Multivariate data coupling formation pressure calculation method and system based on machine learning

Abstract

The invention relates to a multivariate data coupling formation pressure calculation method based on machine learning. The multivariate data coupling formation pressure calculation method comprises the following steps that S1, well logging data and well logging data of multiple wells are collected; s2, simulating actual formation pressure based on a Dc index, drilling efficiency-mechanical specific energy monitoring, gas logging auxiliary judgment, actually measured pressure measurement data and the like; s3, establishing a unary fitting equation between the formation pressure and the logging data; s4, establishing a binary fitting equation between the formation pressure and the logging parameters; s5, establishing a formation pressure calculation model based on logging-logging binary coupling data; s6, calculating formation pressure based on a formation pressure calculation model of logging-logging binary coupling data; s7, correcting the formation pressure obtained in the step S6; s8, establishing a logging-logging parameter thermodynamic diagram of each well; s9, on the basis of a result displayed by the thermodynamic diagram, eliminating by utilizing a correlation principle to obtain a final measuring and recording-well parameter data set; and S10, establishing a formation pressure calculation model based on logging multivariate data coupling.

Description

Technical Field

[0001] The present invention belongs to the field of rock mechanics, and in particular relates to a multi-data coupled formation pressure calculation method based on machine learning. Background Art

[0002] The basement of the Qiongdongnan Basin is mainly composed of granite buried hills of the Indosinian and Yanshanian periods (age: 65Ma). The formation and evolution of the granite buried hills in the Qiongdongnan Basin were affected by multi-period tectonic movements in the Indosinian, Yanshanian and Himalayan periods, which provided good conditions for the formation of weathering crusts and internal fracture reservoirs in the buried hills. The Qiongdongnan buried hills are divided into three buried hill belts: the southern buried hill belt, the central buried hill belt and the northern buried hill belt. In general, from the southern buried hill belt to the northern buried hill belt, the burial depth, burial time and current seawater depth are decreasing, and the location of the buried hill development is rising.

[0003] At present, the research on overpressure in sandstone and mudstone formations has gradually matured at home and abroad. In recent years, the formation pressure monitoring implemented in the Yingqiong Basin has already had preliminary means of formation pressure monitoring based on the causes of pressure, and the pressure monitoring accuracy is high in formations with conventional lithology and common causes of pressure (undercompaction, fluid expansion, pressure conduction, etc.). However, in recent years, bedrock buried hills have gradually become one of the hot areas of exploration and development. The evaluation technology for the special lithology formation pressure of bedrock buried hills in the Qiongdongnan Basin is relatively limited, which seriously restricts the evaluation accuracy of formation pressure.

[0004] At present, the electrical sound wave and logging Dc index are mainly used to evaluate the overpressure of the buried hill formation, but these two parameters cannot accurately indicate the overpressure of the buried hill formation. The main reasons are: 1) The skeleton of the buried hill formation is dense, and the speed of the sound wave mainly depends on the development degree of the rock skeleton and fractures. The overpressure of the pore fluid has little effect on the speed of the sound wave, and the electrical sound wave cannot effectively indicate the overpressure; 2) The logging Dc and Sigma index are mainly used for on-site quantitative monitoring of overpressure caused by undercompaction of sedimentary rocks, but the existence of overpressure can be indicated in bedrock formations. There are some problems, that is, the mechanism of the change of logging parameters is unclear, and the accuracy of the quantitative evaluation results is poor. Therefore, the present invention proposes a multivariate data coupling formation pressure calculation method based on machine learning, which comprehensively considers the interaction between the drill bit and the buried hill formation, and selects or constructs a logging data body that can effectively characterize the overpressure of the buried hill formation, providing a theoretical basis for the accurate evaluation of the Yingqiong Basin and the buried hill formation pressure. Summary of the invention

[0005] A method for calculating formation pressure by coupling multivariate data based on machine learning, characterized in that it comprises the following steps:

[0006] S1: Collect logging data from multiple wells. The logging parameters mainly include acoustic time difference and density. The logging parameters mainly include GR, resistivity, drilling speed, drill string speed, MSE, Dc index, drill bit lateral work, drill bit longitudinal work, drilling fluid circulating equivalent density (ECD), and overburden pressure equivalent density (OBG).

