Shale gas horizontal well fracture pressure prediction method and system
By combining multi-source data fusion and intelligent algorithm optimization, a fracture pressure prediction method is developed. This method integrates machine learning and traditional physical models, solving the problem of capturing the nonlinear relationship between geostress and elements in existing technologies. It achieves high-precision fracture pressure prediction and improves the safety and efficiency of shale gas development.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-11
- Publication Date
- 2026-03-27
AI Technical Summary
In existing technologies, multiple regression models assume linear relationships, making it difficult to capture the complex nonlinear and interactive relationships between geostress and elements, which limits the accuracy of shale gas horizontal well fracture pressure prediction.
A high-precision rupture pressure prediction method is generated by employing an intelligent deep correction model that integrates multi-source data fusion, combining machine learning and traditional physical models, and using multi-scale feature enhancement and nonlinear feature selection. This method includes the combined use of intelligent deep correction based on multi-information fusion, machine learning regression models, deep sequence models, tree models, and XGBoost models.
It has enabled the accurate capture of the complex nonlinear relationship between geostress and elemental characteristics, improved the accuracy and robustness of fracture pressure prediction, reduced engineering risks such as wellbore collapse and well leakage, and improved the efficiency of shale gas development.
Smart Images
Figure CN121744264A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of oil and gas fields, and particularly relates to a shale gas horizontal well fracture pressure prediction method and system. BACKGROUND
[0002] With the deepening of the development strategy of shale gas resources in China, shale gas, as an important part of unconventional oil and gas resources, its efficient, safe and economic development has great strategic significance for guaranteeing national energy security, optimizing energy structure and achieving the "double carbon" goal. As a core technical means of shale gas development, the accurate prediction of the stability of the well wall and the fracture pressure of the horizontal well is directly related to the safety of the drilling engineering, the effectiveness of the fracturing operation and the economy of the overall development. A shale gas horizontal well fracture pressure prediction method and system based on multi-source data fusion, intelligent algorithm optimization and dynamic geomechanical modeling is developed, which aims to integrate geological, engineering, logging and other multidimensional data to build a high-precision, strong-robust and real-time updated fracture pressure prediction model, realize the dynamic quantitative characterization of the stress field and the fracture threshold of the formation around the horizontal well bore, and provide a scientific basis for drilling parameter optimization, fracturing scheme design and well wall stability evaluation, effectively reduce the engineering risks such as well wall collapse, lost circulation and fracturing failure, improve the development efficiency and resource recovery rate of shale gas, promote the upgrading of shale gas development technology to intelligent, accurate and safe, and finally realize the efficient, green and sustainable development of shale gas resources. It has far-reaching and important practical significance and strategic value for guaranteeing national energy supply security, promoting energy industry transformation and upgrading and supporting economic and social high-quality development.
[0003] In the prior art, the relationship between the ground stress and the element combination usually has nonlinear and multi-scale characteristics, and the multiple regression model assumes a linear relationship, which is difficult to capture the complex nonlinear and interactive relationship between the element and the ground stress, resulting in limited prediction accuracy. Therefore, it is particularly important to propose a shale gas horizontal well fracture pressure prediction method and system. SUMMARY
[0004] The purpose of the present application is to solve the problems existing in the prior art and provide a shale gas horizontal well fracture pressure prediction method and system.
[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme: A shale gas horizontal well fracture pressure prediction method, comprising the following steps: Step one: multi-source data acquisition and intelligent depth correction: acquire the logging data, element logging data, logging while drilling data, drilling engineering parameters and well trajectory data of the target horizontal section, use the intelligent depth correction model of multi-information fusion to intelligently match the natural gamma ray measurement results with the comprehensive geological profile composed of multiple logging while drilling curves, correct the depth of all data, and output the standardized data set with unified depth; Step two: multi-scale feature enhancement: take the high-resolution logging data corrected in step one as features and take the low-resolution element logging data as labels, train a machine learning regression model to generate pseudo-element curves matching the resolution of the logging data, and use a depth sequence model to extract local patterns and sequence features from the pseudo-element curves and high-resolution logging data to generate high-dimensional feature sequences; Step three: nonlinear feature screening and mixed ground stress prediction: based on the tree model, the importance of element features in the high-dimensional feature sequence is evaluated, and the SHAP value analysis is used to quantify the contribution of each feature to the ground stress prediction, and the key element combination is screened out, based on the acoustic and density logging data, the initial ground stress profile is calculated by the traditional physical model, and the key element combination and other logging features are used as input to train a machine learning residual prediction model to predict the residual between the calculated value and the actual measured value of the physical model. Add the initial ground stress profile and the residual prediction value to get the final predicted ground stress data; Step four: fracture pressure calculation and verification: according to the ground stress data predicted in step three, calculate the wellbore three-dimensional principal stress, and derive the fracture pressure based on the tensile strength theory, and verify the prediction results by using the actual fracturing test data.
