CN-121998140-A - Oil reservoir physical property intelligent prediction method based on sand body superposition mode

CN121998140ACN 121998140 ACN121998140 ACN 121998140ACN-121998140-A

Abstract

The invention discloses an intelligent prediction method of oil reservoir physical properties based on a sand body superposition mode, which relates to the technical field of electric digital data processing, and comprises the following steps of establishing an oil reservoir grid model, preprocessing existing data, generating a sand body superposition mode library, extracting a well bypass statistical wavelet, generating a sand body superposition mode sample forward, constructing a correlation diagram and an index diagram, screening a high-quality training sample, training a neural network model, predicting oil reservoir physical properties, randomly generating the sand body superposition modes with different geological features by solving disturbance functions under different geological constraint conditions, and combining a forward performance technology and a deep learning technology to realize accurate estimation of the oil reservoir physical properties based on geology, logging and earthquake, wherein the new technology path is more in line with actual conditions, and the accuracy of the estimation of the oil reservoir physical properties is improved.

Inventors

ZHANG YUXIAO
YANG HONGWEI
HOU QINGJIE
WANG YANAN
LI YINLAN
SONG XINGKUN
PAN JULING

Assignees

中国石油化工股份有限公司
中国石油化工股份有限公司胜利油田分公司

Dates

Publication Date: 20260508
Application Date: 20241101

Claims (10)

