Physical trend constraint and hybrid machine learning method for coal reservoir shear wave velocity prediction

Abstract

The application discloses a coal reservoir shear wave velocity prediction method combining physical trend constraint and hybrid machine learning, and belongs to the technical field of coalfield physical exploration. Under the condition of lacking measured shear wave logging data, the method firstly establishes a shear wave velocity trend model by using the physical relationship between the longitudinal wave velocity and the shear wave velocity in a training well; then a residual prediction model based on conventional logging parameters is constructed to perform nonlinear correction on the shear wave velocity trend value; finally, the trend prediction result and the residual prediction result are superposed to obtain a shear wave velocity prediction value of a target well. The application combines physical constraint and data-driven method, effectively improves the precision and stability of shear wave velocity prediction, and is suitable for engineering application of coal reservoirs.

Description

Technical Field

[0001] This invention relates to the field of coalfield physical exploration technology, and in particular to a method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning. Background Technology

[0002] Shear wave velocity (VS) is a key parameter characterizing the elastic mechanical properties and shear strength of coal and rock masses, playing a fundamental role in deep coal seam pressure control, dynamic hazard assessment, and seismic exploration data interpretation. Due to the great depth, high gas content, fractured surrounding rock, and limited downhole construction conditions in deep coal seams, VS logging is costly, time-consuming, and risky. This results in a general scarcity of complete and reliable VS logging data in actual production. Most production wells only have conventional logging and P-wave information, which is insufficient to meet the needs for refined characterization of coal reservoir mechanical zoning and engineering risks.

[0003] Application No. 202510962889.X discloses a transverse wave velocity prediction method based on deep learning and physical constraints, which introduces physical constraint terms such as the Castana relation into the loss function to improve prediction accuracy. However, the aforementioned prior art is essentially still an end-to-end deep learning model, where physical constraints are implemented by adding a regularization term to the loss function of the training target. This approach does not explicitly establish an independent, interpretable, localized physical trend model, but rather implicitly embeds physical laws as penalty terms into the optimization process of the neural network. Specifically, the existing technology has the following drawbacks: First, the physical trend and data-driven features are coupled within the model. When the prediction error term conflicts with the physical constraint term, the final output depends on the weight configuration of each term in the loss function. Physical laws cannot be independently evaluated and corrected as a priori benchmark. Second, the model lacks a hierarchical interpretable architecture. It cannot extract the P-wave and S-wave velocity relationship curves that conform to the regional geological laws, nor can it quantify the contribution of each logging parameter to the deviation of the physical trend. Third, in multi-well joint training scenarios, geological factors such as lithological differences and stress state changes in different well areas can make it difficult for the unified loss function weights to adapt to local features, thereby affecting the model's generalization prediction ability at new well locations. Existing technologies do not yet have a technical solution that clearly defines the residual between the measured VS and the trend value based on the physical trend model of local VP and VS, constructs a residual prediction model with conventional logging curves as input, and comprehensively evaluates the model accuracy and geological interpretability through methods such as cross-validation with one well, target well prediction, and SHAP. This results in problems such as insufficient prediction accuracy of coal reservoir shear wave velocity, limited generalization ability between different wells, and unclear geological meaning within the model.

[0004] Therefore, there is an urgent need for a method for predicting shear wave velocity in coal reservoirs that combines physical trend constraints with hybrid machine learning. Summary of the Invention

[0005] This invention provides a method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning, in order to solve the aforementioned problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning includes:

[0008] S1: Obtain logging data from at least one training well and one target well. The logging data includes P-wave velocity or sonic transit time and at least two conventional logging parameters. The training well also includes measured shear wave velocity data.

[0009] S2: Preprocessing of logging data, including logging curve resampling, outlier removal, noise suppression, and missing data imputation;

[0010] S3: The preprocessed conventional logging parameters are divided into positive and negative indices according to their physical property correlation attributes, and then normalized respectively.

