A cross-well shear wave velocity prediction method based on GR lithology binning alignment

Abstract

This invention discloses a cross-well shear wave velocity prediction method based on GR lithology bin alignment. Addressing the problems of large inter-well lithological differences, inconsistent engineering parameter distributions, and insufficient utilization of multi-source features in existing cross-well prediction methods, this method first collects drilling engineering parameters and logging data from multiple wells. After preprocessing, the data is divided into training and test sets using the well retention method. The optimal GR lithology bin alignment strategy is selected from the training set to achieve inter-well lithology alignment. Based on this, the absolute and relative deviation characteristics of engineering parameters are calculated, and drilling logging features such as P-wave transit time and density are integrated to construct a complete prediction feature set. Finally, the LightGBM regression algorithm is used to establish a cross-well shear wave velocity prediction model, which is optimized using an early stopping mechanism before outputting the prediction results. Practical application shows that the coefficient of determination R0 is significantly improved. 2 The accuracy reached 0.91, which is more than 18% higher than that of traditional modeling methods, and can effectively achieve high-precision and rapid prediction under the condition of no shear wave logging in new wells.

Description

Technical Field

[0001] This invention belongs to the field of petroleum engineering technology, specifically relating to a method for predicting transverse wave velocity across wells based on GR lithology box alignment. Background Technology

[0002] In oil and gas exploration and development, formation elastic parameters are crucial foundational data for formation evaluation, reservoir identification, and wellbore stability analysis. Among these, shear wave velocity (Vs) or formation shear wave transit time (DTS) reflects the elastic properties and rock mechanical characteristics of the formation, playing a vital role in engineering applications such as formation lithology identification, in-situ stress prediction, fracturing design, and seismic inversion. Therefore, obtaining accurate and reliable shear wave velocity data is of paramount importance for oil and gas well engineering design and reservoir evaluation.

[0003] Currently, shear wave velocities are typically obtained through dipole acoustic logging or full-wavetrain acoustic logging. However, these logging operations are usually performed after drilling is completed, which is not only costly but also often results in difficulties in obtaining shear wave velocity data in some wells due to wellbore conditions or operational limitations, leading to a lack of complete formation shear wave information in some wells. Therefore, predicting shear wave velocities using existing logging data or drilling process data has become an important research direction in the current oil and gas exploration and development field.

[0004] With the development of logging while drilling (LWD) technology and drilling data acquisition technology, a large amount of engineering parameter data can be acquired in real time during the drilling process, such as mechanical drilling rate, bit pressure, rotary table speed, torque, pump pressure, and standpipe pressure. Simultaneously, LWD information such as natural gamma ray, density, and P-wave transit time can also be obtained. These data, to a certain extent, reflect the interaction between the drill bit and the formation and the physical properties of the formation, and are therefore widely used in formation parameter prediction research.

[0005] In recent years, with the development of machine learning methods, researchers have gradually utilized data-driven approaches to predict formation elastic parameters. For example, the study "Estimation of shear wave velocity in an Iranian oilreservoir using machine learning methods" uses an artificial neural network model, taking logging parameters such as natural gamma, porosity, density, and P-wave velocity as inputs, to predict shear wave velocity, and validates the effectiveness of this method in terms of prediction accuracy on a large-scale dataset. Another study, "Shear Wave Velocity Prediction Based on the Long Short-Term Memory Network with Attention Mechanism," proposes a shear wave velocity prediction method based on deep learning networks (such as an Inception-Attention-BiLSTM hybrid network). By combining multi-scale feature extraction and temporal feature modeling mechanisms, it improves the accuracy and stability of shear wave velocity prediction under complex logging data conditions.

[0006] In actual oilfield development, there are often significant differences in formation and engineering parameter distribution between different wells. Therefore, cross-well prediction has become an important issue in shear wave velocity prediction research. How to make full use of multi-well data and improve the model's generalization ability across different wells through reasonable feature construction and model training methods is of great significance for achieving stable and reliable shear wave velocity prediction.

