Formation evaluation method and system based on well logging big data

Abstract

The present invention provides a formation evaluation method and system based on well logging big data, which is implemented based on physical simulation and machine learning and belongs to the field of well logging technology. The method includes: acquiring well logging data, and training a machine learning model according to the well logging data to obtain a machine learning model; collecting well logging measurements, and preprocessing the measurements according to preset rules to obtain training sample data; training the training sample data based on the machine learning model, and outputting calculation results; and evaluating the corresponding formation according to the calculation results. The complete well logging data is used to construct a machine learning model, and the well logging data is interpreted based on the machine learning model, so that the associated parameters of the machine learning model are more comprehensive and the model training results are more in line with the actual situation.

Description

Technical Field

[0001] The present invention relates to the field of well logging technology, and in particular to a formation evaluation method based on well logging big data and a formation evaluation system based on well logging big data. Background Art

[0002] Well logging interpretation is a comprehensive technology that studies the physical properties and seepage characteristics of underground rocks, and searches for and evaluates oil and gas based on the logging data collected by logging equipment. The existing well logging interpretation technology mainly includes the selection of interpretation models, establishment of logging response equations, preliminary selection of interpretation parameters for trial interpretation, optimization of interpretation parameters for formal processing, and verification of the reliability of interpretation results. In this process, the interpretation model and interpretation model parameters are selected manually, so the experience of the interpreter plays a key role in the accuracy of processing and interpretation results. At present, well logging data processing software is widely used to analyze the reservoir characteristics of the reservoir and find possible production layers through single well logging data. According to regional experience and the knowledge of well logging interpreters, the main geological parameters reflecting the characteristics of the formation are calculated and analyzed for credibility. The bound water content of the production layer is analyzed to reveal the characteristics of the oil and gas layer and the changes in the oil (gas) saturation limit, and to judge the properties of the oil and gas layer. Other information constraints other than logging are rarely used to analyze and identify the oil, gas, and water content of the reservoir, evaluate the abundance and production capacity of the oil and gas layer, and predict the water content of the production layer. These are models developed using limited data from expert experience, well logs and rock physics experiments.

[0003] Therefore, when determining the lithology identification of complex formations, the use of limited lithology identification logging curves often results in large errors and is difficult to meet the needs of on-site exploration and development. In addition, when determining the geological parameters of the formation, due to the limitations of acoustic, electrical, and nuclear logging methods, it is difficult to accurately reflect parameters such as formation porosity and permeability. When identifying reservoir fluids, the accuracy of logging interpretation of oil, gas, and water layers is not high due to the difficulty in determining bound water. Therefore, the identification accuracy is not high and the model applicability is not strong. Based on the many problems existing in existing logging interpretation technology, it is necessary to create a new formation evaluation method based on logging big data. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a formation evaluation method and system based on well logging big data, so as to at least solve the problems of low accuracy and poor applicability of existing well logging interpretation technologies.

[0005] In order to achieve the above-mentioned objectives, the first aspect of the present invention provides a formation evaluation method based on well logging big data, which is implemented based on physical simulation and machine learning. The method includes: acquiring well logging data, and training a machine learning model according to the well logging data to obtain a machine learning model; collecting well logging measurement values, and preprocessing the measurement values ​​according to preset rules to obtain training sample data; training the training sample data based on the machine learning model and outputting calculation results; and performing corresponding formation evaluation according to the calculation results.

[0006] Optionally, the machine learning model includes: a logging value normalization model, a rock physics prediction model, a complex lithology identification model, an oil and gas reservoir monitoring model, and a mechanical evaluation model.

[0007] Optionally, the acquiring of well logging data includes: acquiring well logging simulation data and well logging big data.

[0008] Optionally, obtaining the logging simulation data includes: obtaining parameters that affect the logging results, and obtaining the linkage relationship between the changed parameters and the corresponding parameters of the observation target layer through the variable retention method; the changed parameter is one of the parameters that affect the logging results; the parameters that affect the logging results include at least: mud filtrate resistivity, well diameter, invasion zone resistivity, invasion zone radius, and thickness of the observation target layer; changing the parameters that affect the logging results one by one, and obtaining the linkage relationship between multiple changed parameters and the corresponding parameters of the observation target layer; and establishing a logging response library as the logging simulation data based on the linkage relationship between the multiple changed parameters and the corresponding parameters of the observation target layer.

[0009] Optionally, the acquiring of big well logging data includes: extracting all historical data related to well logging stored in an actual production process as the big well logging data.

[0010] Optionally, the logging measurement values ​​include one or more of the following: resistivity, mud resistivity, borehole diameter, formation thickness, surrounding rock resistivity, invasion zone diameter, invasion zone resistivity and measurement instrument parameters.

