CN113985493B - Intelligent modeling method for underground multi-information constrained isochronous stratum grillwork - Google Patents

Abstract

The invention provides an intelligent modeling method for an underground multi-information constrained isochronous stratum grid, which is characterized in that high-precision small-layer division results are obtained by utilizing an artificial intelligence method through comprehensive geological information and logging information, and inter-well horizons conforming to geological laws are obtained by means of seismic information based on the results, so that the small-layer point-to-surface division process is realized, and finally, the high-precision isochronous stratum grid model can be obtained. The invention can accelerate the convergence of the neural network model, refine the deposition evolution analysis of complex and other surfaces; the working efficiency can be greatly improved through an automatic procedure under the condition of more wells, and the intelligent small-layer division precision can be effectively improved through the method.

Description

Intelligent modeling method of underground multi-information constrained isochronous stratigraphic framework

Technical field

The invention belongs to the technical field of geological data processing, and specifically relates to an underground multi-information constrained isochronous stratigraphic framework intelligent modeling method.

Background technique

The establishment of traditional isochronous stratigraphic frameworks is based on fine stratigraphic layers. There are two main methods for determining the origin of small layers:

The combination of well and seismic data determines the division of large layers (oil group interface), and then subdivides small layers based on well logging and sand body information. Finally, the division results are adjusted using the reflection characteristics of seismic data; for example, Reference 1, Huo Chunliang, Gu Li, Zhao Chunming , Yan Weipeng, Yang Qinghong. Reservoir fine modeling based on comprehensive integration of seismic, well logging and geology [J]. Acta Petroleum Sinica, 2007(06): 66-71. Proposed the idea of "cycle comparison, hierarchical control": ① Well Seismic combination determines the oil group interface and establishes a "coarse" isochronous stratigraphic framework. ② Taking advantage of the high vertical resolution of logging data, the traditional equal-elevation comparison method and the logging curve similarity comparison method are used to subdivide well points into layers using sand bodies as units; on the basis of well seismic calibration, The seismic inversion data is used as the background to construct well-connected profiles; then the comparison layers are adjusted based on whether the inter-well comparison results are consistent with the seismic reflection characteristics, and a fine stratigraphic framework is established.

The division of small layers is obtained through interpolation based on the oil group layer and cycle characteristic information. For example, reference 2, Gao Boyu, Sun Lichun, Hu Guangyi, Zhang Yuan. Discussion on geological modeling method of fluvial facies reservoir based on single sand body [J]. China Offshore Oil and Gas, 2008(01):34-37., Utilization For well point information, apply the principles of sequence stratigraphy to divide the stratigraphic strata and divide the lower section into four oil groups; then perform well calibration on the seismic data through the time-depth relationship of a single well (such as VSP logging, etc.) to establish seismic reflection characteristics. Corresponding to the well log curve, each isochronous sequence interface is interpreted on the seismic reflection profile. Corresponding to high-resolution sequence stratigraphy, these isochronic interfaces should be at least medium-term base level cycle interfaces; finally, according to the cycle The superposition model and stratigraphic development characteristics in the work area are divided into secondary cycles and compared, and the top trend constraints of the four oil groups are used to interpolate the small layer levels, and the relationship between faults and strata is processed in the form of corner grids to generate structures. Grid model. During the structural modeling process, it is fully ensured that the structural grid reflects the stratigraphic pattern, and the longitudinal grids with the same serial number have isochronous significance.

The method of dividing small layers in Reference 1 makes it difficult to determine the trend of layers between wells. If you encounter a target area with few wells and there is no effective layer information between wells, it is difficult to establish accurate isochronous small layers according to this method ( The longitudinal resolution scale of seismic is much larger than that of well logging, so at the small layer scale, it is impossible to divide small layers only through seismic reflection characteristics); conversely, if multiple well target areas are encountered, this method needs to conduct Manual layer division has high time and labor costs, and there are many human subjective factors, which affects the accuracy of the division.