[0007] The main problems with existing technologies:

[0008] Although existing research has used well logging data or machine learning methods to predict shear wave velocity or formation shear wave transit time, and has achieved good results to some extent, there are still some shortcomings in practical cross-well prediction applications, mainly in the following aspects:

[0009] First, the formation lithology and sedimentary environment of different wells usually vary significantly, resulting in strong inconsistencies in the statistical distribution of logging curves and engineering parameters across different wells. When a unified prediction model is directly constructed using multi-well data, the lack of effective lithological constraints and the differences in data distribution between different wells can easily lead to unstable model training, thereby affecting the accuracy and generalization ability of cross-well predictions.

[0010] Secondly, existing methods typically use raw engineering parameters or logging parameters directly as model input features. However, engineering parameters are affected by various factors such as drilling technology, wellbore conditions, and equipment parameters under different formation conditions, and their absolute values ​​often vary significantly between different wells. Therefore, relying solely on raw engineering parameters for modeling makes it difficult to effectively reflect the changing characteristics of engineering parameters relative to the formation background, thus limiting the model's ability to characterize changes in formation properties.

[0011] Furthermore, in cross-well prediction tasks, some studies rely primarily on logging curves or single-type features for modeling, while failing to adequately utilize the synergistic effects between different types of features. For example, P-wave velocity and density logging parameters in drilling data can reflect formation elastic properties to some extent, but in some methods, they are not effectively integrated with engineering parameter features, resulting in limited information available to the model and affecting the stability and accuracy of the prediction results.

[0012] Therefore, in cross-well prediction scenarios, how to align data from different wells through reasonable lithological classification, construct effective features that reflect the relative changes in engineering parameters, and integrate logging-while-drilling information to improve the model's ability to predict formation shear wave time difference has become a pressing technical problem to be solved. Summary of the Invention

[0013] To overcome the shortcomings of existing shear wave velocity prediction technologies in cross-well prediction applications, such as significant differences in inter-well data distribution, poor stability of direct modeling of engineering parameters, and insufficient utilization of multi-source data features, which lead to limited model generalization ability, unstable prediction accuracy, and difficulty in adapting to the prediction needs of actual new wells, this invention aims to provide a cross-well shear wave velocity prediction method based on GR lithology bin alignment. This method effectively weakens the impact of lithological differences and engineering parameter distribution differences between different wells on model prediction performance by introducing lithological constraints and feature deviation characterization mechanisms, thereby improving the stability and accuracy of cross-well shear wave velocity prediction.

[0014] Specifically, this invention utilizes natural gamma-ray logging parameters as lithological characterization indicators. By performing GR lithological binning analysis on training data, the optimal lithological binning strategy is selected, enabling lithological alignment of data from different wells under the same lithological binning conditions. Based on this, by statistically analyzing the mean values ​​of engineering parameters within the same GR bin, absolute and relative deviation characteristics of engineering parameters are constructed, thereby characterizing the variation of engineering parameters relative to the lithological background. Furthermore, P-wave velocity and density logging characteristics from drilling data are integrated to construct a complete predictive feature set. A LightGBM regression model is used to establish a nonlinear mapping relationship between engineering parameters and formation shear wave time difference, achieving effective prediction of shear wave velocity under cross-well conditions. The technical solution adopted by this invention is:

[0015] 1. A method for predicting trans-well shear wave velocity based on GR lithology box alignment, characterized by comprising the following steps:

[0016] 1) Collect drilling engineering parameters and logging data from multiple wells to construct a cross-well basic dataset;

[0017] 2) Preprocess the basic dataset by labeling well names, removing invalid data, and dividing the training and test sets according to the well retention method;

[0018] 3) Select the optimal GR lithology binning strategy based on the training set;

[0019] 4) Based on the optimal binning strategy, perform GR bin alignment on the training set and test set respectively, calculate the engineering parameter deviation features, and fuse LWD features to form a complete prediction feature set;

[0020] 5) The LightGBM regression algorithm was used to construct a cross-well formation shear wave time difference prediction model and configure the model parameters;

[0021] 6) Train the model using training set data, learn the nonlinear mapping relationship between engineering parameters and formation shear wave time difference, and optimize the model through early stopping mechanism;

[0022] 7) Input the test set data into the trained model and output the formation shear wave time difference prediction results for each depth point.