[0007] S2: Simulate the actual formation pressure based on Dc index, drilling efficiency-mechanical specific energy monitoring, gas measurement auxiliary judgment, measured pressure data, etc.

[0008] S3: Draw an intersection diagram between the well logging data and the formation pressure value, and use the univariate linear regression method to obtain a univariate fitting equation between the formation pressure value and the well logging data.

[0009] S4: Draw an intersection diagram between the logging data, the mud logging data and the formation pressure value, and use the binary linear regression method to obtain the binary fitting equation between the formation pressure value and the logging-mud logging parameters.

[0010] S5: A binary fitting equation between the formation pressure value and the logging parameters is integrated to establish a formation pressure calculation model based on the logging-logging binary coupling data.

[0011] S6: Formation pressure calculation model based on well logging binary coupling data, calculation of formation pressure, including PP→DEN&Dc, PP→DEN&MSE, PP→VP&Dc, PP→VP&MSE.

[0012] S7: The new formation pressure obtained in S6 is comprehensively corrected by the program to obtain PP-correction.

[0013] S8: Based on the collected logging data and PP-corrected data, a logging parameter thermogram of each well is established to obtain the correlation between the logging parameters and the formation pressure.

[0014] S9: Based on the results shown in the heat map, the code is used to delete one of the two parameters with strong correlation to obtain the well logging parameter data set after elimination.

[0015] S10: Based on the data set obtained after elimination, data training is performed using a machine learning algorithm to establish a formation pressure calculation model based on the coupling of multivariate logging data. The machine learning algorithm is a random forest method, a multivariate linear regression method, a decision tree method, or a support vector machine method.

[0016] In the above S5, it is characterized in that the model of the new method for calculating formation pressure based on the coupling of binary data of well logging is:

[0017] P p=m×VP+n×logging parameters (Dc, SIGA, MSE, longitudinal work, transverse work)+O

[0018] P p =j×DEN+k×logging parameters (Dc, SIGA, MSE, longitudinal work, transverse work)+L

[0019] Where: m, n, O, j, k, L are coefficients, which need to be determined by block.

[0020] In the S6, the PP→DEN&Dc refers to a new formation pressure calculated by using the formation pressure calculation model of logging-logging binary coupling data using the logging density and the logging parameter Dc index; the PP→DEN&MSE refers to a new formation pressure calculated by using the formation pressure calculation model of logging-logging binary coupling data using the logging density and the logging parameter MSE; the PP→VP&Dc refers to a new formation pressure calculated by using the formation pressure calculation model of logging-logging binary coupling data using the logging parameter sonic time difference and the logging parameter Dc index; the PP→VP&MSE refers to a new formation pressure calculated by using the formation pressure calculation model of logging-logging binary coupling data using the logging parameter sonic time difference and the logging parameter MSE.

[0021] In the above S7, outliers are first identified and deleted, and then the data set is checked and the rows with null values ​​and NaN values ​​are deleted using the detect outliers method. Then several new formation pressure results are combined to obtain the PP-correction.

[0022] In the above S8, the heat map display results are mainly distinguished by color to determine the size of the correlation, and the heavier the color, the greater the correlation coefficient.

[0023] In the above-mentioned S9, attributes having a correlation (absolute value) of 0.9 or more are deleted by code.

[0024] In the above-mentioned S10, the results obtained by the heat map and the machine learning algorithm are integrated to select the model with the best effect (the highest prediction correlation coefficient).

[0025] The present invention thus provides a system for calculating formation pressure by coupling multivariate logging data based on machine learning, characterized in that it includes a module for executing steps S2 to S10 in the method according to any one of claims 1 to 8, specifically, a data entry module for entering data from step S1; a data processing module for executing data calculations in steps S2 to S10 (specifically, a binary data fitting module for executing data calculations in steps S2 to S6; a pp-correction module for executing data correction and strong correlation elimination in steps S7 to S9; a multivariate data coupling module for outputting a data coupling model obtained in step S10); and a result output module for outputting the result obtained by the data processing module. The above can be implemented by obtaining a relevant computer program through programming in a computing language.

[0026] Preferably, the computer program of the multivariate data coupling formation pressure calculation module based on the random forest method in the multivariate data coupling module is as follows:

[0027]

[0028] Optionally, a result display module is also included, such as displaying through a screen or a remote terminal.

