Phase control model construction method for geostatistical inversion

Abstract

The invention relates to a geostatistical inversion phase control model construction method, and belongs to the field of oil-gas exploration and development. Single seismic attributes are permutated and combined on the basis of a multi-attribute recursive optimization method according to the known sand-to-ground ratio of a single well and the corresponding seismic attributes through the obtained logging data and seismic data, a combination of multiple seismic attributes is obtained, a machine learning model is established, and according to the sand-to-ground ratio of the known well, the sand-to-ground ratio of the known well is calculated. Inputting the multiple seismic attribute pairs of the corresponding wells into a machine learning model in sequence for training and verification, selecting an optimal seismic attribute combination of sand-to-ground ratio correlation based on a training error and a verification error, and inputting attribute values in the optimal seismic attribute combination in the target area into the trained machine learning model, according to the method, the corresponding sand-to-ground ratio is obtained, a phase control plane graph of the area is formed according to the output sand-to-ground ratio, geostatistical inversion is carried out by using the phase control plane graph, the precision of an inversion result is improved, and the multiplicity of solutions is reduced.

Description

Technical Field

[0001] The invention relates to a method for constructing a phase-controlled model of geostatistical inversion, and belongs to the field of oil and gas exploration and development. Background Art

[0002] Geostatistical inversion is an inversion technology developed with the prediction of thin-layer reservoirs. It introduces the idea of ​​random simulation into seismic inversion, and uses the variogram, probability density function and other logging and geological knowledge of sedimentary sequences, lithology and physical properties to carry out random simulation and seismic forward modeling, and with actual seismic constraints, obtain high-resolution lithology and high-resolution elastic bodies, and then obtain porosity and permeability bodies, thereby realizing the prediction of thin-layer reservoirs. However, geostatistical inversion is a model-based inversion, which is constrained and controlled by the model and is more suitable for seismic work areas with small well spacing and relatively uniform distribution. When the well network on the plane is sparse and unevenly distributed, the reliability of the model is reduced, the inversion results are highly random, and the deviation from the actual geological body is large.

[0003] The invention patent document with publication number CN113406702A discloses a geostatistical inversion method and device based on phase-constrained variable grid. The method generates a variable grid according to the seismic phase data of the target area, determines multiple target points from the variable grid according to preset data and extracts virtual wells, performs normal transformation on the actual wells and virtual wells to be processed by Gaussian simulation, and obtains normal transformation result data; constructs a Gaussian distribution based on the mean and variance of the known attribute values ​​within the preset distance range of each target point and obtains the cumulative probability distribution of the Gaussian distribution; simulates the realization of reservoir parameters of unknown points within the preset distance range based on the cumulative probability distribution, and takes the realization that best matches the seismic data as the geostatistical inversion result. The stability of the results obtained by the inversion method is limited and the results have multiple solutions. Summary of the invention

[0004] The purpose of the present invention is to provide a method for constructing a phase-controlled model of geostatistical inversion, so as to solve the problem that the results obtained by the existing geostatistical inversion method are of low accuracy and have multiple solutions.

[0005] To achieve the above object, the solution of the present invention includes:

[0006] A method for constructing a phase-controlled model for geostatistical inversion of the present invention comprises the following steps:

[0007] 1) Obtain well logging information and seismic data information in the target area, obtain the single-well sand-to-ground ratio of the target layer based on the well logging information, and obtain the seismic attributes of the corresponding layer segment based on the seismic data information;

[0008] 2) According to the known single well sand-to-ground ratio and the corresponding seismic attributes, the single seismic attributes are arranged and combined based on the multi-attribute recursive optimization method to obtain a combination of multiple seismic attributes;

[0009] 3) Establish a machine learning model, input multiple seismic attribute pairs of the corresponding wells into the machine learning model in sequence for training and verification according to the sand-to-ground ratio of the known wells, and select the best seismic attribute combination with sand-to-ground ratio correlation based on the training error and verification error;

[0010] 4) Input the attribute values ​​of the best seismic attribute combination in the target area into the trained machine learning model to obtain the corresponding sand-to-ground ratio, form a phase-controlled plane map of the area based on the output sand-to-ground ratio, and use the phase-controlled plane map to perform geostatistical inversion.