In Reference 2, large-scale layer divisions (reservoir groups) are directly based on seismic data and well logging information. The small layers are obtained based on interpolation. This result is obviously insufficient in accuracy because the large layers and small layers are on the geological plane. Distribution trends (referring to fluctuations in depth) are highly inconsistent due to factors such as stratigraphic lithology and geological structure, and even erroneous layer division results may occur. At this time, manual correction is required, which greatly increases work costs.

Therefore, the fine layers obtained through the traditional modeling process cannot obtain the results of small layer division very well, or it requires a lot of manpower, material resources and time costs to complete, which is inefficient; at the same time, there is not yet a good technology to obtain the interwell layers. Distribution, that is, how to convert the layers on each well into the two-dimensional plane of the entire target work area in a point-to-plane manner. The current processing method is generally to divide larger-scale sand group layers based on well data or seismic data. Then an isochronous stratigraphic framework model is established through interpolation and linear bisection techniques. However, the accuracy and geological consistency of the isochronous stratigraphic framework are based on the precise division of actual sub-layers, and the above problems need to be solved urgently.

Regarding the existing technology of intelligent horizon division, literature: Shang Fuhua, Li Jincheng, Yuan Ye, Cao Maojun, Du Ruishan. Stratum division method based on improved BP neural network [J]. Computer Technology and Development, 2020, 30(09): 148-153. The author’s essential technology for stratigraphic division is to use neural networks to classify lithology based on well logging curves. The most basic three-layer BP neural network is selected, and the lithology category is used as the basis for stratigraphic division. The sampling points for sudden changes in lithology are the stratification points of small layers. Therefore, the identification of lithology is the focus of the work. At the same time, the author chose the L-M algorithm to improve and optimize the neural network training effect; the author chose the natural potential curve and the natural gamma curve as the main curves for lithology classification, and then referred to the resistivity, acoustic time difference, density, and neutron curve characteristics, Carry out small layer division. The author chose an improved BP neural network. This type of network is a relatively mature basic neural network at present, but it has many shortcomings: 1. When dealing with some complex nonlinear problems, the training of the network can easily fall into local extremes. Small value, which leads to training failure and a converged neural network model cannot be obtained; 2. The convergence speed is slow, that is, the efficiency of the algorithm is very low; 3. The structure selection of the neural network does not have a systematic theoretical guidance and is all based on expert experience. Therefore, for different target work areas, different geological conditions, and different stratigraphic divisions, its generalization ability is insufficient; 4. This network is highly dependent on samples and has high requirements for data quality.

Literature, Liu Yingjie. Intelligent stratigraphic correlation technology methods and applications [D]. Yanshan University, 2013. Using the Gaussian model to obtain more accurate stratification results through credibility calculation and control the selection of stratification interfaces; determination of lithological intervals After that, according to the characteristics of the well logging attributes, the parameters such as natural gamma, natural potential, acoustic time difference, sand layer thickness and rhythm of each layer are extracted, and these are used as the characteristic attributes of each small layer, using the Probabilistic Neural Network (PNN). Compare the stratigraphy between wells and connect them. The data preprocessing process of this method is relatively complex: the logging curve rhythm types and feature vector tables are summarized through the logging attribute characteristics, that is, the image type parameters are converted into numerical types so that the computer can recognize them. In addition, the natural gamma curve is also processed (GR), acoustic transit time curve (AC), and natural potential curve (SP) are filtered to filter the signals, thereby converting the logging data into square waves. The establishment of the learning sample refers to a standard well, and the verification of the neural network performance only selects 4 wells. For the oil field work area, its applicability cannot be verified, and the use of too few data sets cannot be widely and accurately used. Provide effective and comprehensive geological information for subsequent modeling work through layer division.

Therefore, the existing intelligent layer division methods generally only focus on single well information. When dividing layers, the geological information and seismic information around the target well are not considered. At most, these two types of information are used after the preliminary division results are obtained. There are some common challenges in constraint correction and the application of artificial intelligence in the field of geophysical interpretation: data set size, data quality, model convergence speed, final application effect, etc.