[0029] Compared with the prior art, the present invention has the following advantages: the present invention discloses a multivariate data coupling formation pressure calculation method based on machine learning, which makes full use of logging data, establishes a formation pressure calculation model based on logging-logging multivariate data coupling data through a multivariate data fitting module, and uses logging data to evaluate and eliminate the influence of fracture development degree on rock breaking efficiency. Further, according to the logging data analysis, the interaction between the drill bit and the buried hill formation is comprehensively considered to construct a logging data body that can effectively characterize the overpressure of the buried hill formation. The present invention discloses a multivariate data coupling formation pressure calculation method based on machine learning, which eliminates the influence of strongly correlated parameters through a pp-correction module and a comprehensive thermal map analysis result to simplify the data model, and has higher applicability. The method of the present invention comprehensively considers the interaction between the drill bit and the buried hill formation, establishes a new method for monitoring the buried hill formation pressure, and can realize the quantitative evaluation of the buried hill overpressure. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Flowchart of the formation pressure calculation method based on machine learning coupled multivariate logging data.

[0031] Figure 2-1 to Figure 2-3 The logging data collected in step 1.

[0032] Figure 3 An intersection plot of well logging data and formation pressure.

[0033] Figure 4An intersection plot of logging data and formation pressure.

[0034] Figure 5-1 to Figure 5-3 The formation pressure is calculated by the formation pressure calculation model based on the binary coupling data of well logging and well logging.

[0035] Figure 6 Data processing in step 7.

[0036] Figure 7-1 to Figure 7-3 PP-corrected result in step 7.

[0037] Figure 8 Thermal map of correlation of logging data.

[0038] Fig. 9 Attribute correlation bar chart (without absolute values).

[0039] Fig.10 Random forest method prediction results.

[0040] Fig.11 Multiple linear regression was used to predict the results.

[0041] Fig.12 Decision tree method predicts the results.

[0042] Fig.13 Support vector machine prediction results. DETAILED DESCRIPTION

[0043] The following is a more detailed description of the multivariate data coupling formation pressure calculation method based on machine learning of the present invention in conjunction with the accompanying drawings and examples, wherein the formation pressure calculation method flow chart is as follows: Figure 1 The following is an example to illustrate the specific implementation process.

[0044] Step 1: Select an overpressure well A in the buried hill formation and query the corresponding logging parameters VP, DEN and logging parameters GR, resistivity, drilling speed, drill string speed, MSE, Dc index, drill bit transverse work, drill bit longitudinal work, ECD, OBG. The specific collection information is as follows: Figure 2-1 to Figure 2-3 .

[0045] Step 2: simulate the actual formation pressure based on Dc index, drilling efficiency-mechanical specific energy monitoring, gas measurement auxiliary judgment, measured pressure data, etc. The specific steps are as follows:

[0046] 1) For the same well A, select the logging data used to calculate the formation pressure;

[0047] 2) Monitor formation pressure based on Dc index and drilling efficiency-mechanical specific energy, and calculate formation pressure at different depths;

[0048] 3) Assist in determining formation pressure based on gas logging data;

[0049] 4) Obtain the formation pressure at the test point through actual on-site testing;

[0050] The final simulated actual formation pressure is obtained by mutual correction between the Dc index, drilling efficiency-mechanical specific energy monitoring, gas measurement auxiliary judgment, and measured pressure data evaluation formation pressure results.

[0051] Step 3: draw an intersection diagram between the well logging data and the formation pressure value, and use the univariate linear regression method to obtain a univariate fitting equation between the formation pressure value and the well logging data. The specific steps are as follows:

[0052] 1) Based on the logging data collected in step 1 and the formation pressure data calculated in step 2, draw an intersection diagram of well A logging data and formation pressure, such as Figure 3 .

[0053] 2) Using the univariate linear regression method, the binary fitting equation between the formation pressure value and the logging data is obtained.

[0054] Step 4: Draw an intersection diagram between logging parameters, logging data and formation pressure values, and use a binary linear regression method to obtain a binary fitting equation between the formation pressure value and the logging parameters. The specific steps are as follows:

[0055] 1) Based on the logging data collected in step 1 and the formation pressure data calculated in step 2, draw an intersection diagram of the logging data and formation pressure of well A, such as Figure 4 .

