A reservoir facies modeling method based on sequential neural network

Abstract

The application discloses a reservoir facies modeling method based on a sequential neural network, which comprises the steps of facies mode data volume edition establishment of the sequential neural network, a sequential neural network facies modeling technical method and sequential neural network facies modeling result visualization; the facies mode data volume edition establishment of the sequential neural network further comprises the sub-steps of multi-type facies mode data volume edition establishment and each volume edition statistical characteristic analysis and preprocessing; the sequential neural network facies modeling technical method further comprises the sub-steps of sequential neural network sample chain establishment, sequential neural network structure establishment and sequential neural network parallel training and combination inference; and the application realizes fast, reliable and intelligent establishment of complex facies system modeling.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of pixel-based algorithms in reservoir stochastic modeling, and particularly relates to a reservoir facies modeling method based on a sequential neural network. BACKGROUND

[0002] Reservoir modeling refers to using existing geological data including logging, seismic, core analysis and testing data, following the spatial correlation, heterogeneity and other property characteristics of the reservoir, using relevant algorithms to establish an inter-well attribute model, including porosity model, permeability model, sedimentary facies model, etc., and finally forming a three-dimensional model that completely represents the attribute distribution characteristics in the spatial reservoir. Facies modeling is modeling of discrete attributes such as sedimentary facies, lithofacies, fluid facies, etc. The entire reservoir modeling research field can be divided into three parts: ① research on reservoir modeling input data. ② research on specific modeling algorithms. ③ research on post-processing algorithms of modeling results.

[0003] The first part mainly studies how to extract geological features from reservoir modeling input data (logging, seismic, core analysis, etc.), and perform basic geological information interpretation, including lithology, lithofacies, sedimentary facies interpretation, rock mechanics parameter interpretation, reservoir physical property parameter and horizon interpretation, etc. The second part is to study modeling algorithms. Existing reservoir modeling algorithms can be divided into deterministic modeling and stochastic modeling. Stochastic modeling can generate multiple models, which is of great significance for evaluating reservoir heterogeneity and development risk, and is widely used in production practice. The third part of post-processing refers to the process of solving the problems existing in the model obtained by the modeling algorithm. The model obtained directly by the modeling algorithm usually has problems such as interface or boundary not smooth, attribute distribution not continuous, "noise points", etc. The main method of post-processing is to take some algorithms based on probability statistics to perform secondary operation on the model obtained by the modeling algorithm to obtain a model with better continuous distribution and smoother interface.

[0004] The three parts have integrated certain artificial intelligence algorithms, but the related technical methods are not yet mature. In the field of modeling algorithms, the current mainstream artificial intelligence method is to use a generative adversarial network (GAN) for prediction. This method requires a large amount of reservoir model data, and also needs a long training time, and its inference speed is also slow. In the case of needing fast modeling and having relatively small amount of known data, more two-point geostatistical stochastic modeling algorithms are used.

[0005] In actual engineering, the data that can be collected or directly established before establishing the spatial facies distribution model usually includes: 1) the facies distribution model of a single well, and the data structure is one-dimensional; 2) the facies distribution pattern provided by geologists, and the data structure is two-dimensional. The algorithm of two-point geostatistics can make full use of the data in each dimension, and has the characteristics of fast calculation; the algorithm of multiple-point geostatistics can usually make good use of the plane data, and better simulate the complex facies distribution model, but the calculation is slower; and the conventional single model deep learning algorithm usually needs a large amount of data, and it is difficult to use multi-dimensional data at the same time. The application provides a facies modeling technical method based on a sequential neural network, which can comprehensively utilize the advantages of the above three algorithms, and realizes fast, reliable and intelligent establishment of a complex facies system modeling. SUMMARY

[0006] The technical problem solved by the application is to overcome the shortcomings of the existing modeling technology, and to provide a reservoir facies modeling method based on a sequential neural network.