[0023] 2. The method for predicting transverse wave velocity based on GR lithology box alignment according to claim 1, characterized in that the drilling engineering parameters in step 1) include mechanical drilling speed ROPAVG, drilling pressure WOBAVG, rotary table speed SURFRPM, torque TORQUE, pump pressure PUMP, and mechanical specific energy MSE; the logging data includes natural gamma ray GR, P-wave transit time DTC, and density logging parameters BDCFM; and the cross-well basic dataset also includes well depth data TDEP.

[0024] 3. The cross-well shear wave velocity prediction method based on GR lithology bin alignment according to claim 1, characterized in that the preprocessing of the basic dataset in step 2) includes: adding a well name identifier field to the data of different wells to distinguish the data sources of different wells, removing missing values, outliers and invalid data, and dividing the dataset according to the well retention method, that is, selecting one well as the test well and the data of the remaining wells as the training set to realize the training and evaluation of the cross-well prediction model.

[0025] 4. The method for predicting trans-well shear wave velocity based on GR lithology bin alignment according to claim 1, characterized in that step 3) of selecting the optimal GR lithology binning strategy based on the training set includes the following process: selecting natural gamma logging parameters as lithology characterization parameters in the training set data, setting a combination of 3 GR binning methods and 4 binning quantities, performing multi-group binning on the natural gamma data, wherein the binning methods include equal frequency binning, quantile binning and binning based on K-means clustering, and the number of bins is set to 2-5;

[0026] Under each binning strategy, the engineering parameters of the training set are grouped and statistically analyzed according to the binning results, the mean of each engineering parameter in the corresponding GR bin is calculated, and the absolute deviation feature and relative deviation feature of the engineering parameters are constructed.

[0027] Further, the correlation index between the deviation characteristics and the shear wave time difference of the target stratum is calculated, and the correlation index obtained under different binning strategies is statistically compared. The average correlation is used as the evaluation standard, and the GR lithology binning strategy that maximizes the correlation between the engineering parameter deviation characteristics and the shear wave time difference of the target stratum is selected as the optimal binning strategy.

[0028] 5. The method for predicting trans-well shear wave velocity based on GR lithology bin alignment according to claim 1, characterized in that, the construction of the prediction feature set according to the optimal bin alignment strategy in step 4) includes the following process: according to the optimal GR lithology bin alignment strategy, the natural gamma logging parameters in the training set and the test set are binned respectively, so that the data from different wells are aligned under the same GR lithology bin alignment conditions; within each well, the engineering parameters are statistically calculated based on the GR bin alignment category as the grouping unit to obtain the mean value of the engineering parameters within the corresponding GR bin. Where X i Let represent the i-th engineering parameter, and b represent the corresponding GR bin; construct the absolute deviation characteristics and relative deviation characteristics of the engineering parameters based on the mean, wherein the absolute deviation characteristics of the engineering parameters are:

[0029]

[0030] The relative deviation characteristics of engineering parameters are as follows:

[0031]

[0032] Where ε is a small constant to prevent the denominator from being zero; then the original engineering parameter features, engineering parameter absolute deviation features, engineering parameter relative deviation features and logging-while-drilling features are fused to form a complete prediction feature set, wherein the logging-while-drilling features include P-wave velocity DTC and density logging parameters BDCFM.