Contents of the invention

In view of the above technical problems, the present invention provides an underground multi-information constrained isochronous stratigraphic framework intelligent modeling method, which mainly focuses on well logging and geological information. It performs intelligent division of small layers based on well logging curves and simultaneously creates through data sets. This method adds geological constraints and allows geological information to participate in the entire division process, thereby obtaining results that are more consistent with geological information. Based on the fine layer division, with the help of seismic inversion information and neural network, we solve the problem of layer division between wells and establish a high-precision isochronous stratigraphic framework model. The entire method can effectively improve process efficiency through artificial intelligence technology, and the production of geologically constrained data sets can break the limitations of neural networks on data set size and model convergence speed.

The specific technical solutions are:

The underground multi-information constrained isochronous stratigraphic framework intelligent modeling method is characterized by including the following steps:

S1. Analyze the target work area based on expert experience, select lithology-sensitive logging curves, determine the curve type as the neural network data set as C and the number of sample points for each curve as L, and determine the neural network data set as the neural network data set based on the amount of existing logging information. There are N wells as samples in the network training set, and it is determined that the prediction set samples are M blind wells that need to be divided into small layers;

S2. Conduct geological analysis on the work area, determine the source direction α and vertical source direction β, β-α = 90 ^° , and divide the entire work area logging into 4 parts numbered ①-④ based on α and β;

S3. According to the above-mentioned divided work areas, prepare the neural network training data set;

The production process is as follows: determine that a connected well sample consists of Q wells, and C logging curves for each well have been determined. Two sets of 25% size data of the training set samples are produced according to the direction of the source and the direction perpendicular to the source. The remaining 50% of the data samples are randomly combined;

S3.1. Along the source direction: number the well training samples in the ①③ area N _i , i=1,2,3...,N ₁ , and the number of samples is N ₁ ; number the well training samples in the ②④ area N _i , i=1,2,3…,N ₂ , the number of samples is N ₂ ; where N=N ₁ +N ₂ , the well number W _k that constitutes a sample is generated according to formula (1), k=1,2…, Q;

I _a =[N _j *random(0～1)+0.5] (1)

[] means rounding up, j=1,2. When j=1, take the well samples from the ①③ area for random combination. At this time, N _j =N ₁ . When j=2, take the well samples from the ②④ area for random combination. This When N _j =N ₂ ;

S3.2. Vertical source direction: number the well training samples in the ①② area N _i , i=1,2,3...,N ₃ , and the number of samples is N ₃ ; number the well training samples in the ③④ area N _i , i=1,2,3...,N ₄ , the number of samples is N ₄ ; where N=N ₃ +N ₄ , the well number W _k that constitutes a sample is generated according to formula (2), k=1,2.. ,Q;

I _v =[N _j *random(0～1)+0.5] (2)

[] means rounding up, j=3,4. When j=3, take the well samples from the ①② area for random combination. At this time, N _j =N ₃ ; when j=4, take the well samples from the ③④ area for random combination. This When N _j =N ₄ ;

S3.3. Random combination: All training samples in the work area are uniformly numbered N _i , i=1,2,3...,N; according to formula (3), the well sequence number W _k that constitutes a sample is generated, k=1,2. .,Q;

I _r =[N*random(0～1)+0.5] (3)

[] means rounding up, N is the total number of training samples;

The neural network training set is composed according to the above three methods.

S4. Number the prediction samples M _j , j = 1, 2, 3...M, and number all the training samples _Ni , i = 1, 2, 3..., N. Generate well numbers W _k and M according to Equation (4) _j forms a sample, k=1,2..,Q-1; that is, Q-1 wells in the training set and 1 well in the prediction set form a prediction sample;

I _p =[N*random(0～1)+0.5] (4)

[] means rounding up, N is the number of training samples;

Finally, the final neural network data set is formed based on the training set obtained in S3 and the generated prediction set;

S5. Construct the feature pyramid network FPN and use the training set in S4 for model training;

S6. Use the model obtained in S5 to divide the prediction set in the target work area, that is, the blind well into batches of small layers;