[0011] Optionally, the preprocessing of the measured values ​​according to preset rules includes: performing logarithmic transformation of the measured values; deleting outlier samples in the measured values ​​after logarithmic transformation; and performing linear scale transformation of the measured values.

[0012] Optionally, the training sample data is trained based on the machine learning model and the calculation results are output, including: dividing the training sample data into three parts according to a preset ratio, namely training data, verification data and test data; performing model training based on the three divided parts of data, and the model training process is implemented based on the Encoder-Decoder model, classification tree regression tree machine learning algorithm and enabling learning method; and outputting the calculation results.

[0013] The second aspect of the present invention provides a formation evaluation system based on well logging big data, the system comprising: an acquisition unit, used to obtain well logging data and to acquire well logging measurement values; a model building unit, used to perform machine learning model training based on the well logging data to obtain a machine learning model; a processing unit, used to perform preprocessing of the measurement values ​​according to preset rules to obtain training sample data; a training unit, used to train the training sample data based on the machine learning model and output calculation results; the processing unit is also used to perform corresponding formation evaluation based on the calculation results; the model building unit, the processing unit and the training unit are all constructed based on a private cloud computing platform.

[0014] On the other hand, the present invention provides a computer-readable storage medium having instructions stored thereon, which, when executed on a computer, enables the computer to execute the above-mentioned formation evaluation method based on well logging big data.

[0015] Through the above technical solution, a machine learning model suitable for actual working conditions is obtained based on well logging data, and then the model is trained based on the collected well logging data to obtain accurate evaluation parameters for evaluating the formation. Because the complete well logging data is used, the associated parameters of the machine learning model are more comprehensive, and the model training results are more in line with the actual situation.

[0016] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention. In the accompanying drawings:

[0018] Figure 1 It is a flowchart of the steps of a formation evaluation method based on well logging big data provided by one embodiment of the present invention;

[0019] Figure 2 It is a schematic diagram of a process of acquiring logging data provided by one embodiment of the present invention;

[0020] Figure 3 is a flow chart of preprocessing of well logging measurement data provided by one embodiment of the present invention;

[0021] Figure 4 It is a system structure diagram of a formation evaluation system based on well logging big data provided by one embodiment of the present invention.

[0022] Description of Reference Numerals

[0023] 10-acquisition unit; 20-model building unit; 30-processing unit; 40-training unit. DETAILED DESCRIPTION

[0024] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0025] In the existing logging methods, it is mainly based on expert experience and rock physics models. The conventional processing and interpretation steps are divided into: selecting an interpretation model, establishing a logging response equation, preliminarily selecting interpretation parameters for trial interpretation, optimizing interpretation parameters for formal processing and verifying the reliability of the interpretation results. In this process, the interpretation model and interpretation model parameters are selected manually, and the experience of the interpreter plays a key role in the accuracy of the processing and interpretation results. The comprehensive interpretation of logging data is mainly used to determine the properties of the fluid produced by the reservoir and evaluate the quality of the oil and gas layer, including the reservoir storage and permeability parameters of the reservoir and the evaluation of the production capacity of the reservoir. Due to the large number of interpretation selection parameters and the single rock physics model attributes, the required parameters cannot accurately reflect the actual situation of the downhole formation.

[0026] At present, logging data processing software is widely used. It analyzes the reservoir characteristics of the reservoir and finds possible production layers through single well logging data. According to regional experience and the knowledge of logging interpretation personnel, the main geological parameters reflecting the characteristics of the formation are calculated and analyzed for credibility. The bound water content of the production layer is analyzed to reveal the characteristics of the oil and gas layer and the changes in the oil (gas) saturation limit, and to determine the properties of the oil and gas layer. Other information constraints other than logging are rarely used to analyze and identify the oil, gas and water content of the reservoir, evaluate the abundance and production capacity of the oil and gas layer, and predict the water content of the production layer. These are models established based on the limited data of expert experience, logging and rock physics experiments. When determining the lithology identification of complex formations, the use of limited lithology identification logging curves often causes large errors and is difficult to meet the needs of on-site exploration and development. In addition, when determining the geological parameters of the formation, due to the limitations of acoustic, electrical and nuclear logging methods, it is difficult to accurately reflect the parameters such as formation porosity and permeability. When identifying reservoir fluids, due to the difficulty in determining bound water, the accuracy of logging interpretation of oil, gas and water layers is not high. Therefore, the recognition accuracy is not high and the model applicability is not strong. The solution of the present invention is based on the powerful computing power of the private cloud computing platform. By integrating logging big data, physical simulation and machine learning models, the true parameters of the formation are obtained, the logging knowledge implicit in the big data is discovered, knowledge is learned from data and expert experience, and the complex problems in oil and gas field exploration and development are solved.

