A deep coal bed gas fracture network density prediction method
By constructing a fracture density prediction model based on geological and engineering parameters, and utilizing XGBoost decision trees and semi-supervised learning, the problem of obtaining fracture density data in deep coalbed methane wells was solved. This enabled high-precision fracture density prediction and fracturing effect evaluation, optimized fracturing design, and improved recovery rate.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTHWEST PETROLEUM UNIV
- Filing Date
- 2026-02-09
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies make it difficult to efficiently and cost-effectively obtain fracture network density data in deep coalbed methane wells, resulting in inaccurate evaluation of fracturing effects. Traditional methods rely on expensive microseismic monitoring or subjective interpretation, which cannot quantify fracture network density.
A machine learning approach based on geological and engineering parameters is used to construct a seam density prediction model. XGBoost gradient boosting decision tree is used, and the average seam spacing is used as a representation label for seam density. Multi-source data is fused through a semi-supervised learning framework to achieve intelligent prediction of seam density.
It achieves high-precision, low-cost prediction of fracture density, enabling accurate assessment of fracturing effects, optimization of fracturing design, and improvement of oil recovery.
Smart Images
Figure SMS_1 
Figure SMS_2 
Figure SMS_3
Abstract
Description
Technical Field
[0001] This invention relates to the field of reservoir development technology, and also to intelligent prediction technology based on machine learning and reservoir stimulation effect evaluation technology, specifically a method for predicting the density of fracture networks in deep coalbed methane. Background Technology
[0002] Deep coalbed methane, as an important unconventional natural gas resource, often employs hydraulic fracturing technology to enhance its recovery rate. During fracturing, high-pressure fluids form complex fracture networks (fracture networks) within the coal and rock. The complexity and density of these fracture networks determine the reservoir's stimulation volume and final gas production. Compared to conventional sandstone and shale reservoirs, deep coalbed methane reservoirs exhibit more developed cleavage fractures, making it easier to form orthogonal, complex fracture networks along these cleavage fractures.
[0003] Currently, microseismic monitoring is the most direct means of characterizing fracture morphology. However, due to the high cost and complex deployment of microseismic monitoring technology, it is difficult to apply it on a large scale in every deep coalbed methane well, resulting in extremely limited channels for directly obtaining fracture network spatial morphology data. Furthermore, traditional methods often rely on engineers' experience to qualitatively interpret microseismic event cloud maps, or use simple geometric models (such as elliptical models) based on the planar fracture assumption for approximation, which are highly subjective and cannot be quantified. While existing methods that evaluate fracture network density using microseismic event density (i.e., the number of microseismic events per unit volume) can intuitively reflect the degree of rock fragmentation, fracture network density is also affected by many other factors. This leads to inconsistencies between the quantification results of using microseismic event density and the evaluation of fracturing effectiveness. For example, two locations may have similar microseismic event densities (i.e., similar fracture densities), but the reservoir stimulation volume after fracturing may differ significantly.
[0004] In summary, given the scarcity of microseismic data and the inaccuracy of existing quantification methods, it is necessary to develop an intelligent prediction method that can predict fracture network density by migrating and integrating geological and engineering parameters without relying on expensive on-site monitoring. Summary of the Invention
[0005] To address the aforementioned issues, this invention provides a method for predicting the fracture density of deep coalbed methane. This method uses the average fracture spacing to characterize the fracture density of coalbed methane and predicts the average fracture spacing based on geological and engineering parameters. This enables high-precision, automated, and intelligent prediction, providing key technical support for optimizing fracturing design and improving oil recovery.
[0006] To achieve the above objectives, the present invention provides the following solution:
[0007] A method for predicting the density of fracture networks in deep coalbed methane includes the following steps:
[0008] S1. Collect historical data for each deep coalbed methane fracturing section, including engineering parameters, geological parameters, and microseismic data;
[0009] S2. Using the fracturing section as the sample unit, a dataset is constructed based on historical data. Each sample in the dataset includes engineering parameters and geological parameters. The label of the sample is the average fracture-network spacing. The average fracture-network spacing is the average of the nearest neighbor distances of all microseismic event points. The nearest neighbor distance of each microseismic event point is the minimum Euclidean distance between that microseismic event point and other microseismic event points.
