Automatic classification and prediction method for researching natural gas exploration workload

Abstract

The invention discloses an automatic classification prediction method for researching natural gas exploration workload, and the method comprises the steps: building an automatic classification model based on multiple regression: supposing that all sample data are x1n, x2n... xpn, the number of influence indexes is p, the number of all sample data is n, and the exploration workload is y, classifying according to different linear relationships between all exploration workloads and all sample data; and constructing prediction sample data, and calculating the exploration workload of the prediction sample data by the automatic classification model. The problem that an existing prediction method is based on experience or simple statistical analysis and lacks systematicness and accuracy is solved. The method is simple in principle, easy to implement and short in calculation time consumption, adopts automatic classification, automatic optimization and automatic prediction, and has very wide application prospects.

Description

Technical Field

[0001] The invention relates to a natural gas exploration workload classification method, and in particular to a method for studying automatic classification prediction of natural gas exploration workload. Background Art

[0002] Natural gas exploration is an important part of energy development, and the accuracy of its workload forecast directly affects exploration decisions and investment benefits. Traditional forecasting methods are often based on experience or simple statistical analysis, lacking systematicity and accuracy. Therefore, developing a method that can automatically classify and predict natural gas exploration workload is of great significance for improving exploration efficiency and reducing costs. Summary of the invention

[0003] The purpose of the present invention is to provide a method for studying automatic classification and prediction of natural gas exploration workload, which solves the problem that the existing prediction methods are based on experience or simple statistical analysis and lack systematicity and accuracy. It adopts automatic classification, automatic optimization and automatic prediction and has very broad application prospects.

[0004] In order to achieve the above object, the present invention provides a method for studying automatic classification prediction of natural gas exploration workload, the method comprising:

[0005] Step 1: Establish an automatic classification model based on multivariate regression

[0006] Assume that all sample data are x 1n , x 2n ……x pn , the number of influencing indicators is p, the number of all sample data is n, and the exploration workload is y. If all exploration workloads and all sample data satisfy a linear relationship, all sample data are divided into one category, as follows:

[0007] y j =a0+α1x 1j +a2x 2j +a3x 3j +…+a p x pj (1)

[0008] In formula (1), j represents the number of the jth sample data randomly selected from n samples, j = 0, 1...n; n represents the total number of all sample data; p represents the number of influencing indicators; a p represents the coefficient of the pth influencing indicator; x pj represents the jth sample data in the pth influencing indicator; y j Represents the jth exploration workload.

[0009] If all exploration workloads and all sample data satisfy different linear relationships, the data points that satisfy the same linear relationship are classified into one category. Assume that all sample data can be divided into k categories, and the number of samples in each category is n. k , n k ≥p+1, then each class satisfies the following formula:

[0010] y j，k =a 0，k +a 1，k x 1j，k +a 2j，k x 2j , k +..+a pj，k x pj，k (2)

[0011] In formula (2), k represents the total number of categories, k∈[1,n); j represents the number of the jth sample data in the kth category, j=0,1…nk; nk represents the number of sample data in the kth category; p represents the number of influencing indicators, y j，k represents the fitted value of the j-th exploration workload of the k-th category; x pj，k represents the jth sample data of the pth influencing indicator in the kth category; a p，k represents the regression parameter of the p-th influencing indicator in the k-th category;

[0012] According to the principle of least squares method, the optimal solution of the regression parameter a in formula (2) is obtained: for:

[0013]

[0014] In formula (3):

[0015]

[0016] The steps to identify which sample data can be classified into one category among all sample data are as follows:

[0017] (1) Set the number of sample data n for each category k ≥p+1;

[0018] (2) Through the loop The more samples n, the greater the number of cycles. Therefore, the number of random sample cycles is set to 1000, and n non-repeating samples are randomly selected from all sample data. k Perform multiple regression analysis on sample data and obtain regression parameters Minimize the sum of squares of the fitting errors between the fitting value and the actual value of the exploration workload of the sample data Calculate the maximum value R of the absolute value of the relative error between the fitting value and the actual value of the exploration workload of the sample data in the kth category max =max{|R 1，k |,|R 2，k |,…|R j，k |}; the relative error of the exploration workload for calculating k-classified sample data is: Wherein jk All represent the fitting values ​​of the exploration workload of the kth category calculated by the formula (2); Both represent the actual value of the jth exploration workload of the kth category

[0019] (3) Using the regression parameters in step (2) Then calculate the absolute value of the relative error between the fitting value and the actual value of the exploration workload of the remaining sample data | R o |, wherein the |R o |The o in is the number of remaining samples, o = 1, 2, ..., nk.