[0027] Figure 4 FIG. 1 is a system structure diagram of a formation evaluation system based on well logging big data provided by an embodiment of the present invention. Figure 4 As shown, an embodiment of the present invention provides a formation evaluation system based on well logging big data, and the system includes: an acquisition unit 10, used to obtain well logging data and to acquire well logging measurement values; a model building unit 20, used to perform machine learning model training according to the well logging data to obtain a machine learning model; a processing unit 30, used to perform preprocessing of the measurement values ​​according to preset rules to obtain training sample data; a training unit 40, used to train the training sample data based on the machine learning model and output calculation results; the processing unit 30 is also used to perform corresponding formation evaluation according to the calculation results; the model building unit 20, the processing unit 30 and the training unit 40 are all constructed based on a private cloud computing platform.

[0028] In an embodiment of the present invention, the model building unit 20, the processing unit 30 and the training unit 40 are all built on a private cloud computing platform. The well logging big data private cloud computing platform established by the present invention includes four layers, the first is the computing power support layer of the platform, namely IAAS; the second is the platform support layer, namely PAAS layer; the third is the software support layer, namely SAAS layer; the fourth is the application support layer, APPS, which satisfies the user's problem data upload and result feedback through the API interface. On this private cloud computing platform, the machine learning algorithm is applied to quickly learn the true parameters of the actual formation, and the machine learning well logging interpretation model is established using the well logging big data cloud computing to solve the problems of well logging instrument design, rock physics analysis, intelligent well logging interpretation, complex lithology and unconventional oil and gas evaluation. The existing technology is to use the powerful computing power of the well logging big data private cloud computing, well logging big data, physical simulation and machine learning models to obtain the true parameters of the formation, discover the well logging knowledge implicit in the big data, learn knowledge from the data and expert experience, and improve the technology for solving complex problems in the exploration and development of oil and gas fields.

[0029] Its simulation data comes from the simulation module of the PAAS platform layer, which provides the establishment of a logging response library through simulation. The numerical simulation module supports numerical simulation projects such as microsphere focusing, dual sensing, array sensing, various lateral and through-casing resistivity logging. The cloud computing platform includes the management of simulation module projects, how to perform calculations, the management of calculation results (Session), and the operation of formation models. Another source of data is logging big data and other related test data.

[0030] Preferably, the private cloud computing platform is configured with hardware computing power support: Xeon E-7 4-core processor, RAS supports hot plugging, 512GB DDR memory, 10 expandable SDD hard disks, and data exchange 10G network card. The private cloud consists of four parts of the architecture, basic software and hardware support, platform support, software development layer, and application layer.

[0031] Provide data protocols between IAAS / PAAS / SAAS / APPS cloud service layers based on various service level agreements (SLAs). Cloud service SLAs between cloud service providers and end users or user organizations of services address issues such as service availability, data integrity, privacy, security, and data protection. Infrastructure layer based on IAAS. The PAAS service layer built supports the cloud platform layer, and the PAAS layer is a proprietary software support layer. The application layer of SAAS applications is used to provide different types of cloud services, such as logging interpretation processing applications, logging simulation programs, etc.

[0032] The well logging big data cloud computing platform is a private cloud platform that provides dedicated software, end-user management, task requests and feedback for specialized users. The basic service platform provides basic hardware support, including CPU, GPU, memory, high-speed data exchange, SSD hard disk for distributed big data storage and processing, and software including well logging big data coordination scheduling, resource management, online and offline machine learning algorithm library, structured and unstructured databases.

[0033] The virtual platform service layer mainly provides proprietary user interfaces, user platform resource management and configuration, and develops and tests cloud platform application software. The software service layer is used to provide application stores for proprietary users, who access the application stores through the Internet, with the cloud as the center. The application platform includes the user's data interface terminal and the processing result display terminal.

[0034] Figure 1 FIG. 1 is a flow chart of a formation evaluation method based on well logging big data provided by an embodiment of the present invention. Figure 1 As shown, an embodiment of the present invention provides a formation evaluation method based on well logging big data, the method comprising:

[0035] Step S10: Acquire well logging data, and perform machine learning model training based on the well logging data to obtain a machine learning model.