[0010] S3. Construct a seam density prediction model based on XGBoost gradient boosting decision tree, and train it using the dataset to obtain the trained seam density prediction model.
[0011] S4. Collect the engineering and geological parameters of the fractured section to be predicted, input them into the trained fracture network density prediction model, and predict the average fracture network spacing of the fractured section. The smaller the average fracture network spacing, the greater the fracture network density.
[0012] This invention relates to geological parameters, engineering parameters, and microseismic data. Geological parameters refer to static parameters describing the inherent geological properties of the target coal seam and its surrounding rock. Engineering parameters refer to dynamic operational parameters that can be monitored and controlled in real time during hydraulic fracturing operations. Microseismic data are dynamic monitoring parameters describing the rock fracture response characteristics of the target coal seam under the action of hydraulic fracturing forces. As a specific embodiment of this invention, geological parameters include anisotropy, geostress field, Young's modulus, and Poisson's ratio; engineering parameters include fracturing displacement, cumulative fluid volume, sand addition volume, and fracturing pressure; and microseismic data includes the spatial coordinates of event points.
[0013] As a specific embodiment of the present invention, the mesh density prediction model is composed of multiple XGBoost models, and the average value of the output of all XGBoost models is the final output.
[0014] Each XGBoost model includes:
[0015] Prediction formula:
[0016]
[0017] Objective function calculation formula:
[0018]
[0019]
[0020] In the formula: Indicates the first t After the first iteration i Predicted values for each sample; Indicates the precedingt -1 round of iterations i Predicted values for each sample; Indicates the first t The prediction results of each decision tree; This represents the learning rate, which controls the contribution of each tree. Indicates the first t The overall objective function during rounds of training; This represents the mean squared error loss function term, which measures the difference between the model's predicted value and the true label value. y i The differences between them; This represents a regularization term used to penalize the complexity of the model; This represents the complexity control parameter, the minimum loss reduction required for leaf node splitting; the larger the value, the more conservative the model. T Indicates the first t The number of nodes in the tree; Indicates the L2 regularization weights; Indicates the first j The weight of each leaf node.
[0021] In one specific embodiment of the present invention, step S3 adopts a semi-supervised training process, including the following steps:
[0022] S31. Divide the samples in the dataset into two categories according to whether they contain microseismic data. Those containing microseismic data are assigned to the labeled training set, and those without microseismic data are assigned to the unlabeled training set.
[0023] S32. Train the seam density prediction model with the labeled training set to obtain a temporary prediction model. If the prediction accuracy of the temporary prediction model meets the requirements, stop; otherwise, proceed to step S33.
[0024] S33. Predict the labels of samples in the unlabeled training set for each XGBoost model in the temporary model, and calculate the mean square error of all labels for that sample.
[0025] S34. Repeat step S33 to traverse all samples in the unlabeled training set and obtain the mean squared error for each sample.
[0026] S35. Select samples with small mean square error from the unlabeled training set according to a preset ratio, transfer them to the labeled training set, thereby updating the labeled training set and the unlabeled training set, and then return to step S32.
[0027] This invention selects the average nearest neighbor distance as the average fracture network spacing to characterize the fracture network density of deep coalbed methane reservoirs. This selection is based on the inventors' long-term research findings: the fractures in deep coalbed methane reservoirs exhibit orthogonal characteristics. During fracturing, microseismic event points reflect the fracturing of coal and rock. The spatial distance between microseismic event points can effectively reflect the fracture network distance between each fracture in the final fracture network. The smaller the fracture network spacing value, the denser the fracture network, and the larger the contact area between the fracturing fluid and the coal matrix. This is more conducive to the "permeation-displacement" production enhancement mechanism of deep coalbed methane, thus indicating a more ideal fracturing effect. The average nearest neighbor distance has a high consistency with the fracturing effect and can accurately evaluate the fracturing effect. Therefore, this parameter is suitable for characterizing the fracture network density.