[0007] The object of the application is achieved by the following technical scheme:

[0008] A reservoir facies modeling method based on a sequential neural network, which realizes fast, reliable and intelligent facies modeling of a complex oil and gas reservoir facies system based on lithofacies data of an oil and gas reservoir and corresponding logging curves, and comprises the following steps:

[0009] A sequential neural network facies modeling technical method;

[0010] The sequential neural network facies modeling technical method comprises a sequential neural network sample chain establishment substep, a sequential neural network structure establishment substep and a parallel training and combination inference substep of the sequential neural network facies modeling. Specifically, let the data volume version be F={f j |j=1,2,…,M;f j =1,2,…,c},M positions and c facies types, f j represents the facies type discrete value of the jth position in the volume space, and the sequential neural network sample chain establishment substep further comprises:

[0011] Generating a random path of a sample chain: a series of random numbers are generated by using a preset seed number, each random number represents a position in space, the random numbers corresponding to the positions in the initial known data are removed from the random number set, and then the random numbers in the set are sorted to obtain a random path p={p i |i=1,2,…,N;p i =1,2,…,M}(N<M)

[0012] Wherein p i represents the position represented by the ith node on the path;

[0013] Initialize the sample chain and determine the first sample on the sample chain: initialize the sample chain

[0014] D = {D i |i = 1, 2, …, N}

[0015] wherein, represents the sample represented by the i-th node on the random path, represents the phase type value of the j-th position in the sample represented by the i-th node on the random path;

[0016] Generate the initial sample D1 according to the initial known data, and set the initial known position set as K, then assign the phase type value of all positions in the space represented by the data version as 0, and assign the corresponding phase type discrete value to the initial known position, i.e.

[0017]

[0018] Iteratively calculate all samples of the sample chain: sequentially generate the entire sample chain based on the initial sample D1, and the sample represented by the i+1-th node is i assign the phase type value of the p-th position in the sample space represented by the i-th node as the sample generated thereafter, i.e.

[0019]

[0020] Specifically, the sequential neural network structure establishing sub-step comprises:

[0021] The learning of the one-dimensional data sample chain is trained by using a deep neural network, the learning of the two-dimensional or three-dimensional data sample chain is trained by using a two-dimensional convolutional or three-dimensional convolutional neural network respectively, and the convolutional kernels used are all irregular shape convolutional kernels.

[0022] Specifically, the sequential neural network structure establishing sub-step further comprises:

[0023] Determining the structure shape of the regular convolutional kernel: for a planar sample chain, the structure shape of the regular convolutional kernel needs to be determined, a certain angle is set as an interval to select several directions to calculate the experimental variation functions of these directions, the direction with the largest variation range is selected as the direction of the main variation range, the length of the convolutional kernel is the distance of twice the main variation range direction variation range plus one pixel, the direction perpendicular to the direction is the secondary variation range direction, and the width of the convolutional kernel is the distance of twice the secondary variation range direction variation range plus one pixel;

[0024] The non-regular convolution kernel structure shape is determined by the nearest principle of the regular convolution kernel, that is, the center of the regular convolution kernel obtained in the step of determining the regular convolution kernel structure shape is overlapped with the sample pixels, the distances between the remaining convolution kernel units and the sample pixels are calculated respectively, and the nearest pixel is selected as the corresponding pixel of the convolution kernel unit;

[0025] The above steps are repeated to find the corresponding pixel of each convolution kernel unit, and some convolution kernel units may correspond to the same pixel. After the repetition, the actual structure shape of the non-regular convolution kernel is obtained.

[0026] Specifically, the samples used in the parallel training and combined inference sub-steps of the sequential neural network are extracted from the sample chain, and the total number of samples is the length N of the sample chain. The sample D represented by the i-th node position on the random path is i The input data for training is the node sample, and the label value of the node sample is Specifically, the parallel training and combined inference sub-steps of the sequential neural network modeling include:

[0027] Step 1: Extract the sample D1 represented by the first node p1 on the random path, and the specific phase type of the position represented by the node in the data volume

[0028] Step 2: input the sample D1 into the sub-network corresponding to the sample chain for training;

[0029] Repeat steps 1-2 in the order of the random path until all nodes on the random path have been traversed, that is, one training of the entire sample chain is completed. The network performance is judged by the fitting degree of the sample chain, that is, the loss function value.

