High-resolution Gr prediction method based on geological constraints and XGboost algorithm

Abstract

The present invention provides a high-resolution Gr prediction method based on geological constraints and the XGboost algorithm, comprising the following steps: 1, performing fine calibration of well-seismic data; 2, extracting multiple attributes from seismic data for reservoir segments; 3, extracting seismic attributes and well logging Gr values in the time domain along the well trajectory; 4, optimizing seismic attributes based on geological constraints; 5, optimizing attributes using the XGboost algorithm; 6, based on the optimized attributes, using the optimized attributes and well logging Gr values as feature values and labels, and training a model using XGboost; and 7, predicting a three-dimensional Gr data volume. This high-resolution Gr prediction method based on geological constraints and the XGboost algorithm is unaffected by low-frequency models, has high inter-well reliability, and is more reliable than geostatistical inversion. Its application in actual work areas is not affected by the number of samples, and it has strong universality.

Description

Technical Field

[0001] The present invention relates to the technical field of oilfield development, and in particular to a high-resolution Gr prediction method based on geological constraints and an XGboost algorithm. Background Art

[0002] The prediction of three-dimensional volume of well logging Gr curve is an important task in the description of oil and gas reservoirs.

[0003] Post-stack inversion technology is widely used for 3D prediction of well logging Gr curves, with methods primarily including deterministic inversion and geostatistical inversion. Pre-stack gather data contain more information on lithology and fluids, and therefore can also be used for 3D prediction of well logging Gr curves. Prediction methods include pre-stack elastic impedance inversion and pre-stack geostatistical inversion. Both post-stack and pre-stack inversion methods are limited by the vertical resolution of the inversion technology. Post-stack deterministic inversion and pre-stack elastic impedance inversion are primarily used to analyze the planar distribution of favorable sand bodies. While post-stack geostatistical inversion and pre-stack geostatistical inversion can achieve detailed characterization and description of reservoirs, the reliability of inter-well prediction is relatively weak, and the accuracy of inter-well prediction needs to be further improved.

[0004] With the continuous development of machine learning, intelligent machine learning techniques have begun to be applied to 3D volume prediction of well logging Gr curves. The XGboost algorithm is a commonly used machine learning method. Its unique features include the model's ability to automatically utilize CPU multi-threaded parallel computing to improve computational speed, and a second-order Taylor formula expansion of the loss function for higher prediction accuracy. The XGboost algorithm, used for 3D volume prediction of well logging Gr curves, is data-driven and fully exploits the data relationships between various seismic attributes and well logging Gr curves, maximizing the utilization of seismic attributes. Furthermore, the model's predictions have higher vertical resolution than deterministic inversion and better inter-well reliability than geostatistical inversion.

[0005] However, machine learning algorithms lack automatic feature extraction and cannot effectively utilize local features of seismic data to predict reservoirs. Furthermore, using a single sampling point as input lacks the correlation between adjacent data points. To address this issue, neural network algorithms (including deep belief networks and convolutional neural networks) have begun to be applied to 3D prediction of well logging Gr curves, improving prediction accuracy. However, when sample data is limited, neural network algorithms are prone to overfitting, resulting in false predictions. Therefore, neural network algorithms are less suitable for well areas with low exploration efficiency.

[0006] For areas with limited sample data, how to obtain a more universal three-dimensional prediction method for well logging Gr curves with higher vertical resolution and stronger inter-well reliability is an urgent problem to be solved.

[0007] In the Chinese patent application with application number CN201310311751.0, a logging curve recovery method for tight gas sandstone reservoir prediction is involved. The recovery method is a method for obtaining a recovery curve; by introducing a trend field factor into the target logging curve AC and the curve to be recovered GR, and using spectrum integration technology to match and integrate the high-frequency components of the target logging curve AC and the curve to be recovered GR, a recovery curve with the dimensional characteristics of the target logging curve and retaining the high-frequency components is obtained.