[0036] Specifically, the solution of the present invention is to use well logging big data cloud computing to establish a machine learning well logging interpretation model, which requires obtaining a sufficient amount of training samples for training the obtained interpretation model, that is, the machine learning model. Acquire well logging data and use the well logging data as training samples for building a training model. The well logging data includes well logging simulation data and well logging big data. The well logging simulation data is the well logging data obtained through various existing well logging devices; and the well logging big data is historical data containing the rules and experience of well logging data. Therefore, Figure 2 , the acquisition of logging data includes the following steps:

[0037] Step S101: Acquire well logging simulation data.

[0038] Specifically, the logging simulation data is the logging data obtained through various existing logging devices, electro-acoustic nuclear logging simulation, including different dual lateral instruments, array lateral instruments, induction logging instruments, array induction logging, micro-spherical focusing logging instruments, through-casing resistivity logging instruments, etc. It can realize instrument parameter setting, formation model setting, response digitization, etc., and can obtain the instrument constants (geometric factors), surrounding rocks, mud invasion, borehole effects and other plates of electric logging and response libraries. The response library contains the linkage effect of parameter changes and test results, and has a strong regularity. For example, in induction logging, the established formation model is a five-layer symmetrical formation model. When performing numerical simulation, the remaining formation parameters are kept unchanged, and only the observed target layer parameters are changed. The study uses the variable retention method to change only one parameter at a time. The changed parameters include mud filtrate resistivity, well diameter, invasion zone resistivity, invasion zone radius, target layer thickness, etc. Each time the calculation is completed, the corresponding data in csv format is exported from the simulation platform, and the induction logging response library is established by changing the formation parameters, wellbore and mud parameters. Based on the above change rules, each time a parameter is changed, the logging response library corresponding to each changed parameter is obtained, and the logging response library containing these linkage relationships is used as the logging simulation data.

[0039] Step S102: Acquire historical logging data.

[0040] Specifically, logging technology has evolved from the initial point logging hand-drawing, to photoelectric recording pen, to tracing and printing, 716 format tape recording, to the current disk recording. There are various logging data formats such as interpretation report drawings, data collected by different logging service companies, video and voice data. These historical data stored in different formats contain a large number of logging data rules and have great reference value. Time-lapse production logging data including casing resistivity logging, logging data, core and formation water analysis data, well testing, logging while drilling, logging data, and post-correction and processing interpretation data. New logging sensor technology streams large-capacity, multi-scale and high-dimensional rock physical data into the database in real time. Rock physical logging data such as logging digital curves, array data, logging images, text and tables can all be collected with depth (continuous or discrete) or time correspondence. Logging data also has multi-scale characteristics, for example, core scanning data has a scale of microns to centimeters, while electroacoustic imaging data has a scale of millimeters to centimeters. Similarly, the detection depth range is also obviously different, for example, microresistivity, acoustic wave, density, neutron and gamma are in the centimeter scale, while lateral type is in the 1 meter scale, induction type is in the 2 meter scale, through casing resistivity logging is in the 10 meter scale, and interwell acoustic wave and electromagnetic are in the range of hundreds of meters to 1 kilometer. The above data have both volume attributes and big data characteristics such as diversity, variability, visualization and value, which are the sources of historical logging data.

[0041] After obtaining the logging data, it is necessary to build a machine learning model. First, training begins with simulation data, because the simulation platform can provide an evaluation value of the resistivity of the formation near the well. In the first iteration, the apparent resistivity and the true resistivity of the formation have a large error. Then, the simulation data and logging big data are trained to find the eigenvalues. Based on the eigenvalues, the self-decoding technology of the deep neural network is used to train the data again to obtain the model.

[0042] Step S20: collecting well logging measurement values, and preprocessing the measurement values ​​according to preset rules to obtain training sample data.

[0043] Specifically, the logging measurements are first collected, including resistivity, mud resistivity, borehole diameter, formation thickness, surrounding rock resistivity, invasion zone diameter, invasion zone resistivity and measuring instrument parameters. Represented by n samples related to this parameter. For example, in the borehole resistivity example, each variable x contains n samples of each specific geophysical quantity, such as resistivity, mud resistivity, borehole diameter, formation thickness, surrounding rock resistivity, invasion zone diameter, invasion zone resistivity, and instrument parameters. Each dimension corresponds to a specific value (sample) of the variable. From a linear algebraic point of view, the variable x represents a row of the matrix M or P. After obtaining the collected data, preprocessing is required. Taking resistivity testing as an example, specifically, Figure 3 , including the following steps:

[0044] Step S201: Perform a logarithmic transformation of resistivity.

[0045] Specifically, the resistivity conversion formula is:

[0046]

[0047] For some geophysical variables (such as resistivity), this change in the variable ensures that relative errors of equal size correspond to absolute errors of similar size. Thus, this change in the variable allows us to perform local (within the variable) comparisons.