S7. In the target work area, use the layer division result obtained in S6 as the boundary, perform random optimization inversion of the target well section based on seismic data to obtain the initial wave impedance inversion model, use each sample point in the model as a pseudo well, and select the pseudo well. The wells are trained according to the random combination method in S3 to create a new sample set. Based on the convergence model, the remaining unknown layer information is predicted for the pseudo wells, and then the small two-dimensional layers of the entire work area are obtained, and then the artificial stratification of the real well logs is used. Error calculation is performed on the layer plane, and interpolation is performed on the basis of well point error calculation to obtain the correction result;

S8. Repeat step S7, and continuously optimize and correct the intelligent stratigraphic division results with the well layer as the constraint condition to obtain the final two-dimensional stratigraphic division result that conforms to geological laws;

S9. Based on the results obtained in S8 as horizon constraints, the final isochronous stratigraphic framework model is obtained.

The beneficial effects brought by the technical solution of the present invention are:

1. The beneficial effects of the data set creation method in step S2: it can accelerate the convergence of the neural network model; at the same time, this method abandons the traditional neural network learning method of single well information feature extraction, and transforms one-dimensional well logging data in the form of connected well data. It improves the learning efficiency of the FPN network for two-dimensional data and can better preserve the geological spatial characteristics between wells. It also adds geological information to the data set as a constraint and runs through the entire intelligent layer division process, which is beneficial to the final Obtain layer division results that are more in line with real geological information; when the number of wells is small, data augmentation can also be performed through the production of well-connected data sets to solve the problem of lack of data and inability to train neural network models. When the number of wells is large, In this case, work efficiency can be greatly improved through automated processes.

2. Select the FPN network for the task of dividing small layers, which can simultaneously identify the types of small layers and accurately detect the position of layering lines. Its network structure and multi-scale detection function can not only retain more connected information of layers in geological space It can also improve the operating efficiency of the model and the detection performance of small-scale targets. Compared with choosing an ordinary neural network to achieve simple layer classification tasks, this method can effectively improve the accuracy of intelligent small layer division.

3. Integrate geological information and well logging information and use artificial intelligence methods to obtain high-precision sub-layer division results. Based on this result, seismic information is used to obtain inter-well layers that conform to geological laws, thereby achieving point-to-surface division of sub-layers. process, a high-precision isochronous stratigraphic framework model can finally be obtained.

Description of drawings

Figure 1 shows the work area division results of the embodiment;

Figure 2 is a well-connected sample structure according to the embodiment;

Figure 3 shows the intelligent sub-layer division results of the blind well according to the embodiment;

Figure 4 is the initial model of wave impedance inversion according to the embodiment;

Figure 5 shows the layer correction according to the embodiment.

Detailed ways

The specific technical solutions of the present invention will be described with reference to examples.

Intelligent modeling method of underground multi-information constrained isochronous stratigraphic framework:

S1. Analyze the target work area based on expert experience, select well logging curves that are sensitive to lithology such as sand and mudstone, determine the type of curve as the neural network data set as C and the number of sample points for each curve as L, according to the existing well logging information Determine the number of sample N wells as the neural network training set, and determine the prediction set samples, that is, M blind wells that need to be divided into small layers;

S2. Conduct geological analysis on the work area, determine the source direction α and vertical source direction β, β-α = 90 ^° , and divide the entire work area logging into 4 parts numbered ①-④ based on α and β, as shown in Figure 1 , the solid line is the cutting dividing line, and the direction pointed by the arrow is the source direction;

S3. According to the above-mentioned division of work areas, the neural network training data set is produced. The production process is as follows: determine that a well-connected sample consists of Q wells, and each well has C logging curves (in step S1). It has been determined, then a The samples are shown in Figure 2. Two sets of 25% of the training set samples were produced according to the direction of the source and perpendicular to the source, and the remaining 50% of the data samples were randomly combined.