[0028] Compared with existing calculation methods, the present invention has the following advantages:
[0029] This invention proposes using the average fracture network spacing as a characterization label for fracture network density, and using the average nearest neighbor distance between microseismic event points as the average fracture network spacing. This can better reflect the spatial density of the fracture network, more accurately assess the contact area between fracturing fluid and coal matrix, and thus more scientifically predict the fracturing effect.
[0030] This invention constructs a multi-source data fusion framework with fracturing sections as sample units, integrates geological parameters and engineering parameters into feature vectors, and automatically learns the complex nonlinear relationship between features and average fracture network spacing through machine learning. This method does not rely on expensive microseismic monitoring data, and can achieve intelligent prediction of fracture network density with only conventional geological and engineering parameters, which greatly reduces costs and operational complexity. Detailed Implementation
[0031] The present invention will be further described in detail below with reference to the embodiments, but the implementation of the present invention is not limited thereto.
[0032] A method for predicting the density of fracture networks in deep coalbed methane includes the following steps:
[0033] S1. Collect historical data for each deep coalbed methane fracturing section, including engineering parameters, geological parameters, and microseismic data;
[0034] In this embodiment, geological parameters refer to static parameters that describe the inherent geological properties of the target coal seam and its surrounding rock, mainly including:
[0035] Anisotropy characterizes the differences in the mechanical properties of coal and rock in different directions;
[0036] The geostress field includes the minimum horizontal principal stress ( ), maximum horizontal principal stress ( ) and vertical stress ( Three types of stress (unit: MPa);
[0037] Young's modulus (E) characterizes the resistance of coal and rock to deformation (unit: GPa).
[0038] Poisson's ratio ), representing the ratio of lateral deformation to longitudinal deformation of coal and rock (dimensionless).
[0039] Engineering parameters refer to dynamic operational parameters that can be monitored and controlled in real time during fracturing operations, mainly including:
[0040] Construction displacement (Q), the volume of fracturing fluid injected into the formation per unit time (unit: m³). 3 / min);
[0041] Cumulative liquid volume ( The total volume of fracturing fluid injected into the formation during a single stage or single fracturing operation (unit: m³). 3 );
[0042] Sand addition amount ( Total mass of proppant injected into the formation during a single stage or single fracturing operation (unit: t).
[0043] Construction pressure (P) is the peak pressure during the fracturing cycle of the fracturing section, used to characterize the maximum bottom hole pressure faced during fracturing operations (unit: MPa).
[0044] Microseismic data are dynamic monitoring parameters that describe the rock fracture response characteristics of a target coal seam under the action of fracturing forces. They mainly include the spatial coordinates of event points, that is, the three-dimensional location information of each microseismic event source, including the east-west coordinate X (unit: m), the north-south coordinate Y (unit: m), and the vertical depth Z (unit: m).
[0045] S2. Using the fracturing section as the sample unit, a dataset is constructed based on historical data. Each sample in the dataset includes construction parameters and geological parameters. The label of the sample is the average fracture mesh spacing. The average fracture mesh spacing is the average of the nearest neighbor distances of all microseismic event points. The nearest neighbor distance of each microseismic event point is the minimum Euclidean distance between that microseismic event point and other microseismic event points.
[0046] The formula for calculating the average mesh spacing is as follows:
[0047]
[0048]
[0049] In the formula, The average seam spacing, N This represents the total number of microseismic event points. D i For the firsti The nearest neighbor distance of each microseismic event point; d ij For the first i The microseismic event point and the first j The distance between micro-seismic event points;
[0050] This invention relies on high-quality, structured datasets for model training and validation. It selects a complete and independent fracturing operation segment as the most basic analytical sample unit. During the fracturing operation, each fracturing segment represents an independent reservoir stimulation event. Its corresponding microseismic event point set, geological parameters, and engineering parameters have inherent uniformity, avoiding mutual interference between data from different fracturing segments and ensuring the independence and representativeness of the samples.