[0033] 6. The cross-well shear wave velocity prediction method based on GR lithology box alignment according to claim 1, characterized in that, in step 5), the LightGBM regression algorithm is used to construct a cross-well formation shear wave time difference prediction model, including: using the constructed complete prediction feature set as the model input feature, using the formation shear wave time difference (DTS) as the prediction target variable, establishing a LightGBM regression model based on gradient boosting decision tree, and configuring the model parameters, wherein the model objective function is set to the regression task, the evaluation index is the root mean square error (RMSE), the number of leaf nodes is set to 127, the learning rate is set to 0.005, the feature sampling ratio is set to 0.8, the sample sampling ratio is set to 0.8, the sample sampling frequency is set to 5, and the minimum number of samples per leaf node is set to 20, which is used to construct the cross-well formation shear wave time difference prediction model.

[0034] 7. The method for predicting cross-well shear wave velocity based on GR lithology box alignment according to claim 1, characterized in that the model training process in step 6) includes: randomly dividing the samples in the training set data into a training subset and a validation subset according to a preset ratio, wherein the validation subset accounts for 20% of the number of training set samples; inputting the training subset into the LightGBM regression model for iterative training, simultaneously evaluating the error of the training subset and the validation subset during the model training process, using the root mean square error (RMSE) as the evaluation index, and performing gradient boosting tree iterative training under the condition of a maximum of 2000 iterations; enabling an early stopping mechanism during the training process, automatically stopping the training when the RMSE of the validation subset does not decrease further in 100 consecutive iterations, and recording the corresponding optimal number of iterations as the final model parameters to obtain the optimal cross-well formation shear wave time difference prediction model.

[0035] 8. The method for predicting trans-well shear wave velocity based on GR lithology box alignment according to claim 1, characterized in that, in step 7), the trained LightGBM regression model is used to predict the test well data, the test set feature data is input into the trained LightGBM model, and the predicted shear wave velocity value corresponding to each sample point is calculated through the model prediction function. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the embodiments will be briefly introduced below.

[0037] Figure 1 Schematic diagram of the overall process of cross-well formation shear wave time difference prediction method;

[0038] Figure 2 Schematic diagram of the LightGBM prediction model structure;

[0039] Figure 3 A schematic diagram comparing the predicted and measured values ​​of formation shear wave transit time (DTS). Specific Implementation

[0040] Example 1

[0041] This embodiment demonstrates a cross-well formation shear wave transit time (DTS) prediction method based on lithological bin alignment and engineering parameter deviation characteristics. It has been validated in multiple development wells in the Volve oilfield. This method enables a complete process from raw drilling data processing and cross-well feature alignment to accurate DTS prediction, effectively solving the problem of poor generalization ability caused by systematic differences in engineering parameters between different well locations. See also Figure 1 This is a schematic diagram of the overall process of the prediction method described in this invention, clearly showing the execution logic of each step. This embodiment strictly follows... Figure 1 The process and the steps described in claim 1 are executed sequentially, and the specific implementation details are as follows:

[0042] 1. A method for predicting trans-well shear wave velocity based on GR lithology box alignment, characterized by comprising the following steps:

[0043] 1) Collect drilling engineering parameters and logging data from multiple wells to construct a cross-well basic dataset;

[0044] 2) Preprocess the basic dataset by labeling well names, removing invalid data, and dividing the training and test sets according to the well retention method;

[0045] 3) Select the optimal GR lithology binning strategy based on the training set;

[0046] 4) Based on the optimal binning strategy, perform GR bin alignment on the training set and test set respectively, calculate the engineering parameter deviation features, and fuse LWD features to form a complete prediction feature set;

[0047] 5) The LightGBM regression algorithm was used to construct a cross-well formation shear wave time difference prediction model and configure the model parameters;

[0048] 6) Train the model using training set data, learn the nonlinear mapping relationship between engineering parameters and formation shear wave time difference, and optimize the model through early stopping mechanism;

[0049] 7) Input the test set data into the trained model and output the formation shear wave time difference prediction results for each depth point.