[0056] 2) The binary linear regression method is used to obtain the binary fitting equation between the formation pressure value and the logging parameters, as shown in the following table:

[0057] Table 1 Binary fitting equation between formation pressure value and logging parameters of Well A

[0058]

[0059] Step 5: A binary fitting equation between the formation pressure value and the logging-mud logging parameters is integrated to establish a formation pressure calculation model based on the logging-mud logging binary coupling data.

[0060] P p =m×VP+n×logging parameters (Dc, SIGA, MSE, longitudinal work, transverse work)+O

[0061] P p =j×DEN+k×logging parameters (Dc, SIGA, MSE, vertical attack, horizontal attack)+L Step 6: Calculate the formation pressure based on the formation pressure calculation model of logging-logging binary coupling data ( Figure 5-1 to Figure 5-3), including PP→DEN&Dc, PP→DEN&MSE, PP→VP&Dc, PP→VP&MSE. The specific steps are as follows:

[0062] 1) For the same well A, the corresponding logging data collected in step 1 are substituted into the binary fitting equation obtained in step 4.

[0063] 2) Calculate the new formation pressure: PP→DEN&Dc, PP→DEN&MSE, PP→VP&Dc, PP→VP&MSE, see Figure 5.

[0064] Step 7, the new formation pressure obtained in S6 is comprehensively corrected by the program to obtain PP-correction, such as Figure 6 As shown, the specific steps are as follows:

[0065] 1) Identify and delete outliers, then check the data set and use the detect outliers method to delete rows with null values ​​and NaN values, such as Figure 6 .

[0066] 2) Combining several new formation pressure results, we get the PP-correction, see Figure 7-1 to Figure 7-3 .

[0067] Step 8: Based on the collected logging data and PP-corrected data, a logging parameter thermodynamic map of each well is established. The specific steps are as follows:

[0068] 1) Read the logging data for the same well A and delete the empty columns that are all NaN values.

[0069] 2) Perform attribute analysis on the data after removing outliers and draw heat maps and attribute correlation bar charts.

[0070] Step 9: Based on the results of the heat map display, use the code to delete one of the two parameters with strong correlation to obtain the logging parameter data set after elimination. The specific steps are as follows:

[0071] 1) From the correlation heat map of logging data ( Figure 8 ) It can be seen that the parameters in well A, “drilling rate and ECD” (correlation is -0.94), “drilling rate and OBG” (correlation is -0.96), and “ECD and OBG” (correlation is 0.95), are strongly correlated, and you can choose one of them in the subsequent parameter selection.

[0072] 2) Use the code to delete parameters with strong correlation properties (the absolute value of the correlation coefficient is above 0.9) ( Fig. 9 ).

[0073] Step 10: Based on the data set obtained after elimination, four machine learning algorithms are used for data training. The results are as follows: Fig.10 The random forest prediction results are shown. Fig.11 The prediction results of the multiple linear regression method are shown. Fig.12 The prediction results of the decision tree method are shown. Fig.13 The prediction results of the support vector machine method are shown. By comparing the correlation coefficients of each prediction result, the best machine learning algorithm is selected as the random forest method (correlation coefficient is 0.781). The final method for calculating the multivariate data coupling formation pressure of Well A is the multivariate data coupling formation pressure calculation based on the random forest method. The calculation procedure in the multivariate data coupling module is as follows:

[0074]