[0051] For the fractured sections that underwent microseismic monitoring, a calculation method based on the defined average fracture mesh spacing was selected to calculate the values for each sample. i Microseismic event set S i The corresponding average seam spacing label is used to obtain the output label. Y i , input vector F i = { f i1 , … , f ij} and tags Y i Combined to obtain labeled samples ( F i , Y i ); f ij Indicates the first i The first fracturing stage j The standardized results of the feature parameters, which are the engineering parameters and geological parameters in step S1; in this embodiment, the standardization method of the feature parameters is Z-score standardization, which aims to eliminate the impact of differences in the dimensions and numerical ranges of different features on the stability and convergence speed of model training.
[0052] For the fractured sections that have not undergone microseismic monitoring, only their geological and engineering parameters can be obtained. These parameters are then standardized using Z-scores to obtain unlabeled samples.
[0053] Combine the M labeled samples with the L unlabeled samples to form the complete dataset. D .
[0054] D ={(F 1, Y 1), ..., ( F M , Y M )}U{ F M+1 ,……, F M+L}
[0055] S3. Construct an XGBoost intelligent prediction model for seam spacing based on a semi-supervised algorithm, and train it using a dataset to obtain a trained seam density prediction model.
[0056] In this embodiment, the mesh density prediction model consists of 5 XGBoost models, and the final output is the average of the outputs of all XGBoost models.
[0057]
[0058] In the formula, For the first i The final predicted label for each sample; For the first i The sample at the th b Predicted values in the sub-model (XGBoost model);
[0059] Each XGBoost model includes:
[0060] Prediction formula:
[0061]
[0062] In the formula: Indicates the first t After the first iteration i Predicted values for each sample; Indicates the preceding t -1 round of iterations i Predicted values for each sample; Indicates the first t The prediction results of each decision tree; This represents the learning rate, which controls the contribution of each tree.
[0063] Objective function calculation formula:
[0064]
[0065]
[0066] In the formula, Indicates the first t The overall objective function during rounds of training; This represents the mean squared error loss function term, which measures the difference between the model's predicted value and the true label value. y i The differences between them; This represents a regularization term used to penalize the complexity of the model; This represents the complexity control parameter, the minimum loss reduction required for leaf node splitting; the larger the value, the more conservative the model. T Indicates the first t The number of nodes in the tree; Indicates the L2 regularization weights; Indicates the first j The weight of each leaf node.
[0067] For each XGBoost model, K-fold cross-validation and grid search are used for hyperparameter optimization: the labeled training set is randomly divided into 5 equal parts, and 4 parts are used as the training subset and 1 part as the validation subset in turn, repeating this process 5 times. For each hyperparameter combination, the root mean square error (RMSE) of the 5 validations is calculated, and the hyperparameter combination with the smallest RMSE is selected as the optimal configuration. The core hyperparameters include: the maximum depth of the tree. d (Set to 6), learning rate or (Set to 0.1), regularization parameter l (Set to 1).
[0068] Since microseismic data is scarce and labeled samples are limited in actual engineering projects, this embodiment adopts a semi-supervised learning framework. This framework can simultaneously utilize a small amount of expensive labeled data and a large amount of easily accessible unlabeled data to improve the model's generalization ability and prediction accuracy.
[0069] The specific semi-supervised training process is as follows:
[0070] (1) Sample classification: The samples in the dataset are divided into two categories according to whether they contain microseismic data. The samples containing microseismic data are assigned to the labeled training set (see step S2 for the label calculation process), and the samples without microseismic data are assigned to the unlabeled training set.