[0020] If | R o |<R max , the sample data is classified into this category, then n k =n k +1; when all remaining sample data are classified, the number of sample data that cannot be classified is nn k , if nn k Repeat steps (1) to (3) until nn k <p classification is completed, then nn is satisfied k Each sample data of <p is a single large category;

[0021] Step 2: Construct prediction sample data and calculate the exploration workload of the prediction sample data by the automatic classification model

[0022] Construct prediction sample data, calculate the classification distance between the sample data and the prediction sample data, and obtain the minimum value of the classification distance; obtain the number of the sample data based on the minimum value of the classification distance, query the regression parameters in all classifications in the automatic classification model based on the number of the sample data, and calculate the predicted value of the prediction sample data and the sum of squares of the fitting errors of the prediction sample data based on the regression parameters.

[0023] Preferably, the number of all sample data n = 20, the impact index p = 7, in step 1, assuming n k When =p+2=9, the sample data is classified into four categories, of which there are two categories that meet the requirements of multiple regression, which are recorded as the first category and the second category.

[0024] More preferably, the regression parameter of the first category is The sample data of the first category n1=9 is obtained, and the maximum absolute value of the relative error is 0.017328571%.

[0025] More preferably, in the remaining 20-9=11 sample data, repeat steps (1) to (3) to obtain the sample data of the second largest category n2=9; the regression parameter of the second largest category is The maximum absolute value of the relative error is 0.107075%.

[0026] More preferably, the number of remaining samples is 20-18=2, satisfying n k <p+1, so the remaining data are divided into one category each. There are four categories in total.

[0027] Preferably, the number of all sample data n = 20, the impact index p = 7, in step 1, assuming n k =p+3=10, the sample data is classified into 2 categories, recorded as the first category and the second category.

[0028] More preferably, the regression parameter of the first category is The sample data of the first category n1=10 is obtained, and the maximum absolute value of the relative error is 0.93307%.

[0029] More preferably, the regression parameter of the second category is The sample data of the second largest category n2=10 is obtained, and the maximum absolute value of the relative error is 32.67453333%.

[0030] More preferably, the number of remaining samples is 20-20=0, so the sample data is just divided into two categories.

[0031] Preferably, in step 2, in step 2, the calculation formula of the classification distance is as follows:

[0032]

[0033] In formula (4):

[0034]

[0035] In formula (4), x i，n is the nth item of the i-th influencing indicator in the predicted sample data, i = 1, 2, ... p; x i，j，k is the jth item of data of the i-th influencing indicator in the k-th category of the sample data, j = 1, 2, ..., n k ;D j，k Represents the classification distance from the nth item of data in the predicted sample data to the jth item of data in the kth category of the sample data.

[0036] The minimum value of the classification distance is calculated as follows:

[0037] D jmin =min{D j，1 , D j，2 , …, D j，k} (5)

[0038] In formula (5), D j，1 represents the classification distance from the nth item in the predicted sample data to the jth item in the first category of the sample data, D j，k Represents the classification distance from the nth item of data in the predicted sample data to the jth item of data in the kth category of the sample data;

[0039] The calculation of the predicted value of the predicted sample data is as follows:

[0040] y j，m =a 0，m +a 1，m x 1j +a 2，m x 2j +a 3，m x 3j +…+a p，m x pj (6)

[0041] In formula (6), m represents the mth class in the sample data, m = 1, 2, ..., k; j represents the jth sample data in the predicted sample data, j = 1, 2 ... n k ;n k represents the number of predicted sample data; p represents the number of influencing indicators; y j，m Represents the predicted value of the predicted sample data in the mth class; q p，m represents the coefficient of the pth influencing indicator in the mth category; x pj Represents the jth predicted sample data of the pth influencing indicator

[0042] Preferably, if the sum of squares of the fitting errors of the predicted sample data is minimized, the predicted value of the predicted sample data is the exploration workload of the predicted sample data.

[0043] Preferably, if the number of samples classified in the automatic classification model is n k When <p+1, the regression parameter cannot be obtained. The nearest principle is adopted and the exploration workload of the predicted sample data is the actual value of the exploration workload of the jth sample data.

[0044] The method for studying automatic classification prediction of natural gas exploration workload of the present invention solves the problem that the existing prediction method is based on experience or simple statistical analysis and lacks systematicity and accuracy, and has the following advantages:

[0045] 1. The method of the present invention is simple in principle, easy to implement, and takes a short time to calculate. It adopts automatic classification, automatic optimization, and automatic prediction, and has a very broad application prospect.

[0046] 2. This method can be extended to nonlinear classification prediction, which can provide a scientific basis for exploration decision-making, help decision-makers better understand the distribution and change trend of exploration workload, and thus formulate more reasonable exploration plans and investment strategies. At the same time, this method can also be applied to other similar fields, such as oil exploration and mineral resource development. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a diagram of key factors affecting the workload in the marine conventional gas field of the present invention.

