CN114492627A - A prediction method of shale brittleness index based on improved KNN algorithm - Google Patents

Abstract

The invention discloses a shale brittleness index prediction method based on an improved KNN algorithm, which includes determining training wells and test wells; performing correlation analysis and comparison between each logging parameter in the training wells and the brittleness index, and selecting the absolute correlation coefficient. The three logging parameters whose values are in the top three are used as independent variables, and the brittleness index is used as the dependent variable; the data in the training wells are used to construct training samples and training database; the KNN algorithm is used to iteratively optimize the training database to obtain the optimal training database; The data in the database is then used as training data, the data corresponding to the test wells and the training wells are used as test data, and the cross-validation method is used to obtain the optimal K value; the optimal training database and the optimal K value are used to establish a KNN model to obtain a prediction model. It can improve the accuracy and stability of model prediction, and improve the prediction accuracy.

Description

A prediction method of shale brittleness index based on improved KNN algorithm

technical field

The invention relates to a shale brittleness index prediction method, in particular to a shale brittleness index prediction method based on an improved KNN algorithm.

Background technique

For the characterization of shale reservoir brittleness, on the basis of existing rock brittleness evaluation methods and combined with exploration and development experience, related scholars have successively proposed many evaluation methods to characterize shale brittleness index. However, based on different evaluation purposes, scholars in different fields have proposed different definitions and evaluation formulas. So far, the industry has not had a unified understanding of the definition and evaluation methods of shale brittleness. Therefore, this is also the focus and difficulty of shale-related research.

Morley et al. (1944) defined brittleness as the absence of plasticity. Ramsay (1967) believed that brittleness refers to the destruction of cohesion within the rock; Obert et al. (1967) proposed that brittleness is the property of a material to fracture at or slightly greater than the yield stress. Evans et al. (1990) defined the deformation degree less than 1% as brittleness, more than 5% as ductile, and the rest as brittle-ductile transition. George (1995) defined rock brittleness as the ability of rock to deform continuously without permanent deformation when a stress exceeding that required to generate microcracks is applied to the rock. Goktan and Gunes (2005) defined brittleness as the tendency to fracture without significant deformation under low stress. Holt (2011) believes that there is no specific definition of brittleness, and a considerable number of index parameters can be obtained from laboratory and oilfield field data. The field is closely linked to find the most appropriate definition of brittleness. On the basis of summarizing the macro and micro failure characteristics of shale, Li Qinghui et al. (2012) believed that brittleness is a comprehensive characteristic of the material, which is the internal non-uniform stress generated under its own natural heterogeneity and external specific loading conditions, and lead to local failure. , and then the ability to form multi-dimensional fracture surfaces. Tarasov and PotVin (2013) considered brittleness to be the ability of a rock to self-sustain macroscopic failure in the post-peak failure process due to the accumulation of elastic energy in the material under its own natural heterogeneity and external specific loading conditions. Wang Yu et al. (2014) believed that brittleness is a comprehensive characteristic of rock materials, controlled by both internal and external factors. The internal factor is the heterogeneity of rock materials, mainly referring to the mineral particles, structure and structure of the rock, and the external factor is the rock under specific loading conditions. The ability to generate internal non-uniform stress and lead to local failure, thereby forming a multi-dimensional fracture surface. According to rock mechanics, brittleness refers to the property of an object to rupture with little deformation after being subjected to force. Brittleness is explained in geology when a material exhibits little or no plastic deformation before fracture or failure. Although there is no clear definition of brittleness, the current consensus is that rocks are highly brittle if they exhibit the following characteristics during failure: (1) failure occurs at low strain; (2) fracture-dominated fracture failure; (3) rock is composed of fine grains; ④ High compressive/tensile strength ratio; ⑤ High rebound energy; ⑥ Large internal friction angle; ⑦ Complete crack development during hardness test.

The evaluation of shale brittleness index by logging data is a convenient, economical and practical method, and the rock mechanics elastic parameter method and the brittle mineral method based on rock mineralogy are the most widely used at present. The brittle mineral method has certain advantages in practical application, and the log mineral interpretation results corrected by the test can obtain the brittleness characterization profile of the whole well section, which is highly practical. However, simply considering the mineral composition of the rock, it is easy to ignore the influence of diagenesis. At the same time, under different geological conditions, the composition of brittle minerals in rocks may be different, and due to the different chemical composition of each mineral, its petrophysical brittleness is not the same, and its brittleness is also different, so this method has certain limitations.