[0048] Step S202: Delete outlier samples.

[0049] Specifically, in practice, outlier measurements are usually shown in the sample database. These outliers appear due to measurement errors or physical property errors. These outliers need to be filtered to avoid affecting the subsequent training results. For example, in borehole resistivity measurements, some apparent resistivity measurements are close to infinity, resulting in trumpet-shaped patterns in the well log. When a particular variable xi in any i-th sample, then the entire sample should be removed. Otherwise, the outlier measurement will affect the entire minimization process problem, leading to poor numerical results. The removal process may be automatic using statistical indicators, or determined by the user based on prior physical knowledge about the problem.

[0050] Step S203: Linear scale transformation.

[0051] Specifically, we introduce a linear scaling mapping to the interval [0; 1). We choose this interval because it is a normally (or uniformly) distributed variable with unit length and mean X equal to 1. Let x min is the minimum value of the column feature, x max is the maximum value of the column attribute. It is defined as:

[0052] Rlin(x)=[(x1-x min ) / (x max -x min ),…,(x n -x min ) / (x max -x min )]

[0053] For all possible approximations, the limit x min and x max is fixed. This variation of variables allows for a global comparison of the errors corresponding to different variables, since they all take values ​​in the same interval. Depending on the size of the training dataset and the number of iterations required for convergence, this can result in millions of forward function evaluations. Solving such a large number of forward problems is very time consuming. Even with 1.5D pattern matching simulations. Moreover, most forward solvers are implemented on CPU architectures, while training DNNs normally happens on GPUs. This requires both a GPU and a CPU, which further slows down the training process.

[0054] Step S30: Perform training on the training sample data based on the machine learning model and output calculation results.

[0055] Specifically, preferably, in order to overcome the slow forward calculation and the influence of the simulation method, one DNN network is used to simulate the approximate forward function, and the other is the inversion function, and the loss function is used to build an automatic decoder. Use Encoder-Decoder for model training. The training data is divided into three parts, including training data, verification data and test data. For example, 80% of the samples are used to train the DNN, 10% are used for verification, and the remaining 10% are used for testing. Consider two DNN networks to obtain approximate forward model F and inversion model function I respectively. The forward function F is a well-posed continuous function, and the inversion function I is also suitable for continuous functions. DNN is implemented using Tensorflow 2.0 and Keras libraries or Pytorch. The main applications of machine learning classification and regression algorithms in well logging technology include lithology classification, well logging data repair, well logging formation parameter inversion, formation shear wave extraction, and virtual well logging curves. Classification is mainly used for well logging lithology identification for well logging technology. The function of the classification tree regression tree machine learning algorithm is:

[0056] Y new =predict(tree,X new )

[0057] Among them, Xnew is the new data, predict is the function to be trained and learned, and tree is the selected tree classification method. Cross-validation is used to better understand the predictive accuracy of the tree on new data. Typically, cross-validation splits the training data randomly into 10 parts. It trains 10 new trees, each with 9 tree data. Then, it checks the predictive accuracy of each new tree on the data that was not included in the training tree. This method gives a good estimate of the predictive accuracy of the generated tree because it tests the accuracy or precision of the new tree on new data.

[0058] Preferably, training is also performed based on an enabling learning method, which is to fuse the results of many weak learners into a high-quality overall predictor. These methods can have the same syntax as , so different methods can be tried to explore the most suitable method with the best accuracy according to the needs. Its function is:

[0059] ens=fitensemble(X,Y,model,numberens,learners)

[0060] Where X is the data matrix with each row containing an observation and each column containing a predictor variable. Y is the response with the same number of observations as rows in X. model is the type of method used. The number refers to the number of weak learners in ENS per feature. Therefore, the number of elements of ens is the number of elements in learners. Methods for evaluating the quality of learning. First evaluate the ensemble on an independent test set, or you can evaluate the ensemble by cross-validation. Bagging is a method to put weak learners (such as decision trees) on the dataset and obtain each bootstrap copy with replacement by randomly selecting N observations, where N is the dataset size. Find the maximum correlation coefficient and the minimum correlation coefficient in the deep neural network iteration. Improve the deep neural network input to reduce the cost function error. Set the absolute value of the cost error to 0.001 ohm-meter and terminate the iteration. The calculated apparent resistivity is considered to be the true resistivity of the formation.

[0061] Preferably, the machine learning model includes: a logging value normalization model, a rock physics prediction model, a complex lithology identification model, an oil and gas reservoir monitoring model and a mechanical evaluation model.

[0062] Step S40: performing corresponding formation evaluation according to the calculation results.