[0011] S4: Based on the P-wave velocity and measured S-wave velocity of the training well, establish a physical trend relationship model between P-wave and S-wave velocities;

[0012] S5: Calculate the shear wave velocity trend value at each depth of the training well based on the physical trend relationship model;

[0013] S6: Subtract the measured shear wave velocity from the shear wave velocity trend value in the training well to obtain the shear wave velocity residual;

[0014] S7: Using normalized conventional logging parameters as input and shear wave velocity residuals as output, construct and train a residual prediction model;

[0015] S8: Substitute the target well P-wave velocity into the physical trend relationship model to obtain the target well S-wave velocity trend value, input the normalized conventional logging parameters of the target well into the residual prediction model to obtain the prediction residual, and superimpose the S-wave velocity trend value and the prediction residual to obtain the target well S-wave velocity prediction result.

[0016] Furthermore, step S2 includes:

[0017] Resampling of well logging data at preset depth intervals;

[0018] Identify and remove outlier data points that exceed statistical thresholds;

[0019] Slide window filtering or median filtering can be used to suppress spike noise;

[0020] Missing data points can be filled using linear interpolation, spline interpolation, or interpolation methods based on statistical models.

[0021] Furthermore, in step S3, positive interval mapping is performed on positive indicators based on their minimum and maximum values ​​in the training well, and negative interval mapping is performed on negative indicators based on their minimum and maximum values ​​in the training well, so that the normalized values ​​are mapped to the interval between zero and one.

[0022] Furthermore, the physical trend relationship model can be any one of a linear function model, a polynomial function model, or a piecewise function model.

[0023] Furthermore, when the physical trend relationship model is a linear function model, the relationship between the P-wave and S-wave velocities is:

[0024] ;

[0025] in, This represents the trend value of the shear wave velocity. For the longitudinal wave velocity, a is the slope coefficient, b is the intercept term, and a and b are obtained by fitting the training well data using the least squares method.

[0026] Furthermore, the formula for calculating the shear wave velocity residual in step S6 is as follows:

[0027] ;

[0028] in, For the shear wave velocity residual, To measure the shear wave velocity, This represents the trend value of the shear wave velocity.

[0029] Furthermore, the residual prediction model can be any one of a regression model, an ensemble learning model, a tree model, or a neural network model.

[0030] Furthermore, the residual prediction model is an ensemble regression model built on multiple decision trees. Step S7 also includes: adopting a leave-one-well cross-validation strategy, using single-well data as the validation set and the remaining well data as the training set, performing multiple rounds of validation on the residual prediction model, calculating the root mean square error, mean absolute error and coefficient of determination for each validation set, and selecting the model with the best overall performance.

[0031] Furthermore, the calculation formula for the target well shear wave velocity prediction result in step S8 is as follows:

[0032] ;

[0033] in To predict shear wave velocity, The target well shear wave velocity trend value, To predict residuals.

[0034] Furthermore, conventional logging parameters include at least two of the following: density, resistivity, neutron porosity, natural gamma, borehole diameter, spontaneous potential, and depth.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] This invention, while maintaining the physical trend constraint of VP–VS, introduces machine learning to finely correct the residuals and uses SHAP to provide interpretable analysis, thus realizing an integrated technical solution for high-precision prediction of shear wave velocity in deep coal seams and analysis of geological causal relationships.

[0037] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention.

[0038] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0039] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0040] Figure 1 This is a flowchart of a method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning, as described in an embodiment of the present invention.

[0041] Figure 2 This is a diagram showing the normalization results of multi-well logging data preprocessing in an embodiment of the present invention.

[0042] Figure 3 This is a schematic diagram illustrating the establishment of a local VP–VS physical trend and the definition of shear wave velocity residuals based on training well data in an embodiment of the present invention.

[0043] Figure 4 This is a comparison chart of the prediction errors of different residual machine learning models and the physical trend model under the leave-one-well cross-validation condition in the embodiments of the present invention.

[0044] Figure 5 This is a comparison chart of the target well depth-shear wave velocity prediction curves in an embodiment of the present invention, including the physical trend prediction results and the prediction results of the hybrid model of the present invention.

[0045] Figure 6The bar chart showing the global importance of input features and the summary point cloud map are provided for the RandomForest residual model based on SHAP in this embodiment of the invention.