[0048] Figure 2 It is a multiple regression prediction curve diagram commonly used in the present invention.

[0049] Figure 3 For the present invention k =p+1The first category of fitting curve graph.

[0050] Figure 4 For the present invention k =p+1The second largest type of fitting curve graph.

[0051] Figure 5 For the present invention k =p+2The first category of fitting curve graph.

[0052] Figure 6 For the present invention k =p+2The second largest type of fitting curve graph.

[0053] Figure 7 For the present invention k =p+3The first category of fitting curve diagram.

[0054] Figure 8 For the present invention k =p+3The second largest type of fitting curve diagram.

[0055] Fig. 9 This is a graph of the neural network learning fitting curve of the present invention. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] Example 1

[0058] A method for automatically classifying and predicting natural gas exploration workload is provided, the method comprising:

[0059] 1. Analysis of factors affecting the exploration workload of conventional marine gas reservoirs

[0060] Through the analysis of multiple marine conventional gas reservoirs, the present invention believes that the deployment of exploration wells is closely related to the six conditions of "generation, storage, capping, reservoir, transportation and protection" required for oil and gas formation and the main controlling factors of reservoir formation, among which hydrocarbon source conditions, sedimentary microfacies, effective reservoir thickness, effective porosity, reservoir continuity, development of ancient and modern structures, development degree of hydrocarbon source faults, gas reservoir closure type, reserve abundance, area of ​​favorable zones and success rate of exploration wells are key factors, which jointly affect the deployment of exploration wells. Therefore, the main factors affecting the workload of marine conventional gas field exploration are established.

[0061] like Figure 1 As shown in the figure, the key factors affecting the workload in the field of marine conventional gas in the present invention. Figure 1 It can be seen that the factors affecting the deployment of exploration wells include geological factors and comprehensive factors. The geological factors include reserve abundance, hydrocarbon generation intensity, ancient and modern structural closures, development of hydrocarbon source faults, sedimentary microfacies, reservoir thickness, reservoir porosity, reservoir distribution continuity and favorable zone area; the comprehensive factors include the success rate of exploration wells in gas reservoirs (inside + outside) and the number of exploration wells in gas reservoirs (inside + outside).

[0062] 2. Select and establish characteristic index factors from the factors affecting exploration workload

[0063] Taking structural traps as an example, data are screened out from the factors affecting exploration workload, and characteristic indicators related to natural gas exploration workload are extracted from production. Therefore, the factors affecting the workload in the marine conventional gas field of structural traps are simplified to: effective thickness, porosity, success rate of development wells, reserve abundance, area of ​​favorable areas, total number of exploration wells, gas-bearing area and hydrocarbon generation intensity factors. The statistical results are shown in Table 1. The characteristics that have a significant impact on the prediction results are screened out. The exploration workload mainly predicts how many wells can meet the corresponding workload. The commonly used prediction method is multivariate linear regression prediction.

[0064] Table 1 Main factors affecting workload in the construction field

[0065]

[0066]

[0067] like Figure 2 As shown in FIG. 1 , the commonly used multiple regression prediction curve diagram of the present invention. Figure 2It can be seen that there is a large error between the actual value and the fitted value. If this method is used to predict future workload, it will cause a large error. Therefore, it is necessary to establish an automatic classification prediction model based on multiple regression for prediction.

[0068] 3. Build an automatic classification model based on multivariate regression

[0069] Currently, many classification models have advantages, disadvantages and applicable fields, see Table 2 for details.

[0070] Table 2 Statistics of advantages and disadvantages of different classification models

[0071]

[0072]

[0073] As shown in Table 2, these classification methods do not consider the regularity of the impact of various indicator factors on workload, such as whether it is linear or nonlinear, and these classifications only classify from the discrete aspect of the data.

[0074] However, for the prediction of exploration workload, both the classification of regularity and its predictability must be considered. Therefore, the present invention establishes an automatic classification prediction method based on multivariate regression, and the establishment process is as follows:

[0075] Assume that all sample data are x 1n , x 2n ……x pn , the number of influencing indicators is p, the number of all sample data is n, and the exploration workload is y; if there is a linear relationship between all exploration workloads and all sample data, all sample data are divided into one category, as follows:

[0076] y j =α0+a1x 1j +a2x 2j +a3x 3j +…+a p x pj (1)

[0077] In formula (1), j represents the number of the jth sample data randomly selected from n samples, j = 0, 1...n; n represents the total number of all sample data; p represents the number of influencing indicators; a p represents the coefficient of the pth influencing indicator; x pj represents the jth sample data in the pth influencing indicator; y j Represents the jth exploration workload.