The rock mechanics elastic parameter method can effectively predict the brittleness index by using logging data and seismic data, but it lacks theoretical support, and cannot reflect the influence of confining pressure and the microscopic characteristics of fracture, so it is difficult to characterize the brittleness difference between deep and shallow shale. In addition, when calculating elastic parameters, the results corresponding to different calculation methods or inversion methods may be different, which makes the elastic parameter method unable to have good universality and compatibility.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a shale brittleness index prediction method based on the improved KNN algorithm, which solves the above problems and can greatly improve the prediction accuracy of the KNN algorithm.

In order to achieve the above object, the technical solution adopted in the present invention is as follows: a shale brittleness index prediction method based on an improved KNN algorithm, comprising the following steps;

(1) The original logging data of several wells are known, and the original logging data includes logging parameters and brittleness indices at n different depths in the well; the logging parameters include neutron porosity CNL, shear wave transit time DTS , potassium content K, acoustic time difference AC, natural gamma GR and organic matter content TOC, select one well as training well, and the rest are test wells;

(2) Conduct correlation analysis and comparison between each logging parameter in the training well and the brittleness index, select the three logging parameters whose absolute value of the correlation coefficient is in the top three as independent variables, and mark them as independent variable A, independent variable B, independent variable C, brittleness index as the dependent variable;

(3) Construct training samples and training database;

Three independent variables at the same depth of the training well constitute the training sample x _i ={x _Ai , x _Bi , x _Ci }, i represents the i-th depth, i=1～n, x _Ai , x _Bi , x _Ci respectively Represents the value of the independent variable A, the value of the independent variable B, and the value of the independent variable C at the i-th depth. The fragility index corresponding to this depth is y _i , and y _i is used as the label of the training sample, and all training samples constitute the training database. ;

(4) using the KNN algorithm to iteratively optimize the training database to obtain the optimal training database; including steps (41)-(49);

(41) All training samples are formed into a matrix X={X _A , X _B , X _C }, where X _A , X _B , X _C are column vectors formed by all x _Ai , x _Bi , and x _Ci respectively, and all brittleness The index forms a column vector Y according to the depth; preset an iterative matrix X _t ={X _At , X _Bt , X _Ct }, the number of iterations M, t=1～M; and when t=1, X _At =X _A , X _Bt =X _B , X _Ct =XC;

(42) The elements in the matrix X are respectively normalized to obtain the normalized matrix X';

(43) Normalize the elements in the matrix X _t respectively to obtain the normalized matrix X _t ′;

(44) Calculate the predicted value of the matrix X _t ', including (a1)-(a5);

(a1) Calculate the Euclidean distance between each row of the matrix X' and the first row of the matrix X _t ' respectively, and obtain several Euclidean distance values;

(a2) Use the bubble sort algorithm to sort all Euclidean distance values from small to large to obtain a distance sequence;

(a3) Find the labels of the training samples corresponding to the first K Euclidean distance values in the distance sequence, obtain K labels, and average the K labels as the predicted value of the fragility index in the first row of X _t ';

(a4) According to the method of (a1)-(a3), obtain the predicted value of brittleness index for each row of X _t ';

(a5) Sort all the predicted values of brittleness index by depth to form a column vector Y _t , which is used as the predicted value of the matrix X _t ′;

(45) Calculate the mean square error MSE between Y _t and Y, the allowable absolute error E, and the prediction rate rate;

(46) preset an absolute error, retain the training samples with absolute error≤E in the matrix X, and iteratively update X _t ;

(47) Repeat steps (43)-(46) until the number of iterations is reached;

(48) Compare the mean square error MSE and the prediction rate calculated by each iteration, and select the iteration matrix corresponding to the number of iterations with small mean square error and high prediction rate as the optimal training database;

(5) In the KNN algorithm, the data in the optimal training database is used as the training data, the data corresponding to the test wells and the training wells are used as the test data, and the optimal K value is obtained by the cross-validation method;

(6) Based on the KNN algorithm, a KNN model is established by using the optimal training database and the optimal K value as a shale brittleness prediction model;

(7) Select a well to be logged, obtain the logging data corresponding to the optimal training database in its original logging data, input it into the shale brittleness prediction model, and output its predicted value.