[0006] The above scheme further comprises: Further, in step one, the multi-information fusion intelligent depth correction model uses a dynamic time warping algorithm for depth matching.
[0007] Further, in step two, the machine learning regression model is a gradient boosting tree.
[0008] Further, in step two, the depth sequence model is a long short-term memory network, which extracts features on the depth sequence in a set sliding window.
[0009] Further, in step three, the tree model is a random forest.
[0010] Further, in step three, the machine learning residual prediction model is an XGBoost model.
[0011] Further, in step four, the actual fracturing test data is downhole microseismic monitoring data or fracturing construction pressure curve.
[0012] A shale gas horizontal well fracture pressure prediction system used by a shale gas horizontal well fracture pressure prediction method, comprising: The data acquisition and preprocessing module is used for acquiring logging data, element logging data, logging while drilling data, drilling engineering parameters and wellbore trajectory data of a target well section; The multi-source data intelligent depth correction module is used for integrating the logging while drilling data, drilling engineering parameters and wellbore trajectory data, intelligently matching and correcting depths through a dynamic time warping algorithm or a sequence-to-sequence model, and outputting multi-source data after depth correction; The multi-scale feature enhancement module is used for taking the element logging data as a label and the logging data as a feature, and generating high-resolution pseudo-element curves through a machine learning model; The nonlinear feature screening and ground stress prediction module comprises: The nonlinear feature screening unit is used for screening key element combinations from the pseudo-element curves based on a tree model and SHAP value analysis; The physical-data hybrid modeling unit is used for calculating initial ground stress by using a traditional physical model, and predicting the residual error of the physical model by using a machine learning residual prediction model, and adding the two to obtain final predicted ground stress data; The fracture pressure calculation module is used for calculating fracture pressure according to the predicted ground stress data.
[0013] The present application has the following beneficial effects: 0、In the present application, the formation heterogeneity is finely described through high-resolution pseudo-element curves and depth sequence features, the complex nonlinear relationship between ground stress and element features is effectively captured through the hybrid of a traditional physical model and a machine learning residual prediction model, the basic physical reasonableness of the prediction result is verified, and the data-driven residual prediction learns the deviation caused by complex factors such as local heterogeneity, which cannot be accurately described by the physical model. This “foundation + correction” mode maximizes the prediction accuracy without sacrificing physical consistency.
[0014] 1、In the present application, the intelligent depth correction model of multi-information fusion is used to intelligently match the natural gamma ray measurement results of cuttings with the comprehensive geological profile composed of multiple logging while drilling curves. The model can automatically learn and compensate for systematic depth deviation caused by inaccurate estimation of cuttings return time, measurement error, etc., and realize more robust and more accurate depth homing. BRIEF DESCRIPTION OF DRAWINGS
[0015] Fig. 1 A flowchart of a shale gas horizontal well fracture pressure prediction method proposed by the present application; Fig. 2A system block diagram of a shale gas horizontal well fracture pressure prediction system is provided. DETAILED DESCRIPTION
[0016] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.