[0078] If all exploration workloads and all sample data satisfy different linear relationships, the data points that satisfy the same linear relationship are classified into one category. Assume that all sample data can be divided into k categories, and the number of samples in each category is n. k , n k ≥p+1, then each class satisfies the following formula:

[0079] y j，k =a 0，k +a 1，k x 1j，k +a 2，k x 2j，k +a 3，k x 3j，k +…+a p，k x pj，k (2)

[0080] In formula (2): k represents the total number of categories, k∈[1,n); j represents the number of the jth sample data in the kth category, j=0,1......n k ;n k represents the number of sample data of the kth classification; p represents the number of influencing indicators, y j，k represents the fitting value of the j-th exploration workload of the k-th category (fitting value of the total number of exploration wells); x pj，k represents the jth sample data of the pth influencing indicator in the kth category; a p，k Represents the regression parameter of the p-th influencing indicator in the k-th category.

[0081] The solution method for each type of coefficient matrix is ​​the same as the multivariate regression method. According to the principle of least squares method, the existing literature 1 (Liu Fan. Review of Least Squares Method [J]. Journal of Central China Normal University (Natural Science Edition), 1992, (4): 46-57) obtains the optimal solution of the regression parameter a in formula (2): for:

[0082]

[0083] In formula (3):

[0084]

[0085] The steps to identify which sample data can be classified into one category are as follows:

[0086] (1) Set the number of sample data n for classification k ≥p+1,n k The minimum is p+1;

[0087] (2) Through the loop Second-rate In order to reduce the number of operations, the number of random sample selections is 1000 times, and n non-repeated samples are randomly selected from all n sample data. k The sample data was subjected to multiple regression analysis to obtain the regression parameters of the corresponding groups. Minimize the sum of squares of the fitting errors between the fitting value and the actual value of the exploration workload of the sample data (i.e. At this time, the maximum absolute value of the relative error between the fitting value and the actual value of the exploration workload of each sample data is calculated. max =max{|R 1，k |,|R 2，k |,…|R j，k |}; According to the existing literature 2 (Zhu Jianxin, Li Youfa. Numerical Computation Methods 3rd Edition [M]. Beijing: Higher Education Press, 2012: 196), the relative error between the fitting value and the actual value of the exploration workload of the k-th classification sample data is calculated as follows: in All represent the actual value of the j-th exploration workload of the k-th category, that is, the total number of exploration wells. All yj,k represent the fitted value of the exploration workload of the k-th category calculated by the formula (2), that is, the fitted value of the total number of exploration wells.

[0088] (3) Using the regression parameters in (2) Then calculate the absolute value of the relative error between the fitting value and the actual value of the exploration workload of the remaining sample data | R o |(where o is the number of remaining samples, o=1, 2, ..., n k ), if |R o |<R max , the sample data is classified into this category, then n k =n k +1; when all remaining sample data are classified, the number of sample data that cannot be classified is nn k , if nn k Repeat steps (1) to (3) until nn k <p classification is completed, then nn is satisfied k Each sample data of <p is a separate large category.

[0089] The above calculation process is analyzed as follows:

[0090] According to the previous sample data, the classification calculation is performed. From Table 1, we know that the number of samples is 20 (excluding the prediction samples), the number of cycles is set to 1000, the influencing factors p = 7, and the number of classifications n k The minimum number is 8.

[0091] So first assume that n k=p+1=8, first randomly select 8 sample data from the total sample, according to the principle of permutation and combination, there are a total of The calculation data is too large, and the number of random sample selection cycles is assumed to be 1000. The existing multivariate regression theory (the above existing literature 1) is used to calculate the minimum value of the sum of squares of the fitting errors between the fitting value and the actual value of the exploration workload of the sample data. in represents the actual value of the jth exploration workload of the kth category, that is, the total number of exploration wells, y j，k All represent the fitting values ​​of the exploration workload of the kth category calculated by the formula (2), that is, the fitting values ​​of the total number of exploration wells, which are specifically shown in Tables 3 to 10) (calculated by formula (3)), and calculate the maximum absolute value of the relative error between the fitting value and the actual value of the exploration workload of each sample data, and use the absolute value of the relative error R in step (2) max =max{|R 1，k |,|R 2，k |,…|R j，k |}(The maximum absolute value of the relative error in the 8 randomly selected samples is calculated to be 0.0000333%). ), and x 1j，k 、x 2j，k ......x pj，k The sample data corresponding to the influencing indicators of effective thickness, porosity, development well success rate, reserve abundance, favorable area, gas-bearing area, and hydrocarbon generation intensity in Table 1 are substituted into formula (2) to obtain the fitting value of exploration workload (fitting value of the total number of exploration wells), and the absolute value of the relative error between the fitting value and the actual value of the exploration workload of the remaining sample data is calculated. o |, if |R o |<R max , classify the sample data into this category, and complete the classification of the sample data, see Tables 3 and 4, such as Figures 3-4 The results can be divided into six categories, of which two categories satisfy the law of multiple regression, and the other four individual samples each constitute a category, as follows:

[0092] Regression parameters of the first category

[0093] , n1 = 8, the maximum absolute value of the relative error is 0.0000333%, and the sum of the squares of the fitting error in represents the actual value of the jth exploration workload of the kth category, that is, the total number of exploration wells, y j，kBoth represent the fitting value of the exploration workload calculated by the formula (2) for the kth category, that is, the fitting value of the total number of exploration wells) is 2E-12, indicating that the fitting effect is good. According to the existing literature 3 (Han Ming. Probability Theory and Mathematical Statistics Textbook 3rd Edition [M]. Shanghai: Tongji University Press, 2023), the specific results are shown in Tables 3 and Figure 3 .

[0094] Table 3 n k =p+1The first category table

[0095]

[0096]

[0097] like Figure 3 As shown, the present invention k =p+1The first type of fitting curve graph. Figure 3 It can be seen that the total number of exploration wells (actual value) and the fitted value of the total number of exploration wells (fitted value) are basically consistent, indicating that the fitting effect is very good.

[0098] The second category, regression parameters

[0099]

[0100] n2=7+1=8, the maximum absolute value of the relative error reaches 0.000033%, and the sum of the squares of the fitting error in represents the actual value of the jth exploration workload of the kth category, that is, the total number of exploration wells, y j，k Both represent the fitting value of the exploration workload calculated by the formula (2) for the kth category, that is, the fitting value of the total number of exploration wells) is 5E-12, indicating that the fitting effect is good. The specific results are shown in Tables 4 and Figure 4 .

[0101] Table 4 n k =p+1The second largest category table

[0102]

[0103] like Figure 4 As shown, the present invention k =p+1The second largest type of fitting curve graph. Figure 4 It can be seen that with the same calculation steps, the total number of exploration wells (actual value) and the fitted value of the total number of exploration wells (fitted value) are basically consistent, indicating that the fitting effect is very good.

[0104] The number of remaining samples is 20-16=4, which does not satisfy n k ≥p+1, so the remaining 4 sample data are divided into one category separately, and a total of six categories are divided into six categories, as shown in Table 5.

[0105] Table 5 n k =p+1 The third, fourth, fifth and sixth categories

[0106]

[0107] According to the previous sample data, classification calculation is performed, assuming that n k =p+2=9, the classification of sample data is completed, and it is divided into four categories. The regression parameters of the first category are n1=9, the maximum absolute value of the relative error is 0.017328571%, and the sum of the squares of the fitting error is in represents the actual value of the jth exploration workload of the kth category, that is, the total number of exploration wells, y j，k The fitting value of the exploration workload calculated by the formula (2) for the kth category is 0.000012579, indicating that the fitting effect is good. The specific results are shown in Tables 6 and Figure 5 .

[0108] Table 6 n k =p+2The first category table

[0109]

[0110]

[0111] like Figure 5 As shown, the first type of fitting curve diagram of nk=p+2 of the present invention. Figure 5 It can be seen that the total number of exploration wells (actual value) and the fitted value of the total number of exploration wells (fitted value) are basically consistent, indicating that the fitting effect is very good.

[0112] The second category, regression parameters n2=7+2=9, the maximum absolute value of relative error reaches 0.107075%, and the sum of squares of fitting errors in represents the actual value of the jth exploration workload of the kth category, that is, the total number of exploration wells, y j，k Both represent the fitting value of the exploration workload calculated by the formula (2) for the kth category, that is, the fitting value of the total number of exploration wells) is 0.00005008, indicating that the fitting effect is good. The specific results are shown in Tables 7 and Figure 6 .

[0113] Table 7 n k =p+2The second largest category table

[0114]

[0115] like Figure 6 As shown, the present invention k =p+2The second largest type of fitting curve graph. Figure 6 It can be seen that with the same calculation steps, the total number of exploration wells (actual value) and the fitted value of the total number of exploration wells (fitted value) are basically consistent, indicating that the fitting effect is very good.

[0116] The number of remaining samples is 20-18=2, which does not satisfy n k ≥p+1, so the remaining 2 sample data are divided into one category each, and a total of four categories, as shown in Table 8.