As a preference: in step (1), the method for selecting training wells is: comparing the amount of brittleness index data, brittleness distribution, and high brittleness index ratio of each well, and selecting a large amount of brittleness index data, a uniform brittleness distribution, and a large proportion of high brittleness index well as a training well.

Preferably: in step (2), the abnormal value includes zero value, negative value, and abnormally large value.

As preferably: in the step (31), the following formula is used for normalization;

x represents the element in the vector, x' represents the normalized element, and x _max and x _min represent the maximum and minimum values in the vector, respectively.

As a preference: in step (45), calculate the mean square error MSE, allowable absolute error E, and prediction rate rate between Y _t and Y, and specifically calculate by the following formula:

E=(BI _max -BI _min )×100% (2)

In formula (1), y _i is the brittleness index at the i-th depth of the training well in step (3), and y _ti is the i-th element of the column vector Y _t ;

In formula (2), BI _max and BI _min are the maximum and minimum values in Y _t of this iteration;

In formula (3), M is the number of iterations, time is the number of times the prediction is accurate, and the criterion is: if MSE<E in this iteration, the prediction is considered accurate.

Compared with the prior art, the advantages of the present invention are:

Find out the logging parameters suitable for the prediction model through correlation analysis, and constitute the training sample and training database. The training sample is highly correlated with the brittleness index;

The KNN algorithm has been improved. The traditional category with the highest frequency of occurrence of the first K points is used as the prediction classification to be tested, and the average value of the first K points is changed to be the prediction value of the test, thus forming an improved KNN algorithm. To iteratively optimize the training database to obtain the optimal training database;

The optimal training database is used as the training data, and the corresponding data in the test wells and the training wells are used as the test data, and the optimal K value is obtained by the cross-validation method;

Finally, the optimal training database and the optimal K value are brought into the KNN model for training, and a KNN algorithm-shale brittleness prediction model is obtained.

Description of drawings

Fig. 1 is a flow chart of the present invention.

Fig. 2 is the distribution diagram of original BI value of training well with depth;

Figure 3a is the intersection diagram of neutron porosity CNL and BI;

Fig. 3b is the intersection diagram of shear wave time difference DTS and BI;

Fig. 3c is the intersection diagram of potassium K and BI;

Fig. 4 is the distribution diagram of the BI predicted value obtained by inputting the optimal training database into the shale brittleness prediction model with depth;

Figure 5 is a histogram of the change of the mean square error MSE with the increase of the number of iterations;

Figure 6 is a histogram of the change of the prediction rate rate with the number of iterations;

Fig. 7 is the line graph of the change of mean square error MSE with the change of K value;

Figure 8 is a line graph showing the change of the prediction rate rate with the change of the K value;

Fig. 9 is the distribution diagram of original BI value of test well with depth;

Fig. 10 is a distribution diagram of the predicted BI value of test wells with depth.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings.

Example 1:

Referring to Fig. 1-Fig. 3, a method for predicting shale brittleness index based on improved KNN algorithm, including the following steps;

(1) The original logging data of several wells are known, and the original logging data includes logging parameters and brittleness indices at n different depths in the well; the logging parameters include neutron porosity CNL, shear wave transit time DTS , potassium content K, acoustic time difference AC, natural gamma GR and organic matter content TOC, select one well as training well, and the rest are test wells;

(2) Conduct correlation analysis and comparison between each logging parameter in the training well and the brittleness index, select the three logging parameters whose absolute value of the correlation coefficient is in the top three as independent variables, and mark them as independent variable A, independent variable B, independent variable C, brittleness index as the dependent variable;

(3) Construct training samples and training database;

Three independent variables at the same depth of the training well constitute the training sample x _i ={x _Ai , x _Bi , x _Ci }, i represents the i-th depth, i=1～n, x _Ai , x _Bi , x _Ci respectively Represents the value of the independent variable A, the value of the independent variable B, and the value of the independent variable C at the i-th depth. The fragility index corresponding to this depth is y _i , and y _i is used as the label of the training sample, and all training samples constitute the training database. ;