[0008] In the Chinese patent application with application number: CN202110243211.8, a reservoir prediction method and device based on reservoir prediction factors are involved, and the method includes: obtaining well logging curve data, well logging interpretation results and seismic data body; constructing a reservoir prediction factor based on the well logging curve data; performing a two-parameter correlation analysis on the reservoir prediction factor and the GR logging curve in the well logging curve data to obtain the undetermined coefficient value in the reservoir prediction factor; calculating the reservoir prediction factor curve based on the well logging curve data and the undetermined coefficient value; determining the reservoir prediction factor threshold based on the reservoir prediction factor curve and the well logging interpretation results; calculating the reservoir prediction factor seismic data body based on the seismic data body and the undetermined coefficient value; and performing seismic prediction of oil and gas reservoirs in the study area based on the reservoir prediction factor threshold and the reservoir prediction factor seismic data body.

[0009] The Chinese patent application with application number CN201811403271.6 involves a reservoir property analysis method based on multi-curve joint correction. The reservoir property analysis method based on multi-curve joint correction includes: step 1, clarifying the reservoir property sensitivity curve through curve intersection; step 2, determining the impact of radioactive substances represented by Gr curves on reservoir properties; step 3, using Gr curves to correct the property sensitivity curve to form a new sensitivity curve; step 4, jointly applying all corrected property sensitivity curves to analyze the property development zone.

[0010] The above existing technologies are significantly different from the present invention and fail to solve the technical problem we want to solve. Therefore, we invented a new high-resolution Gr prediction method based on geological constraints and XGboost algorithm. Summary of the Invention

[0011] The purpose of the present invention is to provide a high-resolution Gr prediction method based on geological constraints and XGboost algorithm with high inter-well reliability and better reliability than geostatistical inversion.

[0012] The purpose of the present invention can be achieved by the following technical measures: a high-resolution Gr prediction method based on geological constraints and the XGboost algorithm, the high-resolution Gr prediction method based on geological constraints and the XGboost algorithm includes:

[0013] Step 1: Perform fine calibration of well seismic data;

[0014] Step 2: extract multiple attributes from the seismic data for the reservoir segment;

[0015] Step 3: In the time domain, extract seismic attributes and logging Gr values along the well trajectory;

[0016] Step 4: Optimize seismic attributes based on geological constraints;

[0017] Step 5: Use XGboost algorithm to optimize attributes;

[0018] Step 6: Based on the attribute optimization, the optimized attributes and the well logging Gr value are used as feature values and labels, and the model is trained using XGboost;

[0019] Step 7: Predict the Gr three-dimensional data volume.

[0020] The purpose of the present invention can also be achieved by the following technical measures:

[0021] In step 1, through fine calibration of well logging and seismic data, accurate matching and positioning of well logging and seismic data are obtained, the exact position of the well logging reservoir section on the seismic data is clarified, and the well logging Gr curve is calibrated to the time domain.

[0022] In step 2, for the target layer segment, multiple three-dimensional attributes such as amplitude, frequency, phase, single-frequency volume, and energy half-time are extracted from the seismic data. Based on the bandwidth of the seismic data, single-frequency data volumes within the bandwidth range as well as main frequency and maximum energy data volumes are generated, thereby extracting more geological body information from the seismic data.

[0023] In step 3, in the time domain, the sampling interval is 1 ms, and various seismic attribute curves and well logging Gr curves are extracted along the well trajectory.

[0024] In step 4, for the target layer segment, root mean square attributes are extracted on multiple three-dimensional attribute volumes to obtain planar prediction results of different attributes; based on geological understanding, the meaning of attribute formulas and the results of comprehensive logging interpretation, sensitive attributes are selected from the numerous planar attributes.

[0025] In step 5, the XGboost algorithm is used to optimize attributes based on the logging Gr values and seismic attributes along the well trajectory in the time domain. From the perspective of the XGboost algorithm principle and the data-driven relationship, more suitable attribute data are selected.

[0026] In step 6, the attributes selected in steps 4 and 5 are used as the feature values of the sample data, the time domain logging Gr value is used as the label, and the model is trained using XGboost.