Claims

1. A method for calculating formation pressure based on multivariate coupling of logging and well logging data based on machine learning, characterized in that: The following steps are involved: S1: Collect logging data from multiple wells. Logging parameters include acoustic time difference VP, density DEN, and logging parameters mainly include gamma logging GR, resistivity RT, drilling speed, drill string speed, mechanical specific energy MSE, Dc index, drill bit lateral work, drill bit longitudinal work, equivalent circulating density ECD, and overburden pressure OBG; S2: Simulate the actual formation pressure based on Dc index, drilling efficiency-mechanical specific energy monitoring, gas measurement auxiliary judgment, measured pressure data, etc.; S3: draw an intersection diagram between the well logging data and the formation pressure value, and use a univariate linear regression method to obtain a univariate fitting equation between the formation pressure value and the well logging data; S4: draw an intersection diagram between logging data, mud logging data and formation pressure values, and use a binary linear regression method to obtain a binary fitting equation between the formation pressure value and the logging-mud logging parameters; S5: A binary fitting equation between the formation pressure value and the logging-mud logging parameters is integrated to establish a formation pressure calculation model based on the logging-mud logging binary coupling data; S6: Formation pressure calculation model based on logging-mud logging binary coupling data, calculation of formation pressure, including formation pressure calculation based on density logging-Dc index binary coupling PP→DEN&Dc, formation pressure calculation based on density logging-mechanical specific energy binary coupling PP→DEN&MSE, formation pressure calculation based on acoustic time difference logging-Dc index binary coupling PP→VP&Dc, formation pressure calculation based on acoustic time difference logging-mechanical specific energy binary coupling PP→VP&MSE; S7: The formation pressure obtained in S6 is comprehensively corrected by the program to obtain PP-correction; S8: Based on the collected logging data and PP-corrected data, a logging parameter thermogram of each well is established to obtain the correlation between the logging parameters and the formation pressure; S9: Based on the results displayed in the heat map, the code is used to delete one of the two parameters with strong correlation, and the logging parameter data set after elimination is obtained; S10: Based on the data set obtained after elimination, data training is performed using a machine learning algorithm; a formation pressure calculation model based on the coupling of multivariate logging data is established; the machine learning algorithm is selected from a random forest method, a multivariate linear regression method, a decision tree method or a support vector machine method.

2. The method according to claim 1, characterized in that In the above S5, the formation pressure calculation model based on the coupling of well logging binary data is: P p =m×VP+n×logging parameters (D, SIG4, MSE, longitudinal work, transverse work)+O P p =j×DEN+k×logging parameters (Dc, SIGA, MSE, longitudinal work, transverse work)+L Where: m, n, O, j, k, L are coefficients, which need to be determined by block.

3. The method according to claim 1, characterized in that In the S6, the PP→DEN&Dc refers to a new formation pressure calculated by using the formation pressure calculation model of logging-logging binary coupling data using the logging density and the logging parameter Dc index; the PP→DEN&MSE refers to a new formation pressure calculated by using the formation pressure calculation model of logging-logging binary coupling data using the logging density and the logging parameter MSE; the PP→VP&Dc refers to a new formation pressure calculated by using the formation pressure calculation model of logging-logging binary coupling data using the logging parameter sonic time difference and the logging parameter Dc index; the PP→VP&MSE refers to a new formation pressure calculated by using the formation pressure calculation model of logging-logging binary coupling data using the logging parameter sonic time difference and the logging parameter MSE.

4. The method according to claim 1, characterized in that In the above S7, outliers are first identified and deleted, and then the data set is checked and the rows with null values ​​and NaN values ​​are deleted using the detect outliers method. Then several new formation pressure results are combined to obtain the PP-correction.

5. The method according to claim 1, characterized in that In the above S8, the heat map display results are mainly distinguished by color to determine the size of the correlation, and the heavier the color, the greater the correlation coefficient.

6. The method according to claim 1, characterized in that In the above-mentioned S9, the code is used to delete the parameters having a strong correlation attribute (the absolute value of the correlation coefficient is above 0.9).

7. The method according to claim 1, characterized in that In the S10, it is characterized in that the machine learning algorithm mainly includes the random forest method.

8. A system for calculating formation pressure based on multivariate data coupling of well logging and logging based on machine learning, characterized in that: It includes a module for executing steps S2 to S10 in the method according to any one of claims 1 to 8, and specifically includes a data entry module for entering data from step S1; A data processing module, used to perform the data calculation in steps S2 to S10 (specifically, including a binary data fitting module, used to perform the data calculation in steps S2 to S6; a pp-correction module, used to perform the data correction and strong correlation elimination in steps S7 to S9; A multivariate data coupling module, used to output the data coupling model obtained in step S10); and a result output module, used to output the results obtained by the data processing module; Preferably, the computer program of the multivariate data coupling formation pressure calculation module based on the random forest method in the multivariate data coupling module is as follows:

9. The system according to claim 9, characterized in that Also includes a result display module.

Images