(4) using the KNN algorithm to iteratively optimize the training database to obtain the optimal training database; including steps (41)-(49);

(41) All training samples are formed into a matrix X={X _A , X _B , X _C }, where X _A , X _B , X _C are column vectors formed by all x _Ai , x _Bi , and x _Ci respectively, and all brittleness The index forms a column vector Y according to the depth; preset an iterative matrix X _t ={X _At , X _Bt , X _Ct }, the number of iterations M, t=1～M; and when t=1, X _At =X _A , X _Bt =X _B , X _Ct =X _C ;

(42) The elements in the matrix X are respectively normalized to obtain the normalized matrix X';

(43) Normalize the elements in the matrix X _t respectively to obtain the normalized matrix X _t ′;

(44) Calculate the predicted value of the matrix X _t ', including (a1)-(a5);

(a1) Calculate the Euclidean distance between each row of the matrix X' and the first row of the matrix X _t ' respectively, and obtain several Euclidean distance values;

(a2) Use the bubble sort algorithm to sort all Euclidean distance values from small to large to obtain a distance sequence;

(a3) Find the labels of the training samples corresponding to the first K Euclidean distance values in the distance sequence, obtain K labels, and average the K labels as the predicted value of the fragility index in the first row of X _t ';

(a4) According to the method of (a1)-(a3), obtain the predicted value of brittleness index for each row of X _t ';

(a5) Sort all the predicted values of brittleness index by depth to form a column vector Y _t , which is used as the predicted value of the matrix X _t ′;

(45) Calculate the mean square error MSE between Y _t and Y, the allowable absolute error E, and the prediction rate rate;

(46) preset an absolute error, retain the training samples with absolute error≤E in the matrix X, and iteratively update X _t ;

(47) Repeat steps (43)-(46) until the number of iterations is reached;

(48) Compare the mean square error MSE and the prediction rate calculated by each iteration, and select the iteration matrix corresponding to the number of iterations with small mean square error and high prediction rate as the optimal training database;

(5) In the KNN algorithm, the data in the optimal training database is used as the training data, the data corresponding to the test wells and the training wells are used as the test data, and the optimal K value is obtained by the cross-validation method;

(6) Based on the KNN algorithm, a KNN model is established by using the optimal training database and the optimal K value as a shale brittleness prediction model;

(7) Select a well to be logged, obtain the logging data corresponding to the optimal training database in its original logging data, input it into the shale brittleness prediction model, and output its predicted value.

In this embodiment, the method for selecting training wells in step (1) is as follows: comparing the brittleness index data volume, brittleness distribution, and high brittleness index ratio of each well, and selecting a large brittleness index data volume, uniform brittleness distribution, and high brittleness index ratio The large well acts as a training well. Of course the method is not limited to this.

The abnormal values in step (2) include zero values, negative values, and abnormally large values.

In step (31), normalization is carried out with the following formula;

x represents the element in the vector, x' represents the normalized element, and x _max and x _min represent the maximum and minimum values in the vector, respectively.

In step (45), the mean square error MSE, the allowable absolute error E, and the prediction rate rate between Y _t and Y are calculated, and the following formula is specifically used to calculate:

E=(BI _max -BI _min )×100% (2)

In formula (1), y _i is the brittleness index at the i-th depth of the training well in step (3), and y _ti is the i-th element of the column vector Y _t ;

In formula (2), BI _max and BI _min are the maximum and minimum values in Y _t of this iteration;

In formula (3), M is the number of iterations, time is the number of times the prediction is accurate, and the criterion is: if MSE<E in this iteration, the prediction is considered accurate.

Example 2:

1-10; in order to better illustrate the solution of the present invention, we will further describe Embodiment 1.

For the selection of training wells in step (1), see Fig. 2. In Fig. 2, the abscissa is the depth, and the ordinate is the brittleness index BI. It can be seen from Fig. 2 that the well has a large number of brittle samples, uniform brittleness It is characterized by the large proportion of the number of samples with high brittleness, and the data in it is suitable as the training data of the prediction model.