S3.1. Along the source direction: number N _i (i=1,2,3...,N ₁ ) for the well training samples (number of samples is N ₁ ) in the ①③ area, and number the well training samples (samples) in the ②④ area The number is N ₂ ) and numbered Ni ( _i =1,2,3...,N ₂ ), where N=N ₁ +N ₂ , and the well number W _k (k=1) that constitutes a sample is generated according to the formula (1) ,2…,Q);

I _a =[N _j *random(0～1)+0.5] (1)

[] means rounding up, j=1,2. When j=1, take the well samples from the ①③ area for random combination. At this time, N _j =N ₁ . When j=2, take the well samples from the ②④ area for random combination. This When N _j =N ₂ ;

S3.2. Vertical source direction: number N _i (i=1,2,3...,N ₃ ) for the well training samples (number of samples is N ₃ ) in areas ①②, and number the training samples (samples) of wells (samples) in areas ③④ The number is N ₄ ) and numbered Ni ( _i =1,2,3...,N ₄ ), where N=N ₃ +N ₄ , and the well sequence number W _k (k=1) that constitutes a sample is generated according to the formula (2) ,2..,Q);

I _v =[N _j *random(0～1)+0.5] (2)

[] means rounding up, j=3,4. When j=3, take the well samples from the ①② area for random combination. At this time, N _j =N _3. When j=4, take the well samples from the ③④ area for random combination. This When N _j =N ₄ ;

S3.3. Random combination: Unify number N _i (i=1,2,3...,N) for all training samples in the work area and generate the well sequence number W _k (k=1,2) that constitutes a sample according to equation (3). .,Q);

I _r =[N*random(0～1)+0.5] (3)

[] means rounding up, N is the total number of training samples;

The neural network training set is composed according to the above three methods.

S4. Number the prediction samples M _j (j=1,2,3...M), number all the training samples N _i (i=1,2,3...,N), and generate the well sequence number W _k according to Equation (4) (k=1,2..,Q-1) and M _j form a sample (that is, Q-1 wells in the training set and 1 well in the prediction set form a prediction sample;

I _p =[N*random(0～1)+0.5] (4)

[] means rounding up, N is the number of training samples;

Finally, the final neural network data set is formed based on the training set obtained in S3 and the generated prediction set;

S5. Construct a feature pyramid network (FPN) and use the training set in S4 for model training;

S6. Use the model obtained in S5 to divide the prediction set in the target work area, that is, the blind well into batches of small layers, as shown in Figure 3;

S7. In the target work area, use the layer division result obtained in S6 as the boundary, and conduct stochastic optimization inversion of the target well section based on seismic data to obtain the initial wave impedance inversion model, as shown in Figure 4. Each sample point in the model is used as a pseudo Wells, select pseudo wells to create a new sample set for training according to the random combination method in S3, and predict the remaining pseudo wells with unknown layer information based on the convergence model, thereby obtaining the small two-dimensional layers of the entire work area, and then based on the artificial data of the real well logging Calculate the error of this layer layer by layer, and perform interpolation based on the well point error calculation to obtain the correction result, as shown in Figure 5;

S8. Repeat step S7, and continuously optimize and correct the intelligent stratigraphic division results with the well layer as the constraint condition to obtain the final two-dimensional stratigraphic division result that conforms to geological laws;

S9. Based on the results obtained in S8 as horizon constraints, the final isochronous stratigraphic framework model is obtained.

Claims (3)

1. The intelligent modeling method for the underground multi-information constrained isochronous stratum grillage is characterized by comprising the following steps of:

s1, analyzing a target work area according to expert experience, selecting lithology-sensitive logging curves, determining that the type of the curves serving as a neural network data set is C, the number of sample points of each curve is L, determining N wells serving as a neural network training set according to the number of existing logging information, and determining a prediction set sample, namely M blind wells needing to be divided into small layers;

s2, carrying out geological analysis on the work area, determining an object source direction alpha and a vertical object source direction beta, wherein beta-alpha=90 DEG, and dividing the whole work area well logging into 4 parts numbered (1) - (4) based on alpha and beta;

s3, manufacturing a neural network training data set according to the divided work areas;

s4, numbering predicted samples, and numbering all training samples to generate a sample predicted sample; forming a final neural network data set according to the training set obtained in the step S3 and the generated prediction set;

s5, constructing a feature pyramid network FPN, and performing model training by using the training set in the S4;