[0046] Figure 7 This is a schematic diagram illustrating the SHAP dependency of the main logging features on the prediction results of shear wave velocity residuals in an embodiment of the present invention. Detailed Implementation

[0047] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0048] The embodiments of the present invention provide, as follows Figure 1 As shown, a method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning includes:

[0049] S1: Obtain logging data from at least one training well and one target well. The logging data includes P-wave velocity or sonic transit time and at least two conventional logging parameters. The training well also includes measured shear wave velocity data.

[0050] S2: Preprocessing of logging data, including logging curve resampling, outlier removal, noise suppression, and missing data imputation;

[0051] S3: The preprocessed conventional logging parameters are divided into positive and negative indices according to their physical property correlation attributes, and then normalized respectively.

[0052] S4: Based on the P-wave velocity and measured S-wave velocity of the training well, establish a physical trend relationship model between P-wave and S-wave velocities;

[0053] S5: Calculate the shear wave velocity trend value at each depth of the training well based on the physical trend relationship model;

[0054] S6: Subtract the measured shear wave velocity from the shear wave velocity trend value in the training well to obtain the shear wave velocity residual;

[0055] S7: Using normalized conventional logging parameters as input and shear wave velocity residuals as output, construct and train a residual prediction model;

[0056] S8: Substitute the target well P-wave velocity into the physical trend relationship model to obtain the target well S-wave velocity trend value, input the normalized conventional logging parameters of the target well into the residual prediction model to obtain the prediction residual, and superimpose the S-wave velocity trend value and the prediction residual to obtain the target well S-wave velocity prediction result.

[0057] The following is a detailed description with reference to specific embodiments.

[0058] This embodiment provides a method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning. The method includes the following steps:

[0059] Step S1: Acquire logging data from at least one training well and one target well. The logging data includes P-wave velocity or sonic transit time and at least two conventional logging parameters. In this embodiment, conventional logging parameters include density (DEN), resistivity (deep lateral resistivity (LLD) and shallow lateral resistivity (LLS), neutron porosity (CNL), natural gamma ray (GR), wellbore diameter (CAL), spontaneous potential (SP), depth (DEPTH), and well temperature (TEMP). The training well also includes corresponding measured shear wave velocity data (VS).

[0060] Specifically, at least two coal seam or coal-rock interaction wells with relatively complete logging data within the study area were selected as training wells, including P-wave velocity (VP), S-wave velocity (VS), and the aforementioned conventional logging curves. The logging data acquired from the target well included P-wave velocity (VP) and the aforementioned conventional logging curves, but lacked measured S-wave velocity data.

[0061] Step S2: Preprocess the logging data of the training well and the target well. The preprocessing includes logging curve resampling, outlier removal, noise suppression, and missing data imputation.

[0062] Specifically, the preprocessing steps include:

[0063] (1) Resample the logging data at preset depth intervals to ensure that each curve has consistent data records at the same depth sampling point;

[0064] (2) Identify and remove abnormal data points that exceed the statistical threshold, that is, identify and delete extreme outliers in the logging curve;

[0065] (3) Use sliding window filtering or median filtering to eliminate spike noise and suppress noise interference in well logging curves;

[0066] (4) Use linear interpolation, spline interpolation or interpolation based on statistical models to fill in missing data points and interpolate to fill in the missing points.

[0067] The above preprocessing operations ensure the consistency of data quality between training wells and target wells.

[0068] Step S3: As Figure 2 As shown, the preprocessed conventional logging parameters are divided into positive and negative indices according to their physical property correlation attributes, and then normalized respectively.

[0069] Specifically, based on the physical properties of each curve, curves that are positively correlated with stiffness, such as density (DEN), resistivity (LLD), and LLS, are considered positive indicators, while curves that are related to weakness, such as porosity (CNL) and well diameter (CAL), are considered negative indicators.