[0017] Please refer to Figs. 1-2 The present application is a shale gas horizontal well fracture pressure prediction method, comprising the following steps: Step one: multi-source data acquisition and intelligent depth correction: acquire the logging data, element logging data, logging while drilling data, drilling engineering parameters and well trajectory data of the target horizontal well section, use the intelligent depth correction model of multi-information fusion to intelligently match the natural gamma ray measurement results with the comprehensive geological profile composed of multiple logging while drilling curves, correct the depth of all data, and output the standardized data set with unified depth; Step two: multi-scale feature enhancement: use the high-resolution logging data corrected in step one as features and use the low-resolution element logging data as labels to train a machine learning regression model, generate pseudo-element curves matching the resolution of the logging data, and use a depth sequence model to extract local patterns and sequence features from the pseudo-element curves and high-resolution logging data to generate high-dimensional feature sequences; Step three: nonlinear feature screening and mixed ground stress prediction: based on the tree model, evaluate the importance of element features in the high-dimensional feature sequences, use SHAP value analysis to quantify the contribution of each feature to ground stress prediction, screen out key element combinations, based on acoustic and density logging data, calculate the initial ground stress profile through a traditional physical model, use the key element combinations and other logging features as inputs to train a machine learning residual prediction model to predict the residual between the calculated value of the physical model and the actual measured value, add the initial ground stress profile and the residual prediction value to obtain the final predicted ground stress data; Step four: fracture pressure calculation and verification: calculate the three-dimensional principal stress around the well according to the ground stress data predicted in step three, derive the fracture pressure based on the tensile strength theory, and verify the prediction results using actual fracturing test data.
[0018] In one embodiment, in step one, the intelligent depth correction model of multi-information fusion uses a dynamic time warping algorithm for depth matching.
[0019] In this embodiment: Data preprocessing and feature extraction: Cuttings GR data: Cuttings GR data represents the separation of cuttings from the drilling return mud, measures the gamma ray intensity of cuttings using natural gamma logging technology, reflects the content of radioactive elements (such as uranium, thorium, potassium) in the rock formation, and is standardized. Standardization is obtained by drainage method or caliper measurement to obtain the volume of cuttings, combined with mass to calculate density, and standardized to a unified dimension.
[0020] LWD comprehensive profile construction: Fusion of multiple LWD curves (such as GR, resistivity, density), combined with formation dip angle, porosity and other parameters, to form a three-dimensional geological model. For example, by projecting the contact line and dip angle information to construct the profile, ensuring that the true formation structure is reflected.
[0021] Dynamic time warping (DTW) matching process: Distance matrix calculation: Calculate the Euclidean distance between the cuttings GR sequence and each point of the LWD comprehensive profile to form an n x m matrix.
[0022] Cumulative distance calculation: After initializing the first row and column, recursively fill the matrix, such as: .
[0023] Path backtracking: From the end point to the starting point, determine the best alignment path, such as the matching path of cuttings GR sequence [10, 20, 30, 25, 15] and LWD comprehensive profile [12, 18, 28, 24, 16] may be (1, 1) → (2, 2) → (3, 3) → (4, 4) → (5, 5), corresponding to the minimum cumulative distance.
[0024] Fracture pressure prediction: Calculate the fracture pressure gradient, represented as where, is the density, is the dip angle, is the formation pressure coefficient, is the depth, A and B are the starting / ending points, and 1 is the intermediate reference point for segmenting the dip angle-depth gradient.
[0025] In one embodiment, in step two, the machine learning regression model is a gradient boosting tree.
[0026] In this embodiment: Data preparation: Element logging labels (Si content) and logging features (GR, RT), low-resolution element logging data is interpolated to 0.1m interval by cubic spline, and logging data depth is aligned with logging features Z-score standardization; Model training: Construct a gradient boosting tree model, input features are GR (range 50-150 API) and RT (range 10-1000 Ω·m), and the output is pseudo-Si content. After training, the feature importance is: GR accounts for 40%, RT accounts for 35%, and others account for 25%. Pseudo curve generation: input logging data at 0.1m intervals, model output pseudo-element value at corresponding depth, predict pseudo-Si content every 0.1m from 1000.0-1002.0m, such as 1000.0m predicted value 26.2%, 1000.1m predicted value 26.5%, finally generate pseudo curve resolution 0.1m, completely match with logging data, apply Savitzky-Golay filter (window=5 points, polynomial order=2) to eliminate prediction noise and ensure the smoothness of the generated pseudo curve. Accuracy verification and constraint: Cross-validation: 5-fold cross-validation, R² score ≥0.85, MAE ≤2% Physical constraint: Introduce formation rock physics relationship constraint, such as negative correlation between silicon content and density (ρ=2.65−0.01×Si), to ensure that the prediction result conforms to the geological law. Resolution verification: Fourier transform analysis of pseudo curve spectrum to ensure that the main frequency matches the 0.1m resolution (Nyquist frequency ≤5Hz).