[0063] Specifically, after obtaining accurate calculation results, the calculation results are obtained based on the comprehensive considerations of each machine learning model, and based on the calculation results, the formation evaluation results corresponding to the machine learning model requirements can be obtained, such as whether there is complex lithology, mechanical evaluation results, etc.

[0064] Embodiment 1:

[0065] Application method of logging value normalization model. It is generally believed that logging signals are composed of formation signals, random noise and systematic errors. Well logging normalization is the process of applying correction offsets to logging values ​​to minimize systematic errors. Machine learning methods can solve the problem of multi-well logging value normalization. This method is particularly suitable for hundreds of wells with different data quality and years. The predictive machine learning model is trained using logging data from reference cells in reference wells, and a regression-based optimization algorithm is used to solve the constant offset, which is then used as a normalization correction for the logging of the remaining wells. The machine learning algorithm using natural gamma as depth matching can be used to match other logging values ​​in the same well for depth matching, which is achieved through a supervised neural network machine learning algorithm. This lays a quality foundation for improving the accuracy of well logging interpretation.

[0066] Embodiment 2:

[0067] Application method of rock physics prediction model. When applying rock physics images, pixels are first normalized, which includes unified measurement of pixel parameters, scale value size and resolution. Deep learning methods can not only classify rock images, but also estimate rock physical parameters. Different descriptors are identified directly from 3D images used as input. Convolutional neural network modeling and supervised training are performed using TensorFlow deep learning. The trend of porosity changes verifies that there is good consistency between the training dataset and machine learning test results, which illustrates the potential of deep learning methods in core characterization and can be applied to rock physics databases based on core scan images on a large scale in the future. Rapid prediction of permeability from images directly through image recognition neural networks is a new pore-scale modeling method with great potential. The neural network method was trained for porosity and water saturation using available logging data, which can accurately determine porosity and water saturation, especially when the experimental core test values ​​of argillaceous sandstone reservoirs are used as constraints, and the established correlation works well.

[0068] Embodiment 3:

[0069] Application method of complex lithology identification model. Although tight sandstone has great potential for exploration and development, its complex reservoir composition is becoming more and more complex. At present, the convenient interpretation of well logging faces the problem of more difficult identification of lithology, evaluation and monitoring of reservoirs, and prediction of production. The conventional lithology logging curves that only use the characteristics to identify the lithology of tight sandstone in shale, gypsum, limestone and dolomite have not achieved good results because the intrinsic logging response signal-to-noise ratio is low. Machine learning technology can use more characteristic curves to identify lithology, and regression can be used to repair the missing data in bad wells. Among many characteristic curves, AC (acoustic logging curve), SP (spontaneous potential logging curve), GR (natural gamma logging curve) and CAL (caliper logging curve) are used to train oil well data. Decision tree, support vector machine, KNN, Ensemble, deep neural network and other methods are used for optimization learning. The results show that the training accuracy of bagged trees trained by ensemble method is almost 98.5% and the prediction value is 94.3%, especially, they have excellent prediction performance in a wide range of porosity and pore geometry (especially large pores). The combined neural network classifier and ensemble method provide more satisfactory results for well logging classification and prediction problems.

[0070] Using geophysical logging data to apply machine learning to accurately identify lithology, the best classifier is a neural network, followed by support vector machines and k-nearest neighbor methods. When using multiple curves to classify reservoir quality, the first group uses GR, AC, CNL (neutron logging curve), DEN (density logging curve), RD (deep lateral resistivity), RS, RD / RS, and RD-RS for learning; the second group uses GR, AC, CNL, DEN, RD, and RS for learning; the third group uses GR, AC, CNL, RD, RS, RD / RS, and RD-RS for learning; the fourth group uses GR, AC, CNL, RD, and RS; and the fifth group uses GR and AC for learning. The application of different eigenvalues ​​and machine learning accuracy are shown in Table 1.

[0071] Group\Algorithm Classification tree (%) KNN(%) Ensemble learning (%) Number of characteristic curves 1 63.8 69.7 89.4 8 2 61.3 73.9 73.9 6 3 66.1 80.5 80.5 7 4 58.6 76.3 76.3 5 5 68.4 72.1 72.1 2

[0072] Table 1 Comparison of reservoir classification results under various logging curves and their learning

[0073] Embodiment 4:

[0074] The application method of oil and gas reservoir monitoring model, expanded from single well to gas reservoir, can effectively understand the law of gas saturation change of this gas reservoir, thereby reducing development risks. Machine learning can solve the problem of updating 3D static model and reservoir parameter prediction. The logging data of open hole and cased well are standardized and depth matched. Digital core technology is applied. According to the machine learning correlation analysis of logging lithology curve, the Archie gas saturation calculation method with variable m and n that changes with sedimentary phase is established. The data of different time periods are measured in the cased well to realize the dynamic monitoring of gas reservoir, and then realize the production capacity prediction in time series.

