating the above process until ELBO converges or the preset number of epochs is reached, and finally saving the final model parameters.

9. The method for predicting reservoir fracture development intensity and quantifying uncertainty as described in claim 1, characterized in that, Step S4 includes the following sub-steps: S41: Input the calibration set data into the trained BNN model and process each sample... Sub-Monte Carlo sampling is performed to obtain the position of the true label of each calibration sample in its predicted cumulative distribution function; S42: Construct a mapping relationship between predicted probabilities and empirical probabilities, and calculate the empirical coverage probability at different nominal confidence levels by statistically analyzing the value distribution of all samples in the calibration set. S43: Train an ordinal-preserving regression model to fit the nonlinear mapping from predicted probability to empirical probability, and integrate the trained ordinal-preserving regression model into the back end of the original BNN model to form the final C-BNN model.

10. The method for predicting reservoir fracture development intensity and quantifying uncertainty as described in claim 9, characterized in that, The cumulative distribution function of the C-BNN model Defined as: ； in, The order-preserving regression calibration function is a piecewise constant, monotonically increasing step function; By using the cumulative distribution function, the nominal confidence interval of the original BNN output is corrected to the calibration confidence interval, ensuring that when the C-BNN outputs a 95% confidence interval, the probability of the true value falling into that interval strictly converges to 95%.

Images