[0011] Beneficial effects: The method for constructing a phase-controlled model for geostatistical inversion of the present invention mainly establishes the relationship between seismic attributes and sand-to-ground ratio by means of a recursive optimization method and a machine learning model, obtains a variety of attribute combinations related to the sand-to-ground ratio by means of a multi-attribute recursive optimization method, and sequentially inputs a variety of attributes into a machine learning model for training and verification, thereby obtaining an attribute combination with the highest correlation with the sand-to-ground ratio, and uses the attribute combination to predict the sand-to-ground ratio of the entire target area, and to form a phase-controlled map, which is then used to perform geostatistical inversion. The method gradually integrates seismic data from well points to wells, improves the accuracy of inversion results, reduces multi-solutions, provides a basis for sedimentary phase research, and provides supplements and references for the method for constructing a phase-controlled model for geostatistical inversion of clastic reservoirs. At the same time, the present invention is simple and practical, highly operable, and has a good promotion effect.

[0012] Furthermore, the steps of the multi-attribute recursive optimization method in step 2) are as follows: among the N seismic attributes in the target area, select the attribute with the highest correlation with the sand-to-land ratio as the optimal attribute; among the remaining N-1 single seismic attributes, select the attribute with the highest correlation with the sand-to-land ratio, and form a first attribute combination with the optimal single attribute; from the remaining N-2 seismic attributes, select the attribute with the highest correlation with the sand-to-land ratio, and add it to the first attribute combination, and so on, until the last seismic attribute is added to the first attribute combination, thereby forming multiple attribute combinations in turn.

[0013] Beneficial effect: In the multi-attribute recursive optimization method of the present invention, all seismic attributes in the target area are used, and the attribute with the highest correlation with the sand-to-land ratio is taken as the optimal attribute. The attribute with the highest correlation with the sand-to-land ratio is selected from the remaining N-1 single seismic attributes, and is combined with the optimal single attribute to form a first attribute combination. The remaining seismic attributes in the target area with the highest correlation with the sand-to-land ratio are added to the first attribute combination one by one, and multiple attribute combinations are formed in turn.

[0014] Furthermore, the method further includes fusing the phase control plane map obtained in step 4) with the phase control map obtained according to the logging information to obtain a fused phase control map, and performing geostatistical inversion using the fused phase control map.

[0015] Beneficial effects: The phase control map obtained by the method of the present invention can be fused with the phase control map obtained from logging information, and the fused phase control map can be used for geostatistical inversion. This method can reduce multi-solution and improve inversion accuracy.

[0016] Furthermore, the step 3) also includes a step of correcting the output results of the machine learning model based on a plane space interpolation method.

[0017] Beneficial effect: The present invention also includes the step of correcting the output results of the machine learning model, and the correction is performed through a plane space interpolation method, thereby improving the accuracy of the machine learning model and making the output results more accurate.

[0018] Furthermore, the methods of plane space interpolation include the variogram method and the Kriging interpolation method.

[0019] Furthermore, a neural network model is used in the machine learning model.

[0020] Beneficial effects: The neural network model is used in the machine learning model. The neural network model is based on a nonlinear algorithm and the results obtained are more accurate.

[0021] Furthermore, the neural network model adopts PNN or RBF. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flow chart of the "three-step method" phase-controlled model construction method of geostatistical inversion in the embodiment of the method of the present invention;

[0023] Figure 2 It is a flow chart of the “three-step method” phase-controlled model construction method in the method embodiment of the present invention;

[0024] Figure 3 is a well area plane phase control diagram in an embodiment of the method of the present invention;

[0025] Figure 4 It is a sand body thickness map of the comparative well area geology delineation in the embodiment of the method of the present invention;

[0026] Figure 5 is an inversion plane map obtained by using a conventional geostatistical inversion method in an embodiment of the method of the present invention;

[0027] Figure 6 is an inversion plane map obtained by the geostatistical inversion method in the embodiment of the method of the present invention;

[0028] Figure 7 is a multi-attribute error analysis diagram in an embodiment of the method of the present invention;

[0029] Figure 8 is a property list diagram of multiple attributes in an embodiment of the method of the present invention;