[0075] Machine learning based on cloud computing platform solves gas reservoir monitoring with lower operating cost, more accuracy and faster speed, and its implementation process. Data-driven machine learning predictive analysis has been applied to gas reservoir monitoring and management with large amounts of data. These technologies include K nearest neighbor (KNN), support vector machine regression, kernel ridge regression, random forest, adaptive boosting (Adaboost), deep neural network, etc. The feasibility of using machine learning methods to predict 4D reservoir physical property changes with time series. Combined with acoustic and electrical simulation algorithms and through-casing electroacoustic logging data, dynamic reservoir monitoring and prediction of oil well production capacity and life can be achieved. The vertical gas saturation change distribution results of reservoir monitoring, in which the yellow part is the distribution of gas thickness and saturation in multiple wells in the vertical direction. After production testing, high-yield gas layers were finally found in the first and second layers of well L48 in the first and third gas reservoirs, with a daily gas production of 32.84 million cubic meters and 3.3 cubic meters of water; there are gas-water co-producing layers in the three layers of well L172 in the first, second and third gas reservoirs, with a daily gas production of 31,000 cubic meters and 3.4 cubic meters of water.

[0076] Embodiment 5:

[0077] Mechanical evaluation model application method, the application of machine learning methods can predict shear waves, which are difficult to obtain in soft formations. Machine learning methods can be applied to predict shear wave velocity from various conventional well logging curves, but this method uses several other well logging measurements in addition to the longitudinal velocity log as the feature value, such as natural gamma, density, neutron, resistivity, porosity and saturation well logging curves. It is to train and verify the machine learning model on an existing well logging data set with shear wave velocity. The feature selection process is used to highlight which well logging curves are good predictors of shear velocity (VS). Various regression models are then trained, and the predicted values ​​are compared with the actual values ​​of various models through their root mean square error (RMSE), and the model with the smallest RMSE is selected. Another well in the data set is predicted and used as a validation set. This method will provide higher accuracy in the absence of shear sonic logging and is particularly important for geomechanics, rock physics and other applications. Using regression predictions in machine learning, the shear wave data of the expansion section is repaired for evaluating the anisotropy results of tight gas sandstone reservoirs, see Figure 2 The reservoir anisotropy index indicates that the shear wave velocity varies with direction due to mineral particles, stratification, cracks, stress, etc. in different directions of the reservoir. It is measured by the difference between fast and slow shear wave velocities and can be defined by the following formula:

[0078]

[0079] Among them, S1 is the slow shear wave time difference, S2 is the fast shear wave time difference, and △S = S1-S2. Poisson's ratio and Young's modulus are two parameters necessary for evaluating the brittleness of shale or tight sandstone reservoirs, which can be estimated from dipole acoustic logging and volume density logging. The results of the study show that shear wave slowness is related to the mud or shale content value. Deep neural networks can associate shear logging data with other logging values. Especially in shale reservoirs with clastic rocks, due to thin laminations, the reservoirs show anisotropy. When the wellbore conditions are uneven, there is complex mud rheology and hot gas influx inside the wellbore, it is difficult to obtain reliable horizontal shear wave time differences. In order to eliminate the uncertainty of obtaining horizontal shear slowness from Stoneley waves, the random forest algorithm in the ensemble learning method has greater potential for prediction. In the regression algorithm, the dry weight ratio of calcite, dolomite, quartz, and illite, total porosity, permeability, gamma rays, and longitudinal wave time difference are used as characteristic parameters to better obtain the shear wave parameters.

[0080] In one possible implementation, the scheme of the present invention is also used for simulation. For electroacoustic logging, this type of modeling simulation usually requires a long computing time, as well as a large amount of computer resources and manual intervention. Machine learning can be used to replace the lengthy and expensive model building process. Well logging simulation requires borehole size, number of casings, presence of cement, fluid parameters, lithology, porosity and fluid type parameters in the pore space for instrument response modeling. Under existing hardware conditions, there are problems such as low computing efficiency and difficulty in analyzing response results. Machine learning can replace this process, using existing simulation results to learn, predict unknown wellbore and formation conditions, apply various machine learning algorithms to these existing simulation data, and use the automatic decoder of deep neural networks to generate forward prediction models. When comparing the predictions of the machine learning model with the simulated calculation data in the validation data set, the accuracy can reach more than 90%.