[0070] For positive indicators, a positive interval mapping is performed based on their minimum and maximum values ​​in the training wells, using a linear normalization method based on extreme values. The normalization formula is as follows:

[0071] ;

[0072] For negative indicators, a reverse interval mapping is performed based on their minimum and maximum values ​​in the training wells, using a reverse normalization method. The normalization formula is as follows:

[0073] ;

[0074] Where X is the original logging curve value, and These are the minimum and maximum values ​​of the curve in the training well sample, respectively. These are the dimensionless eigenvalues ​​after normalization. This maps the normalized values ​​to the interval between zero and one.

[0075] Record the minimum and maximum values ​​of each feature in the training wells, and apply the same normalization mapping to the target well data accordingly, thereby establishing a unified feature scale system between the training wells and the target wells.

[0076] Step S4: As Figure 3 As shown, a physical trend relationship model between P-wave velocities and S-wave velocities is established based on the P-wave velocity of the training well and the measured S-wave velocity.

[0077] In this embodiment, the physical trend relationship model is a linear function model. Depth samples with both VP and VS are selected from the training well data, and the local VP-VS empirical relationship is obtained by fitting the data using the least squares linear regression method.

[0078] The relationship between the velocities of longitudinal and transverse waves is:

[0079] ;

[0080] in, ΔV represents the trend value of shear wave velocity, VP represents the P-wave velocity, α is the slope coefficient reflecting the sensitivity of shear wave velocity to changes in P-wave velocity within the study area, and β is the intercept term reflecting the baseline level of shear wave velocity when the P-wave velocity is taken as a reference value. Parameters α and β are obtained by least squares fitting of samples in the training wells that simultaneously possess both VP and VS.

[0081] In this embodiment, the fitting yields a = 0.7260 and b = -0.8620, that is:

[0082] ;

[0083] Compared with the classic Castagna mudstone baseline formula, the local fitting results show differences in both slope and intercept. The coefficient of determination R² increased from 0.798 to 0.939, indicating that the local formula improved the goodness of fit to the training well data. The residual dispersion decreased, suggesting that the local formula better reflects the velocity characteristics of deep coal seams and surrounding rocks in the study area.

[0084] In other implementations, the physical trend relationship model can also be either a polynomial function model or a piecewise function model.

[0085] Step S5: Based on the physical trend relationship model established in Step S4, calculate the shear wave velocity trend value at each depth of the training well.

[0086] Specifically, using a physical trend relationship model: The trend shear wave velocity at each depth point in the training well was calculated. This provides a foundation for the subsequent definition of residuals.

[0087] Step S6: Subtract the measured shear wave velocity from the shear wave velocity trend value in the training well to obtain the shear wave velocity residual.

[0088] The formula for calculating the shear wave velocity residual is:

[0089] ;

[0090] in, For the shear wave velocity residual, To measure the shear wave velocity, This represents the trend value of the shear wave velocity.

[0091] residual This reflects the additional deviation caused by the combined effects of multiple factors, such as lithological assemblage, fracture development, pore structure, and fluid-bearing conditions, after considering the VP-VS linear relationship.

[0092] Step S7: Using normalized conventional logging parameters as input and shear wave velocity residuals as output, construct and train a residual prediction model.

[0093] Specifically, normalized conventional logging curves such as GR, DEN, CNL, LLD, LLS, CAL, SP, DEPTH, and TEMP are used as input features, and the residuals are... As a learning objective, a training sample set is constructed for machine learning modeling.

[0094] In this embodiment, the residual prediction model is an ensemble regression model built based on multiple decision trees, specifically using a random forest regression model to predict the residuals. Perform fitting.

[0095] To comprehensively evaluate the model's generalization ability across different well types, a leave-one-well cross-validation strategy was adopted. Data from a single well was used as the validation set, and data from the remaining wells were used as the training set to perform multiple rounds of validation on the residual prediction model. In each round, data from one well was used as the test set, and the remaining wells as the training set. The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) of the physical trend model and the residual prediction model were calculated on the test well. This multi-dimensional evaluation of the model's prediction accuracy allowed the selection of the model with the best overall performance.