In step (2), when each logging parameter in the training well is correlated and compared with the brittleness index, we find that the neutron porosity CNL, the shear wave transit time DTS and the potassium content K are the three measurements. Well parameters have strong correlation with brittleness index, showing a relatively consistent negative correlation trend, and they can be measured by conventional logging, and the prediction cost is low, so these three logging parameters are selected as independent variables of the prediction model. The intersection diagrams of these three logging parameters and brittleness index BI are shown in Fig. 3a-Fig. 3c. In the three graphs, the abscissa is the brittleness index BI, and the ordinate is the neutron porosity CNL, shear wave transit time DTS and potassium content. K.

Referring to Fig. 4, the optimal training database obtained by the present invention is input into the shale brittleness prediction model, and the distribution diagram of the obtained BI prediction value with depth is shown in Fig. 4, where the abscissa is the depth, and the ordinate is the brittleness index BI, which is the same as Fig. 4. 1 Compared with the original brittleness index, the average absolute error of the prediction result is MAE=0.48, the error rate compared with the curve in Figure 2 is 1.2%, and the prediction accuracy is 98%. The prediction effect of the shale brittleness index in the well is excellent.

Referring to Fig. 5 and Fig. 6, in order to find the optimal number of iterative optimizations, that is, to find the optimal training database, the iterative matrix constructed in step (41) of the present invention is input into the KNN algorithm, and K=4 is taken. The variance MSE and the prediction accuracy rate are used to reflect the optimization effect of each iteration. The results are shown in the clustered histograms 5 and 6. The abscissa in the figure is the number of iterative optimizations. 0 means no optimization. When the number of iterative optimizations is 2 or At 3 times, the MSE reaches a minimum valley value and the rate reaches a maximum peak value, indicating that the model has strong stability and high prediction accuracy, and the optimal database is obtained at this time.

See Figure 7 and Figure 8: In order to find the best K value, the data in the optimal training database is input into the KNN algorithm, and the prediction effect of the model with different K values is reflected by the mean square error MSE and the prediction accuracy rate of the prediction results. , the results are shown in Figure 7 and Figure 8. When the K value is 5-8, the MSE reaches a minimum valley value and the rate reaches a maximum peak value. At this time, the stability and prediction accuracy of the model are better, and the performance is the best. is the optimal K value.

In order to verify the effect of the present invention, see Figure 9 and Figure 10: we select a well to be logged, obtain the logging data corresponding to the optimal training database in its original logging data, and know the original BI depth distribution corresponding to its logging data As shown in Fig. 9, these logging data are sent into the shale brittleness prediction model of the present invention, and the distribution of the obtained BI prediction value with depth is shown in Fig. 10. Compared with the original brittleness index in Fig. 9, the prediction result is The mean absolute error MAE=1.78, the error rate from the curve in Fig. 9 is 4.5%, the prediction accuracy is 85%, and the prediction effect of the interwell shale brittleness index is good.

It is worth mentioning that the training wells, test wells, and wells to be logged are located in the same work area.

The above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (5)

1. A shale brittleness index prediction method based on an improved KNN algorithm is characterized by comprising the following steps: comprises the following steps;

(1) knowing original logging information of a plurality of wells, wherein the original logging information comprises logging parameters and brittleness indexes at n different depths in the wells; the logging parameters comprise neutron porosity CNL, transverse wave time difference DTS, potassium content K, sound wave time difference AC, natural gamma GR and organic matter content TOC, one well is selected as a training well, and the rest are test wells;

(2) performing correlation analysis and comparison on each logging parameter in the training well and the brittleness index, selecting three logging parameters with the absolute values of correlation coefficients in the first three as independent variables, respectively marking the three logging parameters as an independent variable A, an independent variable B and an independent variable C, and using the brittleness index as a dependent variable;

(3) constructing a training sample and a training database;

three independent variables at the same depth of the training well form a training sample x_i＝{x_Ai，x_Bi，x_CiI denotes the ith depth, i is 1 to n, x_Ai，x_Bi，x_CiRespectively representing the value of an independent variable A, the value of an independent variable B and the value of an independent variable C at the ith depth, wherein the brittleness index corresponding to the depth is y_iWill y is_iAs a label of the training sample, all the training samples form a training database;