[0050] In this embodiment, step 1) collects drilling engineering parameters and logging data from multiple wells to construct a cross-well basic dataset.

[0051] Three adjacent development wells (F1-A, F11-A, and F11-T2) were selected from the Volve oilfield to obtain field-acquired engineering parameters and raw logging data. The raw data was in LAS format, which was then read and converted into CSV structured table data using lasio and Pandas tools for subsequent unified processing. The data from the three wells cover the same oil and gas block, with similar formation lithology and drilling techniques. The depth ranges and data entries are shown in Table 1.

[0052] Table 1. Well Location Basic Information for Examples

[0053]

[0054] The data collected in this step is divided into two categories, and the specific parameters are shown in Tables 2 and 3:

[0055] Table 2 Drilling Engineering Parameter Information

[0056]

[0057] Table 3 Well Logging Data Information

[0058]

[0059] After integrating the parameter data from the three wells and removing duplicate fields, a cross-well basic dataset was constructed, with a total of 3,949 valid samples.

[0060] In this embodiment, step 2) involves preprocessing the basic dataset by labeling well names, removing invalid data, and dividing the training and test sets according to the well retention method.

[0061] The cross-well basic dataset constructed in step 1) is subjected to standardization preprocessing, specifically as follows:

[0062] 2.1) Mark well name identifier: Add a unique well name field to each data sample to clearly distinguish F1-A, F11-A, and F11-T2, so as to achieve accurate processing by well grouping and avoid data confusion between wells;

[0063] 2.2) Remove invalid data: Delete data segments with deep missing values, empty GR or DTS target values. For samples with less than 5% missing values ​​for a single feature, use linear interpolation to fill in the missing values. Use the Z-score method to remove extreme outliers beyond 3 standard deviations and replace the outliers with the median of the corresponding feature to ensure data continuity and validity.

[0064] 2.3) Dataset partitioning using the well retention method: A whole-well partitioning strategy is adopted, where each well is used as an independent test set, and the remaining two wells are merged as the training set, forming a total of 3 training-test combinations, as detailed below:

[0065] - Group 1: Training set = F1-A + F11-A, Test set = F11-T2

[0066] - Second group: Training set = F1-A + F11-T2, Test set = F11-A

[0067] - Group 3: Training set = F11-A + F11-T2, Test set = F1-A

[0068] This partitioning method ensures that there is no overlap between the training set and the test set, completely simulating the on-site new well prediction scenario, strictly preventing data leakage, and truly verifying the model's cross-well generalization ability.

[0069] In this embodiment, step 3) selects the optimal GR lithology binning strategy based on the training set.

[0070] GR gamma rays are used as core markers of stratigraphic lithology. This step only uses the training set data to select the optimal binning scheme and does not introduce test set information. Specific operations are as follows:

[0071] 3.1) Setting candidate binning methods: Select three commonly used lithological binning methods: quantile binning, equal frequency binning, and KMeans clustering binning;

[0072] 3.2) Set the range of boxes: Select four gradients: 2 boxes, 3 boxes, 4 boxes, and 5 boxes, and compare and test them one by one;

[0073] 3.3) Determining the optimal strategy: Calculate the average absolute correlation coefficient between the engineering parameter deviation characteristics and the DTS target value under each binning strategy, and select the combination with the largest correlation coefficient as the optimal binning strategy;

[0074] After multiple comparative verifications, the optimal GR lithology binning strategy in this embodiment is the quantile binning method + 3 bins. Under this strategy, the deviation characteristics and the average correlation coefficient of DTS reach 0.78. The lithology intervals are clearly divided after binning, and the alignment effect between wells is optimal.

[0075] In this embodiment, step 4) involves performing GR bin alignment on the training and test sets according to the optimal binning strategy, calculating the engineering parameter deviation features, and fusing LWD features to form a complete prediction feature set.