[0030] In the training process, when the loss rate tends to be stable, the network parameters representing the sample chain converge to the best, that is, the parallel training process is terminated;

[0031] The combined inference is independently predicted by each sub-network model according to the sample data, and then the prediction results of each sub-network are weighted and summed according to certain weight values to obtain the output result of the sequential neural network, that is, the inference result of each sub-network, including the following steps:

[0032] According to the initial data, the input data D1 of each sub-network is determined, and the combined inference phase type value of each sub-network at the first node p1 position on the random path is obtained through the initial input data The next input data is generated:

[0033]

[0034] ​The next input data step is repeatedly generated until the last node on the random path, through which the phase type of the last node on the random path representing the position can be predicted The final prediction result of the model is

[0035]

[0036] The inference results of each sub-network are combined as follows:

[0037]

[0038] Each model is M i , and the weight is w i .

[0039] Specifically, the phase mode data volume establishment of the sequential neural network further comprises the following steps:

[0040] A statistical learning algorithm and prior geological knowledge are used to establish the corresponding relationship between the logging data and the phase type, so as to establish a single-well phase mode volume; a structured grid is used to discretize the planar phase distribution area map established by a geologist, and each grid is assigned a phase according to the principle of maximum area, thereby establishing a two-dimensional planar phase mode volume, and the number of each phase type is counted, which serves as a basis for evaluating the reliability of the final result The phase mode data volume establishment of the sequential neural network comprises the following sub-steps:

[0041] Single-well phase mode volume establishment;

[0042] Planar phase mode volume establishment;

[0043] Analysis and preprocessing of statistical characteristics of each volume.

[0044] Specifically, the single-well phase mode volume establishment sub-step further comprises:

[0045] Determining the feature data corresponding to the phase mode type, which is usually logging data;

[0046] Using a small sample set to select a better solution from commonly used statistical learning algorithms and establishing the structure of the algorithm, which is usually a support vector machine (SVM) algorithm or a multi-layer perceptron (MLP) algorithm;

[0047] After establishing the structure, the core analysis sample and the corresponding feature data value are used as training samples, and the trained algorithm is used to infer the overall lithofacies distribution model, thereby establishing a single-well lithofacies mode volume;

[0048] The planar phase mode volume establishment sub-step specifically comprises:

[0049] By analyzing factors including regional sedimentary environment, belonging layer, structural characteristics and mechanical properties, and combining with experience, a planar facies distribution map is established;

[0050] The planar region is discretized to form a grid surface, and the attribute value of each grid is assigned according to the facies region to which it belongs. The grid containing the boundary line is assigned according to the principle of maximum area, that is, the attribute of the grid is the facies type with the largest contained area.

[0051] Specifically, the facies analysis and preprocessing of each plate are specifically:

[0052] The number of each lithofacies type of the entire plate is statistically analyzed, so that the distribution probability of each facies is obtained, and serves as a basis for the reliability of the final evaluation model.

[0053] Specifically, it further includes the steps of visualizing the sequential neural network facies modeling result, and the sequential neural network facies modeling result visualization includes:

[0054] The rendering data body of the model is defined by using the topological relationship between the point lines and surfaces in the model data body obtained by simulation and between the grid blocks. The specific operation is to first store the vertices of each grid block according to two directions of the plane, then store the vertices of each plane along the depth direction, and then store the topological relationship of each vertex in the form of four points forming a surface, thereby forming a complete rendering data body;

[0055] The rendering data body is established and the open source graphics rendering application program interface is used, and the final visual graph can be obtained according to the fixed rendering pipeline.

[0056] The beneficial effects of the present application are:

[0057] The present application combines the advantages of the algorithms of two-point geostatistics, the algorithms of multi-point geostatistics and the deep learning algorithm of a conventional single model, and realizes rapid, reliable and intelligent establishment of complex facies system modeling. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained according to the structures shown in the drawings without creative labor.