[0027] In step 7, the trained model is applied to the actual 3D work area to predict the Gr 3D data volume.

[0028] The high-resolution Gr prediction method based on geological constraints and the XGboost algorithm in the present invention can predict high-resolution Gr data volumes. Its vertical resolution is higher than that of sparse pulse inversion, and its inter-well reliability is better than that of geostatistical inversion. Moreover, this method has strong universality for predicting three-dimensional volumes of well logging Gr curves in areas with few wells. In addition, this method incorporates geological constraints and the XGboost algorithm for attribute optimization, making the correlation between seismic attributes and well logging Gr stronger, further improving prediction accuracy. Compared with the prior art, the present invention has the following advantages:

[0029] The high-resolution Gr prediction method based on geological constraints and the XGboost algorithm described in the present invention accurately matches seismic attribute data and well logging curves through fine well-seismic calibration and combines them for application. Geological constraints and the XGboost algorithm are used to achieve the optimization of seismic attributes along the well trajectory for the well logging Gr curve. On this basis, sensitive seismic attributes along the well trajectory are used as sample eigenvalues, and the well logging Gr curve is used as the sample label. The model is trained using the XGboost algorithm, and finally the trained model is applied to the three-dimensional seismic work area to obtain the well logging GR three-dimensional data volume.

[0030] This method optimizes seismic attributes through geological constraints and the XGboost algorithm, eliminating the influence of interfering factors during Gr prediction. This strengthens the correlation between seismic attributes and well logging Gr, further improving prediction accuracy. Using the XGboost algorithm to achieve data-driven optimization between seismic attributes and well logging Gr, this method achieves a vertical resolution between seismic and well logging Gr predictions in three-dimensional Gr, achieving high-resolution Gr prediction, higher than that achieved by sparse pulse inversion. Furthermore, this method is unaffected by low-frequency models, exhibits high inter-well reliability, and is superior to geostatistical inversion. The application of this method in actual work areas is not affected by the number of samples, demonstrating strong universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flowchart of a specific embodiment of the high-resolution Gr prediction method based on geological constraints and XGboost algorithm of the present invention;

[0032] Figure 2 A seismic attribute curve diagram extracted along a well trajectory in a specific embodiment of the present invention;

[0033] Figure 3This is a graph showing the order of attribute sensitivity values for performing attribute optimization on a Gr curve using the XGboost algorithm in a specific embodiment of the present invention;

[0034] Figure 4 This is an intersection analysis diagram of 16 attributes in a specific embodiment of the present invention;

[0035] Figure 5 This is a diagram showing the importance of 10 attributes for a well logging Gr curve determined by the XGboost algorithm in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0036] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0037] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations and / or combinations thereof.

[0038] The high-resolution Gr prediction method based on geological constraints and XGboost algorithm of the present invention includes the following steps:

[0039] Step 1: Fine calibration of well logging and seismic data to determine the exact location of the logging reservoir segment on the seismic data; through fine calibration of well logging and seismic data, accurate matching and positioning of the logging reservoir segment on the seismic data are obtained, and the logging Gr (natural gamma) curve is calibrated to the time domain.

[0040] Step 2: For the reservoir segment, multiple attributes such as amplitude, frequency, phase, single frequency volume, and energy half-time are extracted from the seismic data. For the target layer segment, multiple three-dimensional attributes such as amplitude, frequency, phase, single frequency volume, and energy half-time are extracted from the seismic data, laying the data foundation for a series of subsequent operations. Based on the bandwidth of the seismic data, single frequency data volumes within the bandwidth, as well as dominant frequency and maximum energy data volumes, are generated, thereby extracting more geological information from the seismic data.

[0041] Step 3: In the time domain, seismic attributes and logging Gr values are extracted along the well trajectory. In the time domain, the sampling interval is 1 ms, and multiple seismic attribute curves and logging Gr curves are extracted along the well trajectory.

[0042] Step 4: Optimize seismic attributes based on geological constraints. For the target interval, extract root mean square attributes on multiple 3D attribute volumes to obtain planar prediction results for different attributes. Based on expert experience, geological knowledge, the meaning of attribute formulas, and comprehensive logging interpretation results, select sensitive attributes from the numerous planar attributes.