[0027] In one embodiment, in step two, the depth sequence model is a long short-term memory network that extracts features from a depth sequence with a set sliding window.
[0028] In this embodiment: Depth sequence data preprocessing: Data standardization: Z-score standardization of logging data (GR, resistivity, density, etc.) to eliminate dimension differences and improve model convergence speed.
[0029] Depth alignment and resampling: Linear interpolation of logging data with different sampling rates to 0.1m resolution to ensure sequence continuity.
[0030] Sliding window construction and feature extraction: Window division: 5m window length, 2.5m step sliding to generate depth sequence fragments. For example, depth 1000-1005m is a window, and the next window is 1002.5-1007.5m.
[0031] LSTM sequence input: 0.1m resolution data in each window forms a 50-time-step input sequence.
[0032] Feature extraction: The LSTM layer outputs the hidden state of the last time step as a window-level feature vector, capturing local patterns of thin interbeds and fracture zones.
[0033] Heterogeneity pattern recognition: Feature fusion: Concatenate the window features extracted by LSTM with static geological parameters (e.g., porosity, permeability) to form a comprehensive feature vector. Introduce attention weight to assign different contribution degrees to different depth points, highlighting key heterogeneity features.
[0034] Classification / regression layer: Map to the target space through a fully connected layer, such as fracture pressure prediction (regression task) or heterogeneity type identification (classification task). Use mean squared error (MSE) or cross-entropy as the loss function.
[0035] In one embodiment, in step three, the tree model is a random forest.
[0036] Model configuration: Set random forest parameters (number of trees = 500, maximum depth = 15, minimum sample split = 10), and use mean squared error (MSE) as the splitting criterion.
[0037] Feature importance calculation: Extract feature importance scores after training, sorted in descending order. Calculate the reduction in Gini impurity or mean squared error for each feature at the splitting nodes of all decision trees in the random forest to quantify the global importance of the feature. Mathematically expressed as: where, is the parent node Gini coefficient, / is the child node Gini coefficient.
[0038] Local SHAP value calculation: For a single sample, calculate the SHAP value of each feature based on the Shapley value theory, which calculates the contribution value of each feature to the individual prediction result , with the sign indicating the contribution direction (promote / inhibit stress) and the absolute value indicating the contribution size. The global importance is the mean of all sample SHAP values: .
[0039] Global SHAP value aggregation: Calculate the mean of all sample SHAP values to generate a feature contribution heat map. For example: Si content global SHAP value: +0.4MPa (high positive correlation contribution); Ca content global SHAP value: -0.2MPa (high negative correlation contribution).
[0040] Two-dimensional screening strategy: High importance features: Select features with random forest importance scores >0.1 (e.g., Si, Ca, GR).
[0041] Significant contribution features: select features with SHAP value absolute value mean > 0.1 and consistent direction (e.g. Si content positively correlated, Ca content negatively correlated).
[0042] Combination optimization: evaluate the prediction performance of different element combinations through cross-validation, and select the optimal combination. For example: combination 1 (Si+Ca+GR): R²=0.85, MAE=0.5MPa; combination 2 (Si+Al+density): R²=0.82, MAE=0.6MPa.
[0043] Finally, combination 1 is selected as the key element combination.
[0044] In one embodiment, in step three, the machine learning residual prediction model is an XGBoost model.
[0045] In one embodiment, in step four, the actual fracturing test data is downhole microseismic monitoring data or fracturing construction pressure curve.
[0046] A shale gas horizontal well fracture pressure prediction system used by a shale gas horizontal well fracture pressure prediction method, comprising: The data acquisition and preprocessing module is used to acquire well logging data, element logging data, logging while drilling data, drilling engineering parameters and wellbore trajectory data of a target well section; The multi-source data intelligent depth correction module is used to integrate the logging while drilling data, drilling engineering parameters and wellbore trajectory data, and perform intelligent depth matching and correction through a dynamic time warping algorithm or a sequence-to-sequence model, and output multi-source data after depth correction; The multi-scale feature enhancement module is used to take the element logging data as a label and the well logging data as a feature, and generate high-resolution pseudo-element curves through a machine learning model; The nonlinear feature screening and geostress prediction module comprises: The nonlinear feature screening unit is used to screen key element combinations from the pseudo-element curves based on a tree model and SHAP value analysis; The physics-data hybrid modeling unit is used to calculate initial geostress by using a traditional physical model, and predict the residual of the physical model by using a machine learning residual prediction model, and add the two to obtain final predicted geostress data; The fracture pressure calculation module is used to calculate fracture pressure according to the predicted geostress data.