[0071] (2) Model training: The first seam density prediction model (including 5 XGBoost sub-models) is trained using a labeled training set and denoted as M0. This stage is standard supervised learning, the purpose of which is to allow the model to initially establish the basic mapping relationship from geological and engineering features to seam spacing. M0 is the starting point for subsequent pseudo-label generation.
[0072] (3) Pseudo-label generation: Based on each sub-model in the constructed initial model M0, the unlabeled samples are predicted to obtain the five average seam spacings for each unlabeled sample;
[0073] (4) High confidence screening: In order to avoid the pollution of the model by noisy labels, this embodiment adopts a screening strategy based on prediction uncertainty. The relative proportion screening method is adopted. By calculating the prediction variance of all samples in the unlabeled sample set and arranging them in ascending order, the samples with the prediction variance in the top 30% are selected as high confidence pseudo-labels to generate a pseudo-label dataset.
[0074] The formula for calculating the prediction variance is as follows:
[0075]
[0076] In the formula, For the first i The prediction variance of each sample;
[0077] (5) New training set generation: The original labeled training set and the newly generated pseudo-labeled dataset are merged into a new labeled training set, thereby updating the labeled training set. Samples in the newly generated pseudo-labeled dataset are removed from the original unlabeled training set, thereby updating the unlabeled training set. Then, return to step (2) to start a new training until the stopping condition is triggered. In this embodiment, an early stopping condition is also set when the relative change rate of the root mean square error of the model on the test sample set is less than 1% for three consecutive iterations.
[0078] To quantitatively evaluate the prediction accuracy of the model constructed in this invention, the finally selected optimal model was used to verify the independent test set. The root mean square error between the average seam-net spacing predicted by the model and the actual calculated average seam-net spacing was approximately 0.95m. This result indicates that the intelligent prediction method proposed in this invention can achieve high-precision prediction of seam-net spacing and can effectively predict the average seam-net spacing that is difficult to measure directly based on readily available original geological and engineering parameters.
[0079] S4. Collect the engineering and geological parameters of the fractured section to be predicted, input them into the trained fracture network density prediction model, and predict the average fracture network spacing of the fractured section to quantify the fracture network density of the fractured section. The smaller the average fracture network spacing, the greater the fracture network density. This includes the following steps:
[0080] (1) Data acquisition: For any deep coalbed methane fracturing section to be predicted, the required input data mainly includes the original geological and engineering parameters. The geological parameters mainly include the anisotropy of the reservoir section, the geostress field (maximum principal stress, minimum principal stress, vertical stress), elastic modulus, and Poisson's ratio. The engineering parameters mainly include the planned or implemented construction discharge, cumulative liquid volume, sand addition volume, and construction pressure of the reservoir section.
[0081] (2) Data standardization: The collected data needs to be standardized. Each value is transformed to generate a standardized feature vector.
[0082] (3) Prediction result output: The standardized feature vector is input into the trained fracture network density prediction model to output the predicted value of the average fracture network spacing, which is used to characterize the fracture network density. At the same time, based on the prediction results, the optimal combination of construction parameters can be selected by adjusting the engineering parameters (discharge rate, fluid volume) to optimize the fracturing design and improve the recovery rate.
[0083] Furthermore, in deep coalbed methane development, gas production is the most direct reflection of fracturing effectiveness. To verify the applicability of the average fracture spacing as an indicator for characterizing fracture density in deep coal seams, this embodiment also randomly selected 100 samples from the original labeled training set (i.e., a sample set containing microseismic data) (Tables 1-3 show data for some of these samples). Based on the actual gas production 30 days after fracturing, the average fracture spacing and microseismic event point density were calculated respectively. PD The correlation coefficients between the two indicators and gas production.
[0084] The correlation coefficient between average mesh spacing and gas production ( The formula for calculating ) is:
[0085]
[0086] The correlation coefficient between point density and gas production ( The formula for calculating ) is:
[0087]
[0088] In the formula, N test To determine the number of samples in the concentrated fracturing section; ASF i , PD i , Gas i The first i Average seam spacing, microseismic event density, and gas production per sample; , , These represent the average seam spacing, microseismic event point density, and average gas production, respectively.