[0043] In step 5, the well logging Gr values and seismic attributes along the well trajectory are used as input data, and the XGboost algorithm is used to optimize attributes. The well logging Gr values and seismic attributes along the well trajectory in the time domain are used as input data, and the XGboost algorithm is used to optimize attributes. From the perspective of the XGboost algorithm principle and the data-driven relationship, more suitable attribute data is selected without human influence.

[0044] In step 6, based on the attribute optimization, the optimized attributes and the logging Gr value are used as the feature values and labels, and the model is trained using XGboost. The attributes optimized in steps 4 and 5 are used as the feature values of the sample data, and the time domain logging Gr value is used as the label, and the model is trained using XGboost.

[0045] Step 7: Apply the trained model to the actual 3D work area to predict the Gr 3D data volume.

[0046] The following are several specific embodiments of the present invention.

[0047] Example 1:

[0048] like Figure 1 As shown, the high-resolution Gr prediction method based on geological constraints and XGboost algorithm includes the following steps:

[0049] Step 1: Through fine calibration of well logging and seismic data, accurate matching and positioning of well logging and seismic data are obtained, the exact position of the well logging reservoir section on the seismic data is clarified, and the well logging Gr curve is calibrated to the time domain.

[0050] Step 2: For the target layer, multiple 3D attributes, such as amplitude, frequency, phase, single-frequency volume, energy half-time, kurtosis, and torsion, are extracted from the seismic data. This provides the data foundation for subsequent operations. Based on the seismic data bandwidth, single-frequency data volumes, as well as dominant frequency and maximum energy data volumes, are generated at 10-Hz intervals within the bandwidth, thereby extracting more geological information from the seismic data.

[0051] Step 3: In the time domain, the sampling interval is 1ms, and multiple seismic attribute curves and logging Gr curves are extracted along the well trajectory, such as Figure 2 shown. Figure 2This is a list of seismic attribute curves of the pure upper 123 layer group extracted along the well trajectory of a certain well. From left to right, they are kurtosis attribute, weighted frequency attribute, variance attribute, slope attribute, energy half-time attribute, arc length attribute, 10 Hz single frequency attribute, pseudo entropy attribute and seismic trace attribute.

[0052] In step 4, for the target layer segment, root mean square attributes are extracted on multiple 3D attribute volumes to obtain planar prediction results for different attributes. Based on expert experience, geological knowledge, the meaning of attribute formulas, and comprehensive logging interpretation results, sensitive attributes are selected from the numerous planar attributes.

[0053] Step 5: Take the logging Gr value and seismic attributes along the well trajectory in the time domain as input data and use the XGboost algorithm to perform attribute optimization, as shown in the following example: Figure 3 shown. Figure 3 The XGboost algorithm is used to perform attribute optimization on the Gr curves of seven wells to sort the attribute sensitivity values. Figure 3 There are eight columns in total. Column 1 indicates the attribute name. Columns 2-7 indicate the six machine learning algorithms used: decision tree, extreme tree, extreme gradient boosting (XGboost), random forest, adaptive boosting, and gradient boosting. Column 8 shows the sensitivity of each attribute to the well logging Gr value. These six algorithms optimize 23 seismic attributes and provide sensitivity values for each attribute. Based on the average sensitivity value of each attribute, the top four sensitive attributes are kurtosis, weighted frequency, variance, and skew.

[0054] In step 6, the attributes selected in steps 4 and 5 are used as the feature values of the sample data, the time domain logging Gr value is used as the label, and the model is trained using XGboost.

[0055] Step 7: Apply the trained model to the actual 3D work area to predict the Gr 3D data volume.

[0056] Example 2:

[0057] In the specific embodiment 2 of the present invention, the following steps are specifically included:

[0058] Step 1: Through fine calibration of well logging and seismic data, accurate matching and positioning of well logging and seismic data are obtained, the exact position of the logging reservoir section on the seismic data is clarified, and the logging Gr (natural gamma) curve is calibrated to the time domain.