(4) iteratively optimizing the training database by using a KNN algorithm to obtain an optimal training database; comprising steps (41) - (49);

(41) forming all training samples into a matrix X ═ X_A，X_B，X_CIn which X is_A、X_B、X_CAre respectively all x_Ai、x_Bi、x_CiForming a column vector Y by all brittleness indexes according to the depth; presetting an iterative matrix X_t＝{X_At，X_Bt，X_CtThe iteration times M, t is 1-M; and when t is 1, X_At＝X_A、X_Bt＝X_B、X_Ct＝X_C；

(42) Respectively normalizing the elements in the matrix X to obtain a normalized matrix X';

(43) for matrix X_tRespectively normalizing the elements to obtain a normalized matrix X_t′；

(44) Calculating the matrix X_tThe predictive value of' including (a1) - (a 5);

(a1) each row of the matrix X' is respectively connected with the matrix X_tThe first row of' calculates the Euclidean distance to obtain a plurality of Euclidean distance values;

(a2) sorting all Euclidean distance values from small to large by using a bubble sorting algorithm to obtain a distance sequence;

(a3) searching labels of training samples corresponding to the first K Euclidean distance values in the distance sequence to obtain K labels, and averaging the K labels to obtain X_t' predictive value of brittleness index of first line;

(a4) according to the methods of (a1) to (a3), X is obtained_t' prediction of brittleness index for each row;

(a5) all brittleness index predicted values are arranged according to depth to form a column vector Y_tAs a matrix X_t' is predicted value;

(45) calculating Y_tMean square error MSE with Y, allowable absolute error E, and prediction rate;

(46) presetting absolute errors, reserving training samples with absolute errors less than or equal to E in the matrix X, and iteratively updating X_t；

(47) Repeating steps (43) - (46) until the number of iterations is reached;

(48) comparing the mean square error MSE and the prediction rate of each iterative calculation, and selecting an iterative matrix corresponding to the iteration times with small mean square error and high prediction rate as an optimal training database;

(5) in the KNN algorithm, data in the optimal training database is used as training data, corresponding data in the test well and the training well are used as test data, and an optimal K value is obtained by adopting a cross validation method;

(6) establishing a KNN model by utilizing the optimal training database and the optimal K value based on a KNN algorithm to serve as a shale brittleness prediction model;

(7) and selecting a well to be logged, acquiring logging data corresponding to the optimal training database in original logging data, inputting the logging data into the shale brittleness prediction model, and outputting a predicted value of the shale brittleness prediction model.

2. The shale brittleness index prediction method based on the improved KNN algorithm according to claim 1, wherein: in the step (1), the method for selecting the training well comprises the following steps: and comparing the brittleness index data amount, the brittleness distribution and the high brittleness index proportion of each well, and selecting the well with large brittleness index data amount, uniform brittleness distribution and high brittleness index proportion as a training well.

3. The shale brittleness index prediction method based on the improved KNN algorithm according to claim 1, wherein: in step (2), the abnormal values include zero values, negative values, and abnormally large values.

4. The shale brittleness index prediction method based on the improved KNN algorithm according to claim 1, wherein: in the step (31), normalization processing is performed by the following formula;

x represents the element in the vector, x' represents the normalized element, x_maxAnd x_minRespectively representing the maximum and minimum values in the vector.

5. The improved KNN algorithm-based shale brittleness index prediction method according to claim 1, characterized in that: in step (45), Y is calculated_tThe mean square error MSE, the allowable absolute error E and the prediction rate between the prediction rate and the Y are calculated by adopting the following formulas:

E＝(BI_max-BI_min)×100％ (2)

in the formula (1), y_iThe brittleness index at the ith depth of the training well in the step (3) is y_tiIs a column vector Y_tThe ith element of (1);

in formula (2), BI_maxAnd BI_minIs Y of this iteration_tMaximum and minimum values of;

in the formula (3), M is the iteration number, time is the number of accurate prediction, and the judgment standard is as follows: and if the MSE is less than E in the iteration, the prediction is considered to be accurate.

Images