[0030] Fig. 9 It is a sand-to-land ratio plane diagram obtained by combining attributes in an embodiment of the method of the present invention. DETAILED DESCRIPTION

[0031] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0032] Embodiment of the method for constructing a phase-controlled model for geostatistical inversion:

[0033] like Figure 1 The method for constructing a phase-controlled model for geostatistical inversion is shown. By collecting all the completed well information in the target block, the sand-to-ground ratio of the target layer is counted, and the seismic attributes of the corresponding layer segment are extracted; the multi-attribute recursive optimization method is used to fit the initial plane map in the time domain, and the multi-attribute fitting plane results are quality controlled based on the training error and the verification error; the multi-attribute results are used for neural network calculation, and the nonlinear method is used to improve the well-seismic correlation, and the phase-controlled plane map is optimized again. Based on the training error and the verification error, the neural network algorithm is selected to quality control the plane results; the variogram statistics are carried out on the neural network calculation results for the final inter-well plane phase map trend correction, and the variogram correction results are used for geostatistical inversion phase control constraints to improve the inversion results. The specific implementation methods are as follows:

[0034] 1) Obtain well logging information and seismic data information in the target area, obtain the single-well sand-to-ground ratio of the target layer based on the well logging information, and obtain the seismic attributes of the corresponding layer segment based on the seismic data information.

[0035] Specifically, the sand-to-ground ratio of the well points to be logged in the target area is obtained from the logging information, and the sand body phase to be inverted is determined according to the sedimentary background in the target area. The seismic attributes of each formation segment are obtained according to the seismic data information.

[0036] For example, in this embodiment, by analyzing the area (the plane area of ​​the target area is 866km 2 )85 wells were drilled, and the average sand-to-land ratio was 0.288, which clearly showed that the sand bodies were relatively enriched in the northwest, while the sand bodies were underdeveloped in the south, and the well points on the plane had strong heterogeneity.

[0037] 2) According to the known single well sand-to-ground ratio and the corresponding seismic attributes, the single seismic attributes are arranged and combined based on the multi-attribute recursive optimization method to obtain a combination of multiple seismic attributes.

[0038] The attributes in this embodiment are derived from the seismic data body. The seismic layer constrains the extraction time window of the attribute, and the maximum value, minimum value, and root mean square value of the earthquake amplitude, frequency, and phase are statistically calculated. Amplitude, frequency, and phase are three types of attributes that are commonly used to qualitatively describe sandstone thickness.

[0039] Firstly, a multi-attribute recursive optimization method is used to fit the initial plane map in the time domain (initial plane features of sand-land ratio), and the multi-attribute initial fitting plane results are quality controlled based on the training error and validation error.

[0040] Specifically, the steps of fitting the planar features of the sand-to-land ratio based on the multi-attribute recursive optimization method include: selecting a seismic attribute related to the sand-to-land ratio, selecting an attribute with the highest correlation with the sand-to-land ratio from the N seismic attributes in the target area as the optimal attribute, selecting an attribute with the highest correlation with the sand-to-land ratio from the remaining N-1 single seismic attributes, and forming a first attribute combination with the optimal single attribute, selecting the attribute with the highest correlation with the sand-to-land ratio from the remaining N-2 seismic attributes, and adding it to the first attribute combination, and so on, until the last seismic attribute is added to the first attribute combination, and forming multiple attribute combinations in sequence. For example, 15 attributes are selected (for example, a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, where the 15 single attributes are all attributes of the seismic sand-to-ground ratio), and the attribute with the highest correlation with the sand-to-ground ratio of the well point to be measured is selected as the optimal single attribute, for example, g is selected as the optimal attribute; the attribute with the highest correlation with the sand-to-ground ratio of the well point (for example, e) is selected from the remaining 14 attributes, and it is combined with the optimal single attribute to form the first attribute combination (ge); the attribute with the highest correlation with the sand-to-ground ratio of the well point is selected from the remaining 13 single attributes, and it is combined with the first attribute to form the second attribute combination, and so on, according to the size of the correlation, the selected attributes are sorted in turn to form an attribute combination. Based on multiple attributes, the optimal attribute combination is selected, and the sand-to-ground ratio plane features are fitted according to the optimal attribute combination. This embodiment adopts a recursive optimization method for more than 15 attributes such as the maximum frequency attribute and the minimum amplitude attribute to fit the sand-to-ground ratio plane.