1. An intelligent prediction method for physical properties of oil reservoirs based on a sand body superposition mode is characterized by comprising the following steps, Step 101, establishing an oil reservoir grid model, resampling a well logging curve, a sand body label and seismic data based on the oil reservoir grid model, and correspondingly obtaining the well logging curve, the sand body label and the oil reservoir seismic attribute under the model scale at a well point; 102, randomly selecting a disturbance function, randomly generating the starting depth and disturbance points of a disturbed sand-mud boundary based on sample data obtained in the step 101, and resampling the disturbance function according to a set sampling interval by solving the disturbance function, and carrying in a mapping function from a resampling value to a logging curve to generate each sand stacking mode sample to form a sand stacking mode library, wherein the sand stacking mode samples do not contain amplitude data; step 103, extracting and obtaining logging and well bypass data from the complete sample data in the step 101, and optimizing an initial Rake wavelet by adopting a random gradient descent method to obtain a well bypass statistical wavelet; Step 104, forward modeling is carried out on the sand body stacking mode sample generated in the step 102 by adopting the well bypass statistical wavelet in the step 103 to obtain a sand body stacking mode sample after forward modeling, and a sand body stacking mode library after forward modeling is formed; Step 105, based on the forward sand body superposition model library obtained in step 104, extracting and obtaining the amplitude data of each sample, based on the oil reservoir seismic attribute obtained in step 101, extracting and obtaining the longitudinal amplitude data at each channel, carrying out correlation analysis on the two amplitude data obtained by extraction, recording the correlation coefficient and index value of the high correlation value sample amplitude data, and traversing all channels to form a correlation graph and an index graph; Step 106, obtaining a correlation coefficient threshold value, and selecting sand body superposition mode samples above the correlation coefficient threshold value from the correlation diagram obtained in step 105 to form a high-quality training sample; Step 107, training the neural network until convergence by using the high-quality training sample obtained in the step 106 as a training set to obtain a trained neural network; Step 108, inputting the reservoir seismic attribute obtained in step 101 into the trained neural network obtained in step 107, and outputting the predicted reservoir physical property attribute by the neural network.
2. The intelligent prediction method of oil reservoir physical properties based on the sand body stacking mode according to claim 1, wherein in the step 101, resampling a well logging curve according to the space coordinates of an oil reservoir grid model to obtain a well logging curve of a model scale at well points, labeling well logging sand interpretation data, obtaining sand labels under the model scale according to the space coordinates at the well points of the oil reservoir grid model, and resampling seismic data according to the space coordinates at the well points of the oil reservoir grid model to obtain oil reservoir seismic attributes.
3. The intelligent prediction method of oil reservoir physical properties based on a sand body stacking mode according to claim 1, wherein in the step 102, the disturbance function is a linear disturbance function or a nonlinear disturbance function, the depth sampling interval is 0.125 m, the mapping function is a logging curve obtained by extracting from the sample data obtained in the step 101, the logging curve comprises acoustic wave, density, porosity and corresponding space coordinate data, and a piecewise function approximation method is used for establishing a mapping function from space coordinates to acoustic wave, density and porosity, namely formula (1.1); Y=S(z)(1.1) In the formula (1.1), z is a depth coordinate corresponding to a geologic body along a well track, S (-) is a mapping function from depth coordinate data to sound wave, density and porosity data, and Y is a numerical value corresponding to sound wave, density and porosity; The linear disturbance function formula is: Z=f(Z)=aZ′+b (1.2) the formula of the nonlinear disturbance function is: Z=f(Z')=aZ' 2 +bZ'+c(1.3) In the formula, Z is the depth coordinate of the original geologic body, and Z' is the depth coordinate of the geologic body after the position of the geologic body is changed.
4. The method for intelligently predicting physical properties of a reservoir based on a sand stacking mode as recited in claim 3, wherein in said step 102, When the sand body stacking mode sample is copied, a linear disturbance function, namely a formula (1.2), is adopted, the space coordinates of the geologic body are kept unchanged before and after, and under the condition that the starting depth of the sand-mud boundary point is kept unchanged, the linear disturbance function is solved, so that the process of copying the sand body stacking mode sample data is realized; when the thickness of the sand bodies in the stacking mode is uniformly changed, a linear disturbance function, namely a formula (1.2), is adopted, and the disturbance function is solved by changing the starting depth of the boundary points of the sand and mud, so that the uniform change process of the thickness of the sand bodies in the stacking mode is realized; when a sand body is newly added in the existing stacking mode, a nonlinear disturbance function, namely a formula (1.3), is adopted, and a disturbance point is randomly generated in the original boundary point by changing the starting depth of the sand-mud boundary point, so that the disturbance function is solved, and the process of newly adding the sand body in the sand body stacking mode sample is realized; When deleting a section of the sand body in the stacking mode, adopting a nonlinear disturbance function, namely a formula (1.3), and taking the depth value of the disturbance point as the outside of the stacking mode depth range in the process of randomly generating the disturbance point in the original boundary point, so as to realize the removal of the section of the sand body.
5. The intelligent prediction method of oil reservoir physical properties based on sand body stacking mode according to claim 1, wherein in step 103, well bypass amplitude, sound wave and density data are extracted from the complete sample data, and reflection coefficient sequence data are obtained according to the sound wave and density data, wherein the formula is as follows: In the formula (1.4), ρ i represents the density of the i-th layer medium, the unit is g/cm 3 ,v i represents the speed of the i-th layer medium, which is the inverse of the acoustic wave, the unit is m/s, and r i represents the reflection coefficients of the i-th layer and the i+1th layer; In the wavelet extraction process based on data driving, rake wavelets are used as convolution kernel initial parameters, and a function from the reflection coefficient to the amplitude is constructed according to a convolution formula: in the formula (1.5), s t is amplitude data, r t is a reflection coefficient obtained previously, N represents the length of the wavelet ω τ , ω τ is wavelet data to be extracted, an initial value is a Rake wavelet, and the formula is as follows:
6. The method for intelligently predicting physical properties of a reservoir based on a sand stacking mode according to claim 5, wherein in step 103, an initial result is optimized by adopting a gradient descent method with an actual coarsening well bypass as a result constraint based on a data-driven wavelet extraction process; The gradient descent method is as follows: In the formula (1.7), ω t represents the value of the wavelet at the t-th iteration, η represents the learning rate LEARNING RATE, Representing an objective function The first derivative of wavelet parameter omega is gradient, and the optimized wavelet is obtained by iteratively updating wavelet parameters.
7. The intelligent prediction method according to claim 5, wherein in step 104, the sample of the sand stacking pattern generated in step 102 lacks amplitude data, and the obtained sample data is complete sample data by forward modeling the amplitude data according to the equation (1.4) and the equation (1.5) using the statistical wavelet obtained in step 103, and after this step, each sample contains a log, a sand label, and the amplitude data in the constructed sand stacking pattern library.
8. The method for intelligently predicting physical properties of a reservoir based on a sand stacking mode as set forth in claim 1, wherein in said step 105, a formula for calculating a correlation coefficient is as follows: in the formula (1.8), X and Y are respectively amplitude data of stacked samples and longitudinal amplitude data at each channel of oil reservoir seismic attribute; And (3) approximately representing the actual sand body superposition condition of the corresponding channels by the superposition samples with high correlation amplitude, recording the correlation coefficient of the superposition samples, the index in the sand body superposition mode library and the corresponding coordinates, and repeating the process until all the channels are calculated to form a correlation map and an index map.
9. The intelligent prediction method of physical properties of oil reservoirs based on a sand stacking mode according to claim 1, wherein in the step 106, the correlation coefficient threshold is 0.9, points with correlation coefficients greater than 0.9 are obtained by screening from a correlation map, and sample data of the corresponding points in a sand stacking mode library, namely high correlation coefficient samples, are found by combining an index map, so that high-quality training samples which embody actual complex geologic body structure information are formed.
10. The intelligent prediction method of oil reservoir physical properties based on a sand stacking mode according to claim 1, wherein in the step 107, the neural network is a deep neural network, in the training process, amplitude data is extracted from a sand stacking sample as input of the neural network, physical property parameters are extracted from the sand stacking sample as labels, and a mean square error loss function is selected: in the formula (1.9), N is the output length of the neural network, The label sequence is a label sequence, and y is an output sequence of the neural network; A random gradient descent method is selected as an optimizer.