[0047] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.
Claims
1. A method for predicting fracture pressure in shale gas horizontal wells, characterized in that, Includes the following steps: Step 1: Multi-source data acquisition and intelligent depth correction: Acquire logging data, element logging data, logging-while-drilling data, drilling engineering parameters and wellbore trajectory data of the target horizontal well section. Utilize the intelligent depth correction model that integrates multiple information sources to intelligently match the natural gamma ray measurement results of cuttings with the comprehensive geological profile composed of multiple logging-while-drilling curves. Perform depth repositioning correction on all data and output a standardized dataset with unified depth. Step 2: Multi-scale feature enhancement: Using the high-resolution logging data corrected in Step 1 as features and the low-resolution element logging data as labels, a machine learning regression model is trained to generate pseudo-element curves that match the resolution of the logging data. A deep sequence model is then used to extract local patterns and sequence features from the pseudo-element curves and the high-resolution logging data to generate a high-dimensional feature sequence. Step 3: Nonlinear Feature Screening and Hybrid Geostress Prediction: The importance of element features in the high-dimensional feature sequence is evaluated based on a tree model, and the contribution of each feature to geostress prediction is quantified using SHAP value analysis. Key element combinations are screened out. Based on sonic and density logging data, an initial geostress profile is calculated using a traditional physical model. Using the key element combinations and other logging features as input, a machine learning residual prediction model is trained to predict the residual between the physical model's calculated value and the actual measured value. The initial geostress profile is added to the residual prediction value to obtain the final predicted geostress data. Step 4: Calculation and verification of fracturing pressure: Based on the geostress data predicted in Step 3, calculate the three-dimensional principal stress of the well, derive the fracturing pressure based on the tensile strength theory, and verify the prediction results using actual fracturing test data.
2. The method for predicting fracture pressure in a shale gas horizontal well according to claim 1, characterized in that, In step one, the intelligent depth correction model based on multi-information fusion uses a dynamic time warping algorithm for depth matching.
3. The method for predicting fracture pressure in a shale gas horizontal well according to claim 1, characterized in that, In step two, the machine learning regression model is a gradient boosting tree.
4. The method for predicting fracture pressure in a shale gas horizontal well according to claim 1, characterized in that, In step two, the deep sequence model is a long short-term memory network, which uses a set sliding window to slide across the deep sequence to extract features.
5. The method for predicting fracture pressure in a shale gas horizontal well according to claim 1, characterized in that, In step three, the tree model is a random forest.
6. The method for predicting fracture pressure in a shale gas horizontal well according to claim 1, characterized in that, In step three, the machine learning residual prediction model is the XGBoost model.
7. The method for predicting fracture pressure in a shale gas horizontal well according to claim 1, characterized in that, The actual fracturing test data mentioned in step four are downhole microseismic monitoring data or fracturing construction pressure curves.
8. The shale gas horizontal well fracture pressure prediction system used in the shale gas horizontal well fracture pressure prediction method according to claim 1, characterized in that, include: The data acquisition and preprocessing module is used to acquire logging data, elemental logging data, logging-while-drilling data, drilling engineering parameters, and wellbore trajectory data for the target well section; The multi-source data intelligent depth correction module is used to integrate the logging-while-drilling data, drilling engineering parameters and wellbore trajectory data, and to perform intelligent depth matching and correction through dynamic time warping algorithm or sequence to sequence model, and output multi-source data with depth correction. The multi-scale feature enhancement module is used to generate high-resolution pseudo-element curves by using the element logging data as labels and the logging data as features through a machine learning model. The nonlinear feature screening and geostress prediction module includes: The nonlinear feature filtering unit is used to filter out key element combinations from the pseudo-element curve based on the tree model and SHAP value analysis; The physical-data hybrid modeling unit is used to calculate the initial geostress using a traditional physical model and to predict the residuals of the physical model using a machine learning residual prediction model. The two are then added together to obtain the final predicted geostress data. The rupture pressure calculation module is used to calculate the rupture pressure based on the predicted geostress data.