[0117] Table 8 n k =p+2The third and fourth categories

[0118]

[0119] According to the previous sample data, classification calculation is performed, assuming that n k =p+3=10, the classification of sample data is completed, which is divided into two categories. The regression parameter of the first category is n1=10, the maximum absolute value of relative error reaches 0.93307%, and the sum of squares of fitting errors in represents the actual value of the jth exploration workload of the kth category, that is, the total number of exploration wells, y j，k Both represent the fitting value of the exploration workload calculated by the formula (2) for the kth classification, that is, the fitting value of the total number of exploration wells) is 0.0160267, indicating that the fitting effect is good. The classification data are shown in Table 9. The fitting data are shown in Table 9. Figure 7 shown.

[0120] Table 9 n k =p+3The first category table

[0121]

[0122] like Figure 7 As shown, the present invention k =p+3The first type of fitting curve diagram. Figure 7 It can be seen that the total number of exploration wells (actual value) and the fitted value of the total number of exploration wells (fitted value) are basically consistent, indicating that the fitting effect is very good.

[0123] The second largest type of regression parameters are n2=10, the maximum absolute value of relative error reaches 32.67453333%, and the sum of squares of fitting error in represents the actual value of the jth exploration workload of the kth category, that is, the total number of exploration wells, y j，kAll represent the fitting value of the exploration workload calculated by the formula (2) for the kth category, that is, the fitting value of the total number of exploration wells) is 4.583465, indicating that the fitting effect is poor. The classification data are shown in Table 10. The fitting data are shown in Table 10. Figure 8 shown.

[0124] Table 10 n k =p+3The second largest category table

[0125]

[0126] like Figure 8 As shown, the present invention k =p+3The second largest type of fitting curve diagram. Figure 8 It can be seen that the total number of exploration wells (actual value) and the total number of exploration wells (fitted value) are not consistent, indicating that n k = p+3 The second largest category of fitting effects is not n k =p+3The first category has good fitting effect.

[0127] The number of remaining samples is 20-20=0, so the sample data is just divided into two categories.

[0128] Combined with the previous analysis, n k Different values ​​and classifications will result in different fitting effects.

[0129] 4. Construct prediction sample data, and use the automatic classification model to calculate the exploration workload of the prediction sample data

[0130] Construct prediction sample data, calculate the classification distance between sample data and prediction sample data, and obtain the minimum value of the classification distance;

[0131] The number of the sample data is obtained according to the minimum value of the classification distance, and the regression parameters in all classifications in the automatic classification model are queried from the number of the sample data. The predicted value of the predicted sample data and the sum of squares of the fitting errors of the predicted sample data are calculated from the regression parameters.

[0132] The calculation formula of the classification distance is as follows:

[0133]

[0134] In formula (4), x i，n is the nth item of the i-th influencing indicator in the predicted sample data, i = 1, 2, ... p; x i，j，k is the jth item of the i-th influencing indicator in the k-th category of the sample data, j = 1, 2, ..., n k ;D j，k Represents the classification distance from the nth item of data in the predicted sample data to the jth item of data in the kth category of the sample data;

[0135] The minimum classification distance is calculated as follows:

[0136] D jmin =min{D j，1 , D j，2 , …, D j，k} (5)

[0137] In formula (5), D j，1 represents the classification distance from the nth item in the predicted sample data to the jth item in the first category of the sample data, D j，k Represents the classification distance from the nth item of data in the predicted sample data to the jth item of data in the kth category of the sample data;

[0138] The calculation of the predicted value of the forecast sample data is as follows:

[0139] y j，m =a 0，m +a 1，m x 1j +a 2，m x 2j +a 3，m x 3j +…+a p，m x pj (6)

[0140] In formula (6), m represents the mth class in the sample data, m = 1, 2, ..., k; j represents the jth sample data in the predicted sample data, j = 1, 2 ... n k ;n k represents the number of predicted sample data; p represents the number of influencing indicators; Represents the predicted value of the predicted sample data in the mth class; a p，m represents the coefficient of the pth influencing indicator in the mth category; x pj Represents the j-th predicted sample data of the p-th influencing indicator.

[0141] If the sum of squares of the fitting errors of the predicted sample data is the smallest, the predicted value of the predicted sample data is the exploration workload of the predicted sample data. If the number of classified samples in the automatic classification model is nk<p+1, the regression parameter cannot be obtained, and the exploration workload of the predicted sample data is the actual value of the exploration workload of the jth sample data according to the principle of proximity, as follows:

[0142]

[0143] In formula (7): represents the exploration workload of the predicted sample data; j，m The actual value of the exploration workload of the jth sample data is the total number of exploration wells.