[0096] like Figure 4 As shown, this embodiment selects a random forest regression model to fit the residual ΔVS and compares it with other ensemble learning methods. To comprehensively evaluate the model's generalization ability across different wells, a leave-one-well cross-validation method is adopted: in each round, the data from one well is used as the test set, and the remaining wells are used as the training set. The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) of the physical trend model and each residual model on the test well are calculated to evaluate the model's prediction accuracy from multiple dimensions. The accuracy improvement ratio of the residual model relative to the physical trend model is also statistically analyzed, as shown in Table 1.

[0097] Goodness of fit: ;

[0098] Root mean square error: ;

[0099] Mean absolute error: ;

[0100] The verification results show that the random forest residual model can significantly reduce the prediction error in most well types, with an improvement rate close to or exceeding 90%. It has better performance than the model compared in this embodiment. Therefore, random forest is selected as the core algorithm for nonlinear correction of residuals in the hybrid model in this embodiment.

[0101] Table 1

[0102] In this embodiment, the parameters of the random forest model are configured as follows: the number of trees n_estimators is 400, the maximum depth max_depth is not limited (the split depth is determined adaptively by the data), the minimum number of samples for internal nodes min_samples_split is 4, the minimum number of samples for leaf nodes min_samples_leaf is 2, and multi-threaded computation is enabled to improve training efficiency.

[0103] In other implementations, the residual prediction model can also be any of the following: a regression model, other ensemble learning models (such as XGBoost, LightGBM, CatBoost), a tree model, or a neural network model.

[0104] Step S8: As Figure 5 As shown, the target well's P-wave velocity is substituted into the physical trend relationship model to obtain the target well's S-wave velocity trend value. The normalized conventional logging parameters of the target well are input into the residual prediction model to obtain the prediction residual. The S-wave velocity trend value and the prediction residual are superimposed to obtain the target well's S-wave velocity prediction result.

[0105] Specifically, the normalized logging features of the target well are input into a cross-validated random forest residual model to obtain the predicted residuals at each depth point of the target well. And compared with the trend shear wave velocity calculated by the physical trend model. By superimposing the data, the predicted shear wave velocity of the target well can be obtained. Will and By comparing the data plotted at different depths on the same graph, it can be observed that while inheriting the overall trend, the hybrid model makes more precise adjustments to the shear wave velocities at locations such as high-resistivity dense layers, low-resistivity weak layers, and coal-rock transition zones. This makes the predicted curves more closely match the actual mechanical distribution of the coal and rock mass in detail. For well sections with some measured shear wave velocity data, further adjustments can be made at this stage. Error statistics were performed between the measured VS and the actual VS to further verify the predictive reliability of the method of the present invention on the target well. In this embodiment, the hybrid model predicts the shear wave velocity of the target well using conventional logging curves (GR, DEN, CNL, LLD, LLS, CAL, SP, DEPTH, TEMP) as input features, and the residual ΔVS between the measured shear wave velocity and the trend value as the prediction target. A mapping function f(x) is constructed to predict the residual:

[0106] ;

[0107] The final prediction formula is:

[0108] ;

[0109] In this case This is the residual prediction function obtained based on random forest training, and its output approximates ΔVS.

[0110] To improve the geological interpretability of the method, SHAP value analysis was performed on the random forest residual prediction model based on training well data.

[0111] like Figure 6 As shown, the SHAP method is first used to calculate the residuals of each input feature pair. The average absolute contribution was used to rank the global importance of features. From the SHAP global importance bar chart and Summary point cloud, it can be seen that density (DEN) contributes significantly more to the residuals than other features, indicating that variations in rock skeleton density are the main factor driving the deviation of shear wave velocity from the VP-VS linear trend.

[0112] like Figure 7 As shown, further plotting the SHAP dependency diagram of the main features reveals that the residual contribution gradually shifts from negative to positive as density increases. Resistivity exhibits a significant positive lifting effect on the residual after reaching a certain threshold, while the direct impact of natural gamma on the residual is relatively limited. Interpreting these patterns allows us to link the numerical relationships within the machine learning model to geological factors such as lithological changes, fracture development, and fluid-bearing conditions, thereby enhancing the understandability and credibility of the method in engineering practice.