[0059] Figure 1 is a technical flowchart of the present application;

[0060] Figure 2 is a schematic diagram of the size of the region position of the case used in the present application;

[0061] Figure 3 is a schematic diagram for establishing a planar version of a case area used in the present application;

[0062] Figure 4 is a W1 single-well facies volume map of a case area used in the present application;

[0063] Figure 5 is a W2 single-well facies volume map of a case area used in the present application;

[0064] Figure 6 is a W3 single-well facies volume map of a case area used in the present application;

[0065] Figure 7 is a top-bottom bed facies distribution map (left bottom right top) of a target depth section of a case area used in the present application;

[0066] Figure 8 is a schematic diagram for establishing a sample chain of the present application;

[0067] Figure 9 is a schematic diagram for associating a one-dimensional sample chain and a two-dimensional sample chain of the present application;

[0068] Figure 10 is a schematic diagram for establishing an irregular convolution kernel of the present application;

[0069] Figure 11 is a CNN structure diagram of a case area used in the present application;

[0070] Figure 12 is a loss value change diagram of a DNN training process of a case area used in the present application;

[0071] Figure 13 is a loss value change diagram of a CNN training process of a case area used in the present application;

[0072] Figure 14 is a combination inference schematic diagram of the present application;

[0073] Figure 15 is a comparison diagram of different methods (Petrel and sequential neural network comparison diagram) of a case area used in the present application;

[0074] Figure 16 is a comparison diagram of A1-A18 sequential neural network grid simulation results of the present application;

[0075] Figure 17 is a comparison diagram of A19-A22 sequential neural network grid simulation results and B1-B12 Petrel sequential indication simulation results of the present application;

[0076] Figure 18 is a comparison diagram of B13-B22 Petrel sequential indication simulation results of the present application. DETAILED DESCRIPTION

[0077] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0078] To provide a clearer understanding of the technical features, objectives, and beneficial effects of this invention, the following detailed description of the technical solution is provided. Obviously, the described embodiments are only a portion of the embodiments of this invention, not all of them, and should not be construed as limiting the scope of implementation of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the protection scope of this invention.

[0079] Example 1

[0080] This invention proposes, as follows Figure 1 The sequential neural network facies modeling process shown is characterized by its key component, the second step, "Sequential Neural Network Facies Modeling Techniques," and its three key technical steps. The complete process includes: 1) Establishing a data volume through statistical learning and prior geological knowledge to provide a data foundation for the subsequent establishment of sample chains; 2) Generating a random path within the target area and using this path to decompose the previously established volume, forming training samples at different locations along the random path, and linking all training samples in chronological order to form a sample chain. Multiple sequential neural networks are established according to different types of sample chains, and each model is trained in parallel using the sample chains. The results of these models are then combined to form the final facies model; 3) Utilizing the topological relationships between points, lines, surfaces, and grid blocks in space, as well as the facies distribution model obtained from the sequential neural network, the data volume required for graphics rendering is constructed. Rendering is then performed using an open-source graphics interface to obtain the final 3D graphics.

[0081] (1) Establishment of phase mode data plates for sequential neural networks

[0082] Statistical learning algorithms and prior geological knowledge are used to establish the correspondence between logging data and facies types, thereby creating a single-well facies model template. A structured grid is used to discretize the planar facies distribution map created by geologists, and facies values ​​are assigned to each grid according to the principle of maximizing area, thus creating a two-dimensional planar facies model template.