[0041] Specifically Figure 8 Taking the example shown above, based on the multi-attribute recursive optimization method, the sand-to-ground ratio of the CH912 section of the well point is compared. The top ten seismic attribute correlations are, from large to small, the weighted maximum amplitude, the absolute maximum amplitude, the absolute root mean square amplitude, the apparent polarity root mean square, the minimum amplitude envelope, the weighted root mean square amplitude, the maximum amplitude, the maximum frequency, the amplitude envelope root mean square, and the frequency root mean square. According to the above-mentioned multi-attribute recursive optimization method, multiple attribute combinations can be obtained respectively (for example, the first attribute combination: {weighted maximum amplitude, maximum absolute amplitude, absolute root mean square amplitude}, the second attribute combination: {weighted maximum amplitude, maximum absolute amplitude}, etc.). According to Figure 7The quality control measures shown in the error analysis diagram indicate that the first eight attribute combinations are the optimal attribute combinations for describing the CH912 sand layer, namely, the optimal attribute combination: {weighted maximum amplitude, maximum absolute amplitude, root mean square absolute amplitude, root mean square apparent polarity, minimum amplitude envelope, weighted root mean square amplitude, maximum amplitude}, so the well-seismic correlation is increased from 25.6% to 68.4%.

[0042] The above embodiment is an embodiment of the optimal attribute combination obtained only for the well point CH912 segment. As other embodiments, the correlation between the seismic attributes and the sand-to-ground ratio of different well points is different, so the optimal attribute combination of the well points is different, which can be obtained according to the above multi-attribute recursive optimization method.

[0043] Based on the training error and validation error, the plane attribute result with the highest correlation with well seismic is selected, such as Figure 7 The black line in the multi-attribute error analysis is the training error, the red line is the validation error, the ordinate is the error, the smaller the value, the smaller the error, the higher the correlation, and the abscissa is the number of attributes involved in the calculation. Figure 8 The attribute list shown in the figure shows that the first column of the multi-attribute list is the sorting, the second column is the prediction object, the third column is the attribute name involved in the calculation, the fourth column is the training error value, the fifth column is the verification error value, and the sixth column is the correlation coefficient. The following three columns can be ignored for the time being. According to the verification error in the fifth column, when the sorting of the attribute pair is 11, the verification error value is the smallest, indicating that 11 attributes are involved in the calculation, and the sand-to-land ratio plane map obtained has the highest correlation with the sand-to-land ratio value of the well point. On the basis of 11 attributes, continue to add attributes to participate in the calculation, and the verification error value gradually increases, that is, data overflow and over-training occur, indicating that more attributes cannot optimize the results. When the single attribute in the attribute pair is less than 11 attributes, the verification error does not have an inflection point, that is, insufficient training, indicating that this is not the optimal attribute combination.

[0044] Taking the CH813 formation as an example, according to the multi-attribute error analysis table, the attribute combination corresponding to the minimum verification error is selected, that is, 12, indicating that the sand-to-ground ratio plane obtained by the 1-12 attribute combination is relatively consistent with the well point. The attribute plane results are as follows: Fig. 9 As shown. Well-seismic correlation refers to the consistency of multi-attribute results when using single well data and seismic data to describe the same geological body. For example, there are sample data of 100 wells in the region, and the well logging curves on the target layer all show sandstone, and the corresponding 80 sample seismic data show strong amplitude. Therefore, the correlation of using seismic amplitude to predict single well lithology is 80%.

[0045] 3) Establish a machine learning model. According to the known sand-to-ground ratio of the well, multiple seismic attribute pairs of the corresponding well are sequentially input into the machine learning model for training and verification. Based on the training error and verification error, the best seismic attribute combination with sand-to-ground ratio correlation is selected.