[0059] Step 2: For the target layer, multiple 3D attributes such as amplitude, frequency, phase, single-frequency volume, and energy half-time are extracted from the seismic data, laying the data foundation for a series of subsequent operations. Based on the seismic data bandwidth, single-frequency data volumes within the bandwidth, as well as dominant frequency and maximum energy data volumes, are generated, thereby extracting more geological information from the seismic data.

[0060] Step 3: In the time domain, the sampling interval is 1 ms, and multiple seismic attribute curves and well logging Gr curves are extracted along the well trajectory.

[0061] Step 4: Based on geological constraints, perform seismic attribute optimization. For the target layer, extract the root mean square attributes on multiple 3D attribute bodies to obtain plane prediction results of different attributes. Based on expert experience, geological knowledge, the meaning of attribute formulas and comprehensive logging interpretation results, select sensitive attributes from the numerous plane attributes, such as Figure 4 shown. Figure 4 It is an intersection analysis diagram of 16 attributes. Both the vertical and horizontal axes represent 16 attributes. Starting from the upper left corner, the attributes are: seismic data volume, 10hz, 20hz, 30hz, 40hz, 50hz, 60hz, 70hz, 80hz, amplitude, arc length, frequency, instantaneous main frequency, phase, main frequency amplitude, and main frequency, a total of 16 attributes. Figure 4 From the above analysis, the seismic data volume has a strong linear correlation with the 30Hz and 40Hz single-frequency volumes, so the 30Hz and 40Hz single-frequency volumes were eliminated, while the seismic data volume was retained. The 50Hz single-frequency volume has a strong correlation with the 60Hz, 70Hz, and 80Hz single-frequency volumes, so the 60Hz, 70Hz, and 80Hz single-frequency volumes were eliminated, while the 50Hz single-frequency volume was retained. The instantaneous main frequency attribute has a strong correlation with the frequency attribute among the three instantaneous attributes, so the frequency attribute was eliminated, while the instantaneous main frequency attribute was retained. Finally, through geological understanding and the intersection analysis between attributes, 10 attributes were selected for subsequent model training.

[0062] In step 5, the well logging Gr values and seismic attributes along the well trajectory are used as input data, and the XGboost algorithm is used to optimize attributes. The well logging Gr values and seismic attributes along the well trajectory in the time domain are used as input data, and the XGboost algorithm is used to optimize attributes. From the perspective of the XGboost algorithm principle and the data-driven relationship, more suitable attribute data is selected without human influence.

[0063] In step 6, based on the attribute optimization, the optimized attributes and the logging Gr value are used as the feature values and labels, and the model is trained using XGboost. The attributes optimized in steps 4 and 5 are used as the feature values of the sample data, and the time domain logging Gr value is used as the label, and the model is trained using XGboost.

[0064] Step 7: Apply the trained model to the actual 3D work area to predict the Gr 3D data volume.

[0065] Example 3:

[0066] In the specific embodiment 3 of the present invention, the following steps are specifically included:

[0067] Step 1: Through fine calibration of well logging and seismic data, accurate matching and positioning of well logging and seismic data are obtained, the exact position of the logging reservoir section on the seismic data is clarified, and the logging Gr (natural gamma) curve is calibrated to the time domain.

[0068] Step 2: For the target layer, multiple 3D attributes such as amplitude, frequency, phase, single-frequency volume, and energy half-time are extracted from the seismic data, laying the data foundation for a series of subsequent operations. Based on the seismic data bandwidth, single-frequency data volumes within the bandwidth, as well as dominant frequency and maximum energy data volumes, are generated, thereby extracting more geological information from the seismic data.

[0069] Step 3: In the time domain, seismic attributes and logging Gr values are extracted along the well trajectory. In the time domain, the sampling interval is 1 ms, and multiple seismic attribute curves and logging Gr curves are extracted along the well trajectory.