s6, carrying out batch small-layer division on a prediction set, namely blind wells, in the target work area by using the model obtained in the S5;

s7, taking the horizon dividing result obtained in the S6 as a boundary in a target work area, carrying out random optimization inversion on a target well section based on seismic data to obtain a wave impedance inversion initial model, taking each sample point in the model as a pseudo well, selecting the pseudo well to manufacture a new sample set for training according to a random combination method in the S3, and predicting pseudo wells with other unknown horizon information based on a convergence model to further obtain a small two-dimensional layer of the whole work area, carrying out error calculation on the horizon according to manual layering of a real logging, and carrying out interpolation on the basis of well point error calculation to obtain a correction result;

s8, repeating the step S7, and continuously optimizing and correcting the intelligent stratum division result by taking the well horizon as a constraint condition to obtain a two-dimensional horizon division result which finally accords with a geological rule;

s9, obtaining a final isochronous stratigraphic grid model based on the result obtained in the S8 as horizon constraint.

2. The intelligent modeling method for the underground multi-information constrained isochronous stratigraphic framework according to claim 1, wherein the neural network training data set is created in the step S3;

the manufacturing flow is as follows: determining that one well connecting sample consists of Q wells, wherein each well has C well logging curves, respectively manufacturing 25% of the number of two sets of training set samples according to the direction of a material source and a vertical material source, and carrying out the rest 50% of data samples in a random combination mode;

s3.1, following the object source direction: numbering N the well training samples of the (1) (3) region _i ，i＝1,2,3…,N ₁ The number of samples is N ₁ The method comprises the steps of carrying out a first treatment on the surface of the Numbering N the well training samples of the (2) (4) region _i ，i＝1,2,3…,N ₂ The number of samples is N ₂ The method comprises the steps of carrying out a first treatment on the surface of the Where n=n ₁ +N ₂ Generating a well sequence number W constituting a sample according to formula (1) _k ，k＝1,2…,Q；

I _a ＝[N _j *random(0～1)+0.5] (1)

[]Representing a round-up, j=1, 2, where N is the random combination of the well samples taken from the (1) (3) region when j=1 _j ＝N ₁ When j=2, taking (2) (4) area samples for random combination, N _j ＝N ₂ ；

S3.2, vertical object source direction: numbering N the well training samples of the (1) (2) regions _i ，i＝1,2,3…,N ₃ The number of samples is N ₃ The method comprises the steps of carrying out a first treatment on the surface of the Numbering N the well training samples of the (3) (4) region _i ，i＝1,2,3…,N ₄ The number of samples is N ₄ The method comprises the steps of carrying out a first treatment on the surface of the Where n=n ₃ +N ₄ Generating a well sequence number W constituting a sample according to formula (2) _k ，k＝1,2..,Q；

I _v ＝[N _j *random(0～1)+0.5] (2)

[]Representing a round-up, j=3, 4, where N is the random combination of the well samples taken from the (1) (2) region when j=3 _j ＝N ₃ The method comprises the steps of carrying out a first treatment on the surface of the When j=4, taking (3) (4) area samples for random combination, N _j ＝N ₄ ；

S3.3, random combination: uniformly numbering N for all training samples in work area _i I=1, 2,3 …, N; generating a well sequence number W constituting a sample according to formula (3) _k ，k＝1,2..,Q；

I _r ＝[N*random(0～1)+0.5] (3)

[] Representing the upward rounding, wherein N is the total number of training samples;

and forming a neural network training set according to the three modes.

3. The method for intelligent modeling of an underground multi-information constrained isochronous stratigraphic framework according to claim 1, wherein S4 specifically comprises: numbering predicted samples M _j J=1, 2,3 … M, number N for all training samples _i I=1, 2,3 …, N, the well number W is generated according to equation (4) _k And M is as follows _j One sample was composed, k=1, 2., Q-1; namely, the training set takes a predicted sample which is formed by 1 well in the prediction set and 1 well in the Q-1 well;

I _p ＝[N*random(0～1)+0.5] (4)

[] Represents the rounding up, N is the number of training samples.