Description

Oil reservoir physical property intelligent prediction method based on sand body superposition mode Technical Field The invention relates to the technical field of electric digital data processing, in particular to an intelligent oil reservoir physical property prediction method based on a sand body superposition mode. Background In the geophysical oil reservoir industry, three-dimensional physical property prediction is a key work, and physical property parameter estimation is generally carried out by adopting a geostatistical method. However, conventional geostatistical methods have problems in modeling geologic body features with complex geologic structures. Such as two-point geostatistics and multi-point geostatistics, which describe geologic volume structural information using a variational function and a combination pattern of points, respectively. However, their effect in certain scenarios is not ideal. Currently, deep learning methods have made breakthrough progress in many fields, which are good at dealing with complex nonlinear problems and have a great generalization capability. The physical property prediction method based on deep learning can synthesize logging hard data and seismic soft data to obtain a high-resolution three-dimensional physical property estimation model. However, the deep learning method requires a large number of high-quality samples with labels in the supervised learning process, and the data acquisition cost in the geophysical exploration industry is high, so that the number of samples is insufficient, and the development of the deep learning method is hindered. The application publication number is CN116629023A, and the name is a novel three-dimensional oil reservoir characterization method integrating stratum forward modeling and geostatistical science. The method comprises the steps of taking an SFM mudstone model as a constraint condition, establishing a lithology model calibrated under the constraint of the SFM mudstone model by adopting a geostatistical method, dividing a sedimentary microphase according to a rock core and a logging curve, obtaining a sedimentary microphase plane subsection map according to well connection sedimentary phase analysis, converting the sedimentary microphase plane subsection map into a sedimentary phase three-dimensional trend body, and obtaining a sedimentary microphase model according to the three-dimensional trend body as the constraint condition for a geological modeling process. The application publication number is CN113589363A, and the name is a new oil gas prediction method integrating an artificial neural network and geostatistics. According to the novel oil gas prediction method integrating the artificial neural network and the geostatistics, the stratum sections of the earthquake and the well logging data are averaged, and Kendall correlation analysis of the earthquake and the well logging data is carried out. And selecting a proper curve to train an artificial neural network at a well point, then carrying out seismic logging and geological statistics to obtain three-dimensional distribution of physical parameters of a reservoir, and training a second artificial neural network at the well on the basis of the three-dimensional distribution to carry out three-dimensional prediction of oil gas. Application publication number CN107316341a, entitled a multipoint geostatistical depositional phase modeling method. The method is based on the original multipoint geostatistical sedimentary facies modeling method, provides a concept of geometric factors, adopts a geometric factor limiting space to obtain conditional probability reflecting the scale of a sedimentary facies, and adopts a method of correcting proportion identity to correct a conditional probability distribution curve obtained by scanning a training image, thereby improving the original multipoint geostatistical method. Modeling implementation by using an improved algorithm shows that the multi-point geostatistical method for introducing the geometric factors inherits the advantages of the traditional multi-point geostatistical method for reproducing the geometric form of the sedimentary facies, and overcomes the defects of the method in aspects of processing the facies scale, continuity and the like. The method is based on multipoint geostatistics, and the selection of training images affects the final result. The method comprises the steps of establishing five river-phase sand superposition mode geological conceptual models with different forms, forward modeling the five models to screen out seismic waveforms and attribute information as template data, carrying out superposition mode identification on sand of a target interval on a seismic section by using an identification algorithm and a template to obtain mode distribution of the target sand, and selecting logging data in the same mode to participate in the modeling process according to different sand modes obtained by