[0076] Following the optimal binning strategy determined in step 3), the training and test sets are processed synchronously, maintaining independent processing and no information exchange throughout the process. Specific details are as follows:

[0077] 4.1) GR bin alignment: For the training set and test set, 3-bin bin alignment is performed independently according to the GR data distribution of each well, dividing them into three lithological intervals: low GR, medium GR, and high GR. This achieves benchmark alignment between different well locations and lithological strata of the same type, eliminating parameter offset caused by lithological differences between wells.

[0078] 4.2) Calculate the deviation characteristics of engineering parameters: For 11 drilling engineering parameters, including ROPAVG, WOBAVG, and MSE, calculate the absolute and relative deviation characteristics respectively. The calculation formulas are as follows:

[0079] Absolute deviation = Original value of working parameter - Mean value of working parameter within the corresponding GR bin

[0080]

[0081] Add a tiny value of 10 to the formula -5 This avoids calculation errors caused by a denominator of zero; the deviation feature is used to eliminate systematic deviations in the engineering parameters of a single well while retaining effective fluctuation information related to formation response.

[0082] 4.3) Integrating LWD features: The two LWD logging features, DTC (longitudinal wave transit time) and BDCFM (well diameter), are directly spliced ​​and integrated with the original engineering parameters, GR features, and deviation features after bin alignment. After removing duplicate and redundant fields, a complete predictive feature set containing 35 dimensions is formed, providing sufficient and effective information for model training.

[0083] In this embodiment, step 5) uses the LightGBM regression algorithm to construct a cross-well formation shear wave time difference (DTS) prediction model and configures the model parameters. Specifically, the LightGBM gradient boosting regression algorithm, which is suitable for structured data modeling and has good generalization ability, is selected to construct the cross-well DTS prediction model. Based on the data characteristics and cross-well prediction task requirements of this embodiment, the key parameters of the model are reasonably set. See [link to relevant documentation]. Figure 2 , Figure 2 This is a schematic diagram of the LightGBM prediction model structure described in this invention, showing the connection relationships between the model input layer, the ensemble learning structure composed of multiple gradient boosting decision trees, the prediction output layer, and the early stopping optimization module.

[0084] Regarding model parameter settings, this embodiment sets the objective function to a regression task, using the root mean square error (RMSE) as the model evaluation metric; the number of leaf nodes is set to 127 to improve the model's nonlinear fitting ability; the learning rate is set to 0.005 to ensure stable convergence of the model during iterative training; the feature sampling ratio is set to 0.8 to randomly select some features for training in each tree construction process; the sample sampling ratio is set to 0.8, and the bagging sampling frequency is set to 5 to enhance the model's randomness and improve its generalization ability; the minimum number of samples per leaf node is set to 20 to limit excessive splitting and reduce the risk of overfitting; and the maximum number of iterations is set to 2000 to ensure that the model has sufficient training rounds. Through the above parameter configuration, the model's fitting ability is ensured while also considering its stability and generalization performance, making it suitable for small-sample data scenarios in cross-well formation shear wave time difference prediction.

[0085] In this embodiment, step 6) uses training set data to train the model, learns the nonlinear mapping relationship between engineering parameters and formation shear wave time difference, and optimizes the model through an early stopping mechanism.

[0086] Based on the model constructed in step 5) and the complete prediction feature set obtained in step 4), data leakage-free model training is carried out, with the following specific steps:

[0087] 6.1) Feature Standardization: The complete prediction feature set is grouped and standardized. The training set is grouped by well and standardized using its own feature mean and standard deviation. The test set is standardized separately using its own data statistics, as shown in the following formula, to eliminate the impact of differences in feature dimensions on model training:

[0088]

[0089] In the formula: x is the original feature value, μ is the mean of the corresponding feature, and σ is the standard deviation of the corresponding feature;

[0090] 6.2) Training set splitting: 20% of the standardized training set is randomly divided as a validation subset, which is used only for monitoring early stopping of the model and does not participate in forward training;

[0091] 6.3) Model Training and Early Stop Optimization: Input the training subset into the model to learn the nonlinear mapping relationship between engineering parameters, bias features, LWD features and DTS; set the number of early stop rounds to 100, that is, when the RMSE of the validation set does not decrease for 100 consecutive rounds, the training is stopped immediately to retain the optimal iteration weights, effectively prevent the model from overfitting and improve the stability of cross-well prediction.