[0144] Experimental Example 1 verifies the effectiveness of the automatic classification prediction method

[0145] In order to verify the effectiveness of the automatic classification prediction method of Example 1, the average value of all the previous samples (i.e., the predicted sample) is taken for prediction. The results show that: the predicted sample (the average value shown in Table 1) is calculated by formula (5) and the results are shown in Table 11. The minimum distance in Table 11 shows that the predicted sample is closest to sample No. 16 (i.e., sample No. 16 in Table 1). In the above classification process (i.e., as shown in Tables 3 to 10), it can be seen that sample No. 16 is closest to sample No. 16 in n k =p+1, it belongs to the first sample in the fifth category, but the sample data of the fifth category is very small, and the regression parameters of the multivariate regression equation cannot be obtained. Therefore, formula (7) is applied to obtain the fitting value of the total number of exploration wells of the prediction sample. is 12, and its fitting square error is 12.6025. Sample No. 16 has k =p+2, it is the eighth sample in the first category, so the regression parameter of the first category is used. (Right now ), and x 1j，m 、x 2j，m ......x pj，m The sample data corresponding to the effective thickness, porosity, development well success rate, reserve abundance, favorable area, gas-bearing area, and hydrocarbon generation intensity index of the predicted sample (such as the fitting values ​​shown in Table 1) are substituted into formula (6) for calculation to obtain the fitting value of the total number of exploration wells: is 11.29, and the sum of squares of the fitting error is 8.0656. k =p+3, the 10th sample belongs to the first category, and the regression parameter of the first category (i.e. ), and x 1j，m 、x 2j，m ......x pj，m The sample data corresponding to the effective thickness, porosity, development well success rate, reserve abundance, favorable area, gas-bearing area, and hydrocarbon generation intensity index of the prediction sample (such as the prediction sample shown in Table 1) are substituted into formula (6) for calculation to obtain the fitting value of the total number of exploration wells: The value is 7.669, and the sum of squares of the fitting error is 0.609961.

[0146] Table 11 Classification distance calculation results

[0147]

[0148]

[0149] Using the existing neural network method, we conduct deep learning on the same prediction samples mentioned above according to the existing literature 4 (Jiao Licheng. Neural Network System Theory [M]. Xi'an: Xi'an University of Electronic Science and Technology Press, 1990: 284), and compare their learning effects. Fig. 9 , using the value of the corresponding indicator of the predicted sample as the prediction input value, the predicted result is 11.32, and the sum of squares of the fitting error is 8.2369.

[0150] Comprehensive analysis of the above experimental results shows that when n k =p+3, the prediction accuracy is higher than n k =p+1,n k =p+2 and network neural network methods, proving that the method of the present invention is reliable. The change of random sampling data will obtain different classification models, but the overall increase in the number of samples and the number of samplings will better reflect the overall distribution and improve the final classification accuracy of the model.

[0151] like Fig. 9 As shown in FIG. 1 , the neural network learning fitting curve diagram of the present invention. Fig. 9 It can be seen that through comparative analysis of experimental results, the automatic classification prediction method of Example 1 can achieve the same effect as neural network prediction, but the neural network depends on sufficient sample data. If the prediction sample is not within the range of the learning sample, the prediction result may have a large deviation, and the neural network cannot express the regularity between indicators.

[0152] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. A method for automatic classification and prediction of natural gas exploration workload, characterized in that: The method includes: Step 1: Establish an automatic classification model based on multivariate regression Assume that all sample data are x 1n , x 2n ……x pn , the number of influencing indicators is p, the number of all sample data is n, and the exploration workload is y; If all exploration workloads and all sample data satisfy a linear relationship, all sample data are classified into one category, as follows: y j =A0+a1x 1j +a2x 2j +a3x 3j +…+a p x pj (1) In formula (1), j represents the number of the jth sample data randomly selected from n samples, j = 0, 1...n; n represents the total number of all sample data; p represents the number of influencing indicators; a p represents the coefficient of the pth influencing indicator; x pj represents the jth sample data in the pth influencing indicator; y j represents the jth exploration workload; If all exploration workloads and all sample data satisfy different linear relationships, all sample data are divided into k categories, and the number of sample data in each category is n k , n k ≥p+1, then each class satisfies the following formula: y j，k =a 0，k +a 1，k x 1j，k +a 2j，k x 2j，k +..+a pj，k x pj，k (2) In formula (2), k represents the total number of categories, k∈[1,n); j represents the number of the jth sample data in the kth category, j=0,1…nk; nk represents the number of sample data in the kth category; p represents the number of influencing indicators, y j，k represents the fitted value of the j-th exploration workload of the k-th category; x pj，k represents the jth sample data of the pth influencing indicator in the kth category; a p，k represents the regression parameter of the p-th influencing indicator in the k-th category; According to the principle of least squares method, the optimal solution of the regression parameter a in formula (2) is obtained: for: In formula (3): The steps to identify which sample data can be classified into one category among all sample data are as follows: (1) Set the number of sample data n for each category k ≥p+1; (2) Randomly select non-repeating n from all sample data k Perform multiple regression analysis on sample data and obtain regression parameters Minimize the sum of squares of the fitting errors between the fitting value and the actual value of the exploration workload of the sample data Calculate the maximum value R of the absolute value of the relative error between the fitting value and the actual value of the exploration workload of the sample data in the kth category max =max{|R 1，k |,|R 2，k |,…|R j，k |}; The relative error of the exploration workload for calculating k-classified sample data is: Wherein jk All represent the fitting values ​​of the exploration workload of the kth category calculated by the formula (2); All represent the actual value of the jth exploration workload of the kth category; (3) Using the regression parameters in step (2) Then calculate the absolute value of the relative error between the fitting value and the actual value of the exploration workload of the remaining sample data | R o |, wherein the |R o The o in | is the number of remaining samples, o = 1, 2, ..., n k ; If | R o | <R max , the sample data is classified into this category, then n k =n k +1; when all remaining sample data are classified, the number of sample data that cannot be classified is nn k , if nn k Repeat steps (1) to (3) until nn k <p classification is completed, then nn is satisfied k Each sample data of <p is a single large category; Step 2: Construct prediction sample data and calculate the exploration workload of the prediction sample data by the automatic classification model Construct prediction sample data, calculate the classification distance between sample data and prediction sample data, and obtain the minimum value of the classification distance; The number of the sample data is obtained according to the minimum value of the classification distance, and the regression parameters in all classifications in the automatic classification model are queried from the number of the sample data. The predicted value of the predicted sample data and the sum of squares of the fitting errors of the predicted sample data are calculated from the regression parameters.