[0081] The embodiment of the present invention further provides a computer-readable storage medium, on which instructions are stored, which, when executed on a computer, enable the computer to execute the above-mentioned formation evaluation method based on well logging big data.

[0082] Those skilled in the art will understand that all or part of the steps in the method for implementing the above-mentioned embodiments can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium, including several instructions for making a single-chip microcomputer, a chip or a processor (processor) perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0083] The optional embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the technical concept of the embodiments of the present invention, the technical scheme of the embodiments of the present invention can be subjected to a variety of simple modifications, and these simple modifications all belong to the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the embodiments of the present invention will not further describe various possible combinations.

[0084] In addition, various embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the embodiments of the present invention, they should also be regarded as the contents disclosed in the embodiments of the present invention.

Claims

1. A formation evaluation method based on well logging big data, the method is implemented based on physical simulation and machine learning, characterized in that: The method comprises: Acquire well logging data, and perform machine learning model training based on the well logging data to obtain a machine learning model; wherein, The obtaining of logging data comprises: Acquire logging big data including logging simulation data and logging history data; The step of obtaining the well logging simulation data comprises: Obtaining parameters that affect the logging results, and obtaining a linkage relationship between the changed parameters and corresponding parameters of the observed target layer by a variable retention method; the changed parameter is one of the parameters that affect the logging results; The parameters affecting the logging results include at least: mud filtrate resistivity, well diameter, invasion zone resistivity, invasion zone radius, and observation target layer thickness; Changing the parameters affecting the logging results one by one to obtain linkage relationships between the multiple changed parameters and corresponding parameters of the observed target layer; According to the linkage relationship between the plurality of changed parameters and the corresponding parameters of the observation target layer, a well logging response library is established as the well logging simulation data; Collecting well logging measurement values, and preprocessing the measurement values ​​according to preset rules to obtain training sample data; Perform training on the training sample data based on the machine learning model and output calculation results; Corresponding formation evaluation is performed according to the calculation results.

2. The method according to claim 1, characterized in that The machine learning model includes: Well logging value normalization model, rock physics prediction model, complex lithology identification model, oil and gas reservoir monitoring model, and mechanical evaluation model.

3. The method according to claim 1, characterized in that The obtaining of the logging big data includes: All historical data related to well logging stored in the actual production process are extracted as the well logging historical data.

4. The method according to claim 1, characterized in that: The logging measurements include one or more of the following: Resistivity, mud resistivity, borehole diameter, formation thickness, surrounding rock resistivity, invasion zone diameter, invasion zone resistivity and measuring instrument parameters.

5. The method according to claim 4, characterized in that The preprocessing of the measured values ​​according to a preset rule includes: performing a logarithmic transformation of the measured values; Remove outlier samples from the log-transformed measurements; Perform linear scale conversion of measured values.

6. The method according to claim 1, characterized in that The performing training of the training sample data based on the machine learning model and outputting calculation results includes: Dividing the training sample data into three parts according to a preset ratio, namely training data, verification data and test data; Model training is performed according to the three divided parts of data, wherein the model training process is implemented based on an Encoder-Decoder model, a classification tree regression tree machine learning algorithm, and an enabling learning method; Output the calculation results.

7. A formation evaluation system based on well logging big data, characterized in that: The system comprises: The acquisition unit is used to obtain well logging data and to collect well logging measurement values; wherein, The obtaining of logging data comprises: Acquire logging big data including logging simulation data and logging history data; The step of obtaining the well logging simulation data comprises: Obtaining parameters that affect the logging results, and obtaining a linkage relationship between the changed parameters and corresponding parameters of the observed target layer by a variable retention method; the changed parameter is one of the parameters that affect the logging results; The parameters affecting the logging results include at least: mud filtrate resistivity, well diameter, invasion zone resistivity, invasion zone radius, and observation target layer thickness; Changing the parameters affecting the logging results one by one to obtain linkage relationships between the multiple changed parameters and corresponding parameters of the observed target layer; According to the linkage relationship between the plurality of changed parameters and the corresponding parameters of the observation target layer, a well logging response library is established as the well logging simulation data; A model building unit, used to perform machine learning model training according to the well logging data to obtain a machine learning model; A processing unit, used to perform preprocessing of the measured values ​​according to preset rules to obtain training sample data; A training unit, used to perform training on the training sample data based on the machine learning model and output a calculation result; The processing unit is also used to perform corresponding formation evaluation according to the calculation results; The model building unit, the processing unit and the training unit are all built based on a private cloud computing platform.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the formation evaluation method based on well logging big data as described in any one of claims 1 to 6.

Images