[0089] The results showed that the correlation coefficient between the average fracture mesh spacing and gas production was -0.896, indicating a strong negative correlation between the average fracture mesh spacing and gas production, with gas production increasing as the fracture mesh spacing decreased. The correlation coefficient between microseismic event point density and gas production was 0.624, indicating a moderate positive correlation between the two, with gas production increasing as the point density increased. From the correlation analysis, the average fracture mesh spacing showed a stronger correlation and was more applicable than point density.
[0090] Table 1 Geological Data
[0091]
[0092] Table 2 Engineering Data
[0093]
[0094] Table 3. Spacing, Dot Density, and Gas Production
[0095]
[0096] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for predicting the density of fracture networks in deep coalbed methane, characterized in that, Includes the following steps: S1. Collect historical data for each deep coalbed methane fracturing section, including engineering parameters, geological parameters, and microseismic data; S2. Using the fracturing section as the sample unit, a dataset is constructed based on the historical data. Each sample in the dataset includes engineering parameters and geological parameters. The label of the sample is the average fracture-network spacing. The average fracture-network spacing is the average of the nearest neighbor distances of all microseismic event points. The nearest neighbor distance of each microseismic event point is the minimum Euclidean distance between that microseismic event point and other microseismic event points. S3. Construct a seam density prediction model based on XGBoost gradient boosting decision tree, and train it using the dataset to obtain the trained seam density prediction model. S4. Collect the engineering and geological parameters of the fractured section to be predicted, input them into the trained fracture network density prediction model, and predict the average fracture network spacing of the fractured section to be predicted. The smaller the average fracture network spacing, the greater the fracture network density.
2. The method for predicting the density of fracture networks in deep coalbed methane according to claim 1, characterized in that, The geological parameters include anisotropy, geostress field, Young's modulus, and Poisson's ratio; the engineering parameters include construction discharge rate, cumulative liquid volume, sand addition rate, and construction pressure; the microseismic data includes the spatial coordinates of event points.
3. The method for predicting the density of fracture networks in deep coalbed methane according to claim 1, characterized in that, The mesh density prediction model consists of multiple XGBoost models, and the final output is the average of the outputs of all XGBoost models. Each XGBoost model includes: Prediction formula: Objective function calculation formula: In the formula: Indicates the first t After the first iteration i Predicted values for each sample; Indicates the preceding t -1 round of iterations i Predicted values for each sample; Indicates the first t The prediction results of each decision tree; Indicates the learning rate; Indicates the first t The overall objective function during rounds of training; This represents the mean squared error loss function term, which measures the difference between the model's predicted value and the true label value. y i The differences between them; This represents a regularization term used to penalize the complexity of the model; This represents the complexity control parameter; T Indicates the first t The number of nodes in the tree; Indicates the L2 regularization weights; Indicates the first j The weight of each leaf node.
4. The method for predicting the density of fracture networks in deep coalbed methane according to claim 3, characterized in that, Step S3 employs a semi-supervised training process, including the following steps: S31. Divide the samples in the dataset into two categories according to whether they contain microseismic data. Those containing microseismic data are assigned to the labeled training set, and those without microseismic data are assigned to the unlabeled training set. S32. Train the seam density prediction model with the labeled training set to obtain a temporary prediction model. If the prediction accuracy of the temporary prediction model meets the requirements, stop; otherwise, proceed to step S33. S33. Predict the labels of samples in the unlabeled training set using each XGBoost model in the temporary prediction model, and calculate the mean square error of all labels for that sample. S34, Repeat step S33 to traverse all samples in the unlabeled training set and obtain the mean squared error for each sample; S35. Select samples with small mean square error from the unlabeled training set according to a preset ratio, and transfer them to the labeled training set to update the labeled training set and the unlabeled training set. Then return to step S32.