2. The method according to claim 1, characterized in that: Assume that the number of all sample data is n=20, and the influencing index is p=7. In step 1, assuming that nk=p+2=9, the sample data is classified into four categories, of which two categories satisfy the requirements of multiple regression, which are recorded as the first category and the second category.

3. The method according to claim 2, characterized in that The regression parameters of the first category are The sample data n1 of the first category is obtained as 9, and the maximum absolute value of the relative error is 0.017328571%.

4. The method according to claim 2, characterized in that: In the remaining 20-9=11 sample data, repeat steps (1) to (3) to obtain the sample data of the second largest category n2=9; the regression parameter of the second largest category is The maximum absolute value of the relative error is 0.107075%.

5. The method according to claim 1, characterized in that Assume that the number of all sample data is n = 20, and the influence index is p = 7. In step 1, assume that n k =p+3=10, the sample data is classified into 2 categories, recorded as the first category and the second category.

6. The method according to claim 5, characterized in that The regression parameters of the first category are The sample data of the first category n1=10 is obtained, and the maximum absolute value of the relative error is 0.93307%.

7. The method according to claim 5, characterized in that The regression parameters for the second category are The sample data of the second largest category n2=10 is obtained, and the maximum absolute value of the relative error is 32.67453333%.

8. The method according to claim 1, characterized in that: In step 2, the calculation formula of the classification distance is as follows: In formula (4), x i,n is the nth item of the i-th influencing indicator in the predicted sample data, i = 1, 2, ... p; x i,j,k is the jth item of data of the i-th influencing indicator in the k-th category of the sample data, j = 1, 2, ..., n k ;D j,k Represents the classification distance from the nth item of data in the predicted sample data to the jth item of data in the kth category of the sample data; The minimum value of the classification distance is calculated as follows: D jmin =min{D j，1 ,D j，2 ,…,D j，k } (5) In formula (5), D j，1 represents the classification distance from the nth item in the predicted sample data to the jth item in the first category of the sample data, D j，k Represents the classification distance from the nth item of data in the predicted sample data to the ith item of data in the kth category of the sample data; The calculation of the predicted value of the predicted sample data is as follows: y j，m =a 0，m +a 1，m x 1j +a 2，m x 2j +a 3，m x 3j +…+a p，m x pj (6) In formula (6), m represents the mth category in the sample data, m=1, 2, ..., k; j represents the jth sample data in the predicted sample data, j=1, 2...nk; nk represents the number of predicted sample data; p represents the number of influencing indicators; Represents the predicted value of the predicted sample data in the mth class; a p，m represents the coefficient of the pth influencing indicator in the mth category; x pj Represents the j-th predicted sample data of the p-th influencing indicator.

9. The method according to claim 1, characterized in that: If the sum of squares of the fitting errors of the predicted sample data is the smallest, the predicted value of the predicted sample data is the exploration workload of the predicted sample data.

10. The method according to claim 1, characterized in that If the number of samples n classified in the automatic classification model k When < p + 1, the regression parameters cannot be obtained. Using the principle of proximity, the exploration workload of the predicted sample data is taken as the actual value of the exploration workload of the i-th sample data.

Images