[0070] Step 4: Optimize seismic attributes based on geological constraints. For the target interval, extract root mean square attributes on multiple 3D attribute volumes to obtain planar prediction results for different attributes. Based on expert experience, geological knowledge, the meaning of attribute formulas, and comprehensive logging interpretation results, select sensitive attributes from the numerous planar attributes.

[0071] Step 5: Take the logging Gr values and seismic attributes along the well trajectory as input data, and use the XGboost algorithm to optimize attributes. Take the logging Gr values and seismic attributes along the well trajectory in the time domain as input data, and use the XGboost algorithm to optimize attributes. From the perspective of the XGboost algorithm principle and the data-driven relationship, it is not affected by human factors and selects more suitable attribute data, such as Figure 5 shown. Figure 5 This is the XGboost algorithm's judgment of the importance of 10 attributes for the logging Gr curve. The horizontal axis is different attributes, and the vertical axis is the attribute importance. The larger the value, the more important the attribute. Figure 5 From the above results, we can see that the arc length attribute is the most important for the Gr curve, followed by the dominant frequency attribute, amplitude attribute, 20 Hz single frequency body and instantaneous dominant frequency attribute; seismic data is the least important for the Gr curve, so seismic data is removed from the input attribute dataset, and finally 9 attributes are selected for subsequent model prediction.

[0072] In step 6, based on the attribute optimization, the optimized attributes and the logging Gr value are used as the feature values and labels, and the model is trained using XGboost. The attributes optimized in steps 4 and 5 are used as the feature values of the sample data, and the time domain logging Gr value is used as the label, and the model is trained using XGboost.

[0073] Step 7: Apply the trained model to the actual 3D work area to predict the Gr 3D data volume.

[0074] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art may modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features therein. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

[0075] Except for the technical features described in the specification, all other technical features are known technologies to those skilled in the art.

Claims

1. A high-resolution Gr prediction method based on geological constraints and XGboost algorithm, characterized by: The high-resolution Gr prediction method based on geological constraints and the XGboost algorithm includes: Step 1: Perform fine calibration of well seismic data; Step 2: for the reservoir segment, extract multiple seismic attributes from the seismic data; Step 3: In the time domain, extract seismic attributes and logging Gr values along the well trajectory; Step 4: Optimize the seismic attributes from step 2 based on geological constraints; Step 5: Use the XGboost algorithm to optimize the earthquake attributes in step 3; Step 6: Based on the attribute optimization in steps 4 and 5, the optimized attributes and the well logging Gr value are used as feature values and labels, and the model is trained using XGboost; Step 7: Based on the model trained in step 6, predict the Gr three-dimensional data volume; In step 1, through fine calibration of well logging and seismic data, accurate matching and positioning of well logging and seismic data are obtained, the exact position of the well logging reservoir section on the seismic data is determined, and the well logging Gr curve is calibrated to the time domain; In step 2, for the target layer, multiple three-dimensional attributes such as amplitude, frequency, phase, single frequency volume, and energy half-time are extracted from the seismic data. Based on the bandwidth of the seismic data, single frequency data volumes within the bandwidth range as well as main frequency and maximum energy data volumes are generated, thereby extracting more geological information from the seismic data. In step 3, in the time domain, the sampling interval is 1 ms, and multiple seismic attribute curves and well logging Gr curves are extracted along the well trajectory; In step 4, for the target interval, root mean square attributes are extracted from multiple 3D attribute volumes to obtain planar prediction results for different attributes. Based on geological knowledge, the meaning of attribute formulas, and comprehensive logging interpretation results, sensitive attributes are selected from the numerous planar attributes. In step 5, the XGboost algorithm is used to optimize attributes based on the logging Gr values and seismic attributes along the well trajectory in the time domain. From the perspective of the XGboost algorithm principle and the data-driven relationship, more suitable attribute data are selected; In step 6, the attributes selected in steps 4 and 5 are used as the feature values of the sample data, the time domain logging Gr value is used as the label, and the model is trained using XGboost; In step 7, the trained model is applied to the actual 3D work area to predict the Gr 3D data volume.

Images