[0092] In this embodiment, step 7) inputs the test set data into the trained model and outputs the formation shear wave time difference prediction results for each depth point.

[0093] Input the complete features of the standardized test set into the optimal model trained in step 6), perform inference and prediction point by point at depth, directly output the formation shear wave time difference (DTS) value at the corresponding depth, and organize them into a continuous prediction curve according to the depth sequence.

[0094] See Figure 3 This is a schematic diagram comparing the predicted and measured DTS values ​​of the present invention. F11-A test well data is selected to show the fit between the model prediction curve and the field measured DTS curve. At the same time, the prediction curve of the traditional modeling method is marked as a comparison, which intuitively reflects the accuracy advantage of the present method.

[0095] An overall performance evaluation was performed on the results of the three sets of test wells. The model's cross-well prediction average RMSE was 8.26 μS / ft, and the average coefficient of determination R was [missing value]. 2 =0.91, the prediction accuracy meets the standards for field logging interpretation and formation evaluation. Compared with the traditional direct modeling method, the cross-well generalization accuracy is improved by more than 18%, which can realize rapid and accurate DTS prediction in new wells without logging conditions, and reduce downhole operation costs.

Claims

1. A method for predicting trans-well shear wave velocity based on GR lithological box alignment, characterized in that, Includes the following steps: 1) Collect drilling engineering parameters and logging data from multiple wells to construct a cross-well basic dataset; 2) Preprocess the basic dataset by labeling well names, removing invalid data, and dividing the training and test sets according to the well retention method; 3) Select the optimal GR lithology binning strategy based on the training set; 4) Based on the optimal binning strategy, perform GR bin alignment on the training set and test set respectively, calculate the engineering parameter deviation features, and fuse LWD features to form a complete prediction feature set; 5) The LightGBM regression algorithm was used to construct a cross-well formation shear wave time difference prediction model and configure the model parameters; 6) Train the model using training set data, learn the nonlinear mapping relationship between engineering parameters and formation shear wave time difference, and optimize the model through early stopping mechanism; 7) Input the test set data into the trained model and output the formation shear wave time difference prediction results for each depth point.

2. The method for predicting trans-well shear wave velocity based on GR lithology box alignment according to claim 1, characterized in that, The drilling engineering parameters mentioned in step 1) include mechanical drilling speed ROPAVG, drilling pressure WOBAVG, rotary table speed SURFRPM, torque TORQUE, pump pressure PUMP, and mechanical specific energy MSE. The logging data includes natural gamma ray GR, P-wave transit time DTC, and density logging parameters BDCFM. The cross-well basic dataset also includes well depth data TDEP.

3. The method for predicting trans-well shear wave velocity based on GR lithology box alignment according to claim 1, characterized in that, The preprocessing of the basic dataset in step 2) includes: adding a well name identifier field to the data of different wells to distinguish the data sources of different wells, removing missing values, outliers and invalid data, and dividing the dataset according to the well retention method, that is, selecting one well as the test well and the data of the remaining wells as the training set to realize the training and evaluation of the cross-well prediction model.