[0113] Through the above steps, this embodiment, while maintaining the VP-VS physical trend constraint, introduces machine learning to finely correct the residuals and utilizes SHAP to provide interpretable analysis, thus achieving an integrated technical solution for high-precision prediction of shear wave velocity in deep coal seams and analysis of geological causal relationships. The shear wave velocity prediction results described in this embodiment can be used in at least one application in coal seam mechanical parameter analysis, brittleness evaluation, wellbore stability analysis, or geological engineering evaluation.

[0114] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from the spirit and scope of this invention.

Claims

1. A method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning, characterized in that, include: S1: Obtain logging data from at least one training well and one target well. The logging data includes P-wave velocity or sonic transit time and at least two conventional logging parameters. The training well also includes measured shear wave velocity data. S2: Preprocessing of logging data, including logging curve resampling, outlier removal, noise suppression, and missing data imputation; S3: The preprocessed conventional logging parameters are divided into positive and negative indices according to their physical property correlation attributes, and then normalized respectively. S4: Based on the P-wave velocity and measured S-wave velocity of the training well, establish a physical trend relationship model between P-wave and S-wave velocities; S5: Calculate the shear wave velocity trend value at each depth of the training well based on the physical trend relationship model; S6: Subtract the measured shear wave velocity from the shear wave velocity trend value in the training well to obtain the shear wave velocity residual; S7: Using normalized conventional logging parameters as input and shear wave velocity residuals as output, construct and train a residual prediction model; S8: Substitute the target well P-wave velocity into the physical trend relationship model to obtain the target well S-wave velocity trend value, input the normalized conventional logging parameters of the target well into the residual prediction model to obtain the prediction residual, and superimpose the S-wave velocity trend value and the prediction residual to obtain the target well S-wave velocity prediction result.

2. The method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning according to claim 1, characterized in that, Step S2 includes: Resampling of well logging data at preset depth intervals; Identify and remove outlier data points that exceed statistical thresholds; Slide window filtering or median filtering can be used to suppress spike noise; Missing data points can be filled using linear interpolation, spline interpolation, or interpolation methods based on statistical models.

3. The method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning according to claim 1, characterized in that, In step S3, positive indicators are mapped to positive intervals based on their minimum and maximum values ​​in the training wells, and negative indicators are mapped to negative intervals based on their minimum and maximum values ​​in the training wells, so that the normalized values ​​are mapped to the interval between zero and one.

4. The method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning according to claim 1, characterized in that, The physical trend relationship model can be any one of a linear function model, a polynomial function model, or a piecewise function model.

5. The method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning according to claim 4, characterized in that, When the physical trend relationship model is a linear function model, the relationship between the P-wave and S-wave velocities is: ； in, This represents the trend value of the shear wave velocity. For the longitudinal wave velocity, a is the slope coefficient, b is the intercept term, and a and b are obtained by fitting the training well data using the least squares method.

6. The method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning according to claim 1, characterized in that, The formula for calculating the shear wave velocity residual in step S6 is: ； in, For the shear wave velocity residual, To measure the shear wave velocity, This represents the trend value of the shear wave velocity.

7. The method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning according to claim 1, characterized in that, The residual prediction model can be any one of the following: regression model, ensemble learning model, tree model, or neural network model.

8. The method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning according to claim 7, characterized in that, The residual prediction model is an ensemble regression model built on multiple decision trees. Step S7 also includes: adopting a leave-one-well cross-validation strategy, using single-well data as the validation set and the remaining well data as the training set, performing multiple rounds of validation on the residual prediction model, calculating the root mean square error, mean absolute error and coefficient of determination for each validation set, and selecting the model with the best overall performance.

9. The method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning according to claim 1, characterized in that, The formula for calculating the target well shear wave velocity prediction result in step S8 is as follows: ； in To predict shear wave velocity, The target well's shear wave velocity trend value, To predict residuals.

10. The method for predicting shear wave velocity in coal reservoirs based on physical trend constraints and hybrid machine learning according to claim 1, characterized in that, Conventional logging parameters include at least two of the following: density, resistivity, neutron porosity, natural gamma, borehole diameter, spontaneous potential, and depth.

Images