[0046] The neural network algorithms used in this embodiment are PNN (probabilistic neural network) and RBF (radial basis function), where PNN (probabilistic neural network) is a branch of radial basis network, including 4 layers, namely input layer, hidden layer, summation layer and output layer. The feature vector input by the input layer is the plane feature obtained by fitting multiple attribute recursive optimization methods; the hidden layer is connected to the input layer through the connection weight, and the matching degree between the input feature vector and each pattern in the training set, that is, the similarity, is calculated, and the distance is sent to the Gaussian function to obtain the output of the pattern layer. The number of neurons in the hidden layer is the number of input sample vectors, that is, the number of samples is the number of neurons in this layer; the summation layer is responsible for connecting the pattern layer units of each class, and the number of neurons in this layer is the number of sample categories; the output layer outputs the sample category with the highest score in the summation layer.

[0047] RBF (Radial Basis Function) is a basis function in a function space, and these basis functions are all radial functions. A radial function must satisfy one condition: for a fixed point, the function value is the same for the independent variable x that is equidistant around the fixed point. For example, the Gaussian function is one of the radial functions.

[0048] The main purpose of using PNN (probabilistic neural network) and RBF (radial basis function) is to improve the correlation between the predicted sand-to-ground ratio and the actual statistical sand-to-ground ratio of the well point, and try to make the predicted sand-to-ground ratio closer to the actual statistical value. The nonlinear fitting method is used to change the well-seismic correlation, optimize the phase control plane results, and select the optimal neural network algorithm based on the training error and verification error to quality control the plane results. Based on the multi-attribute results, this embodiment selects the algorithm with the highest correlation from the two neural network algorithms of probabilistic neural network and radial basis function, and then quality controls the plane results.

[0049] When both the training error and the validation error are relatively small, it indicates that the attribute combination is optimal, and the validation error has a higher priority than the training error, that is, the validation error value is given priority. If the training error and the validation error are not used, the calculation result will be under-trained or over-trained, resulting in insufficient training or data overflow.

[0050] Since the original well-seismic correlation is good, the correlation does not change significantly after the neural network, the plane transition is more natural, and the influence of some faults on the sand-to-ground ratio is removed.

[0051] Finally, the variogram statistics are carried out on the neural network calculation results. Based on the variogram, reasonable parameters such as range, azimuth, base value, etc. are selected for the final trend correction of the inter-well plane phase diagram.

[0052] Among them, the variogram method is a method for calculating interpolation weights based on the distance and difference between data points. This method is usually used to process irregularly distributed data points, such as geophysical data, drilling data, etc. The advantage of the variogram method is that it can reflect the spatial relationship and change trend between data points, but the parameters of the variogram need to be adjusted to obtain the best interpolation effect. As another implementation method, the Kriging interpolation method can also be used to calculate the interpolation weights in this embodiment, and the choice depends on the distribution of data points and the purpose of interpolation. In practical applications, the best method can be selected by comparing the interpolation results of different methods. This embodiment uses a variogram to correct the results of the optimal neural network output. The formula of the variogram is:

[0053]

[0054] Where h is the lag distance, Z(u) is the regionalized variable, and Z(u+h) is the regionalized variable after lag h. The function is described by parameters such as range, sill value, azimuth, and function type. The range can generally reflect the average scale of the carrier of the regionalized variable (such as sandstone), and the extension length of the sand body in a certain direction can be predicted using the range. The sill value can reflect the amplitude of the change of the variable. The azimuth (nugget value) can reflect the continuity of the calculation result. Therefore, by setting reasonable values ​​of parameters such as range, azimuth, and sill value, the plane results output by the neural network are corrected.

[0055] 4) Input the attribute values ​​of the best seismic attribute combination in the target area into the trained machine learning model to obtain the corresponding sand-to-ground ratio, form a phase-controlled plane map of the area based on the output sand-to-ground ratio, and use the phase-controlled plane map to perform geostatistical inversion.

[0056] In order to reduce the multi-solution and further improve the inversion accuracy, this embodiment also includes fusing the phase control plane map obtained in step 4) with the phase control map obtained according to the logging information to obtain a fused phase control map, and using the fused phase control map to perform geostatistical inversion.

[0057] In this embodiment, the correction result of the variogram is used to constrain the geostatistical inversion. Figure 2 As shown, the sand body phase and logging information obtained from the deterministic inversion of the sedimentary phase in the target area are respectively integrated into the corrected plane map to obtain the final phase-controlled plane map. The final phase-controlled plane map is used to constrain the geostatistical inversion, thereby improving the inversion result.