4. The method for predicting trans-well shear wave velocity based on GR lithology box alignment according to claim 1, characterized in that, The optimal GR lithology binning strategy for screening based on the training set in step 3) includes the following process: Select natural gamma logging parameters as lithology characterization parameters in the training set data, set 3 combinations of GR binning methods and 4 combinations of binning numbers, and perform multi-group binning on the natural gamma data. The binning methods include equal frequency binning, quantile binning and binning based on K-means clustering, and the number of bins is set to 2-5. Under each binning strategy, the engineering parameters of the training set are grouped and statistically analyzed according to the binning results, the mean of each engineering parameter in the corresponding GR bin is calculated, and the absolute deviation feature and relative deviation feature of the engineering parameters are constructed. Further, the correlation index between the deviation characteristics and the shear wave time difference of the target stratum is calculated, and the correlation index obtained under different binning strategies is statistically compared. The average correlation is used as the evaluation standard, and the GR lithology binning strategy that maximizes the correlation between the engineering parameter deviation characteristics and the shear wave time difference of the target stratum is selected as the optimal binning strategy.

5. The method for predicting trans-well shear wave velocity based on GR lithology box alignment according to claim 1, characterized in that, The construction of the prediction feature set based on the optimal binning strategy in step 4) includes the following process: according to the optimal GR lithology binning strategy, the natural gamma logging parameters in the training set and the test set are binned respectively, so that the data of different wells can be aligned under the same GR lithology binning conditions. Within each well, the engineering parameters were statistically calculated based on the GR sub-box category, and the mean value of the engineering parameters within the corresponding GR sub-box was obtained. Where X i Let represent the i-th engineering parameter, and b represent the corresponding GR bin; construct the absolute deviation characteristics and relative deviation characteristics of the engineering parameters based on the mean, wherein the absolute deviation characteristics of the engineering parameters are: The relative deviation characteristics of engineering parameters are as follows: Where ε is a small constant to prevent the denominator from being zero; then the original engineering parameter features, engineering parameter absolute deviation features, engineering parameter relative deviation features and logging-while-drilling features are fused to form a complete prediction feature set, wherein the logging-while-drilling features include P-wave velocity DTC and density logging parameters BDCFM.

6. The method for predicting trans-well shear wave velocity based on GR lithology box alignment according to claim 1, characterized in that, Step 5) involves constructing a cross-well formation shear wave time difference prediction model using the LightGBM regression algorithm. This includes: using the constructed complete prediction feature set as the model input feature, and the formation shear wave time difference (DTS) as the prediction target variable, establishing a LightGBM regression model based on a gradient boosting decision tree, and configuring the model parameters. The model objective function is set to the regression task, the evaluation index is the root mean square error (RMSE), the number of leaf nodes is set to 127, the learning rate is set to 0.005, the feature sampling ratio is set to 0.8, the sample sampling ratio is set to 0.8, the sample sampling frequency is set to 5, and the minimum number of samples per leaf node is set to 20. This is used to construct the cross-well formation shear wave time difference prediction model.

7. The method for predicting trans-well shear wave velocity based on GR lithology box alignment according to claim 1, characterized in that, The model training process in step 6) includes: randomly dividing the samples in the training set data into a training subset and a validation subset according to a preset ratio, wherein the validation subset accounts for 20% of the number of training set samples; inputting the training subset into the LightGBM regression model for iterative training; simultaneously evaluating the errors of the training subset and the validation subset during the model training process, using the root mean square error (RMSE) as the evaluation index, and performing gradient boosting tree iterative training under the condition of a maximum of 2000 iterations; enabling an early stopping mechanism during the training process, automatically stopping the training when the RMSE of the validation subset does not decrease further in 100 consecutive iterations, and recording the corresponding optimal number of iterations as the final model parameters to obtain the optimal cross-well formation shear wave time difference prediction model.

8. The method for predicting trans-well shear wave velocity based on GR lithological box alignment according to claim 1, characterized in that, In step 7), the trained LightGBM regression model is used to predict the test well data. The test set feature data is input into the trained LightGBM model, and the predicted shear wave velocity value corresponding to each sample point is calculated through the model prediction function.

Images