[0058] like Figure 3 The color scale range of the plane phase control map shown is 0-1. The larger the value, the more developed the sand body is, corresponding to the warm color area on the map. The smaller the value, the less developed the sand body is, corresponding to the cold color area on the map. The black dots correspond to the actual well point locations. From the phase control map, it can be seen that the sand body is mainly developed in the central and northwest areas of the work area, and is distributed along the northeast-southwest direction. Figure 4 The figure shows the sand body thickness map drawn by geologists based on the actual drilling logging curve. The yellow area indicates that the sand body is thick and well developed, and the white area indicates that the sand body is not developed. From the figure, the sand body gathers in the middle and gradually thins from southwest to northeast. Figure 3 and Figure 4 After comparison, it is found that the two have certain similarities. Therefore, the phase control map formed by the method in this embodiment is consistent with geological knowledge and can be used for subsequent seismic inversion work. The phase control map shows that the main sedimentary body used for seismic inversion is prominent, the well point is consistent with the distribution law of sand-to-land ratio, and the river boundary is gradually clear.

[0059] The addition of plane phase control makes up for the problem of insufficient constraints on plane lithology trends caused by the lack of low-frequency seismic data. Figure 5 , Figure 6 The following are the time thickness planes of sandstone predicted by conventional geostatistical inversion method and this method. By comparison, it can be seen that the lithology distribution results obtained by this method are consistent with the sedimentary trends such as source direction, catchment area orientation, and main sedimentary area, and can also depict the details of lithology changes and present the heterogeneity of the reservoir.

Claims

1. A method for constructing a phase-controlled model for geostatistical inversion, characterized in that: The following steps are involved: 1) Obtain well logging information and seismic data information in the target area, obtain the single-well sand-to-ground ratio of the target layer based on the well logging information, and obtain the seismic attributes of the corresponding layer segment based on the seismic data information; 2) According to the known single well sand-to-ground ratio and the corresponding seismic attributes, the single seismic attributes are arranged and combined based on the multi-attribute recursive optimization method to obtain a combination of multiple seismic attributes; 3) Establish a machine learning model, input multiple seismic attribute pairs of the corresponding wells into the machine learning model in sequence for training and verification according to the sand-to-ground ratio of the known wells, and select the best seismic attribute combination with sand-to-ground ratio correlation based on the training error and verification error; 4) Input the attribute values ​​of the best seismic attribute combination in the target area into the trained machine learning model to obtain the corresponding sand-to-ground ratio, form a phase-controlled plane map of the area based on the output sand-to-ground ratio, and use the phase-controlled plane map to perform geostatistical inversion.

2. The method for constructing a phase-controlled model for geostatistical inversion according to claim 1, characterized in that: The steps of the multi-attribute recursive optimization method in step 2) are as follows: among the N seismic attributes in the target area, select the attribute with the highest correlation with the sand-to-land ratio as the optimal attribute; among the remaining N-1 single seismic attributes, select the attribute with the highest correlation with the sand-to-land ratio, and form a first attribute combination with the optimal single attribute; from the remaining N-2 seismic attributes, select the attribute with the highest correlation with the sand-to-land ratio, and add it to the first attribute combination, and so on, until the last seismic attribute is added to the first attribute combination, and multiple attribute combinations are formed in sequence.

3. The method for constructing a phase-controlled model for geostatistical inversion according to claim 1, characterized in that: The method further includes fusing the phase control plane map obtained in step 4) with the phase control map obtained according to the logging information to obtain a fused phase control map, and performing geostatistical inversion using the fused phase control map.

4. The method for constructing a phase-controlled model for geostatistical inversion according to claim 1, characterized in that: The step 3) also includes a step of correcting the output results of the machine learning model based on a planar space interpolation method.

5. The method for constructing a phase-controlled model for geostatistical inversion according to claim 4, characterized in that: Plane space interpolation methods include the variogram method and the Kriging interpolation method.

6. The method for constructing a phase-controlled model for geostatistical inversion according to claim 1, characterized in that: Neural network model is used in machine learning model.

7. The method for constructing a phase-controlled model for geostatistical inversion according to claim 6, characterized in that: The neural network model uses PNN or RBF.

Images