[0083] A single-well facies model quantity establishment

[0084] The target case area is a sub-region consisting of three wells in an oilfield in Jimsar, Junggar Basin. Figure 2), the simulated facies type is lithofacies, according to all core analysis test results of Jimsar, the lithofacies can be classified into 6 types, respectively: micritic dolomite, calcarenite, dolomitic mud shale, fine sandstone, dolomitic siltstone and silty mud shale. In addition, 7 logging curve data including spontaneous potential (SP), natural gamma (GR) are provided, therefore, the corresponding relationship between logging curve and lithofacies can be established by using algorithms such as multi-layer perception (MLP), support vector machine (SVM), so as to infer the lithofacies distribution type of non-coring section. Taking the MLP model as an example, the number of layers and the number of perception machines of the model are reasonably designed, and the learnable parameters of MLP are determined by using the sample data at the coring section. In the case area, the MLP model is trained by using the core analysis samples of the three wells to make the parameters converge to the optimal value, and then the overall lithofacies distribution model of the three wells is inferred, thereby establishing the lithofacies distribution volume profile of single well Figure 4 、 5 、6)。

[0085] B Planar facies model volume profile establishment

[0086] Geologists establish a planar facies distribution map by analyzing regional sedimentary environment, belonging horizon, structural characteristics, mechanical properties and other factors, combined with experience. The map mainly includes a group of sub-regions represented by closed curves, and each sub-region represents a facies distribution range. After the map is obtained, the planar region can be discretized to form a grid surface by setting the size of the structured network, and the attribute value of each grid is assigned according to the facies region it belongs to. The grid containing the boundary line can be assigned by the area maximum principle, that is, the attribute of the grid is the largest facies type it contains. Figure 3 )。

[0087] The grid of the case area is set to 50x50x1, and after the entire space is discretized, a 37x28x22 three-dimensional grid body is formed. At the same time, the planar facies distribution maps of the top surface (k=22 grid surface) and the bottom surface (k=1 grid surface) are also collected. According to the above method, two planar facies model volume profiles can be obtained. Figure 7

[0088] C Analysis of statistical characteristics of each volume profile

[0089] In this case area, there are only single well and planar volume profile data, so there is no need to establish three-dimensional volume profile data. According to the law of large numbers, when the number of trials is sufficient, the frequency of an event is approximately equal to the probability of the event occurring. Therefore, the number of grids of each facies type is counted, and the ratio of the number of grids to the total number of grids is approximately the probability of the facies type occurring in the simulation area. Whether on the plane or in the vertical direction, the sub-region does not contain micritic dolomite facies, so only the number of the remaining five types of lithofacies needs to be counted, and the final result is shown in Table 1.

[0090] ​Table 1 proportion of each lithofacies type

[0091]

[0092] (2) Sequential neural network facies modeling technique

[0093] The modeling method includes three key technical links of establishing a sample chain, establishing a network structure, parallel training and combined inference. The sample chain provides a data basis for subsequent network training and determines its sequential training method, the network structure determines the number and type of neural networks, and the parallel training and combined inference specify the calculation process method during final modeling.

[0094] A sequential neural network sample chain is established

[0095] The previously established volume provides a data basis for the establishment of the sample chain. The establishment of the sample chain is similar to the sequential simulation idea, and the basic calculation process of the latter is: according to the known data set, the facies type data of the target position is predicted and inferred, and the inference prediction result is integrated into the known data set to infer and predict the facies type of the next position. According to this idea, a series of positions with time sequence are determined to establish samples for each position, and then link the samples in time sequence to form a sequential neural network sample chain (the sample chain establishment of the planar data volume is shown in Figure 8 j |j=1,2,…,M;f j =1,2,…,c},f j represents the facies type discrete value of the jth position in the volume space (assuming there are M positions and c facies types), and the specific steps of establishing the sample chain are as follows:

[0096] ① Generate a random path for the sample chain. Set a seed number, generate a series of random numbers using the seed number, and each random number represents a position in space. Therefore, according to the position of the initial known data, remove the random numbers corresponding to these positions from the random number set, and then sort the random numbers in the set to obtain a random path p={p i |i=1,2,…,N;p i =1,2,…,M}(N<M), wherein p i represents the position represented by the ith node on the path.

[0097] ② Initialize the sample chain and determine the first sample on the sample chain. Initialize the sample chain D={D i |i=1,2,…,N}, represents the sample represented by the ith node on the random path, ​represents the phase type value of the jth position in the sample represented by the ith node on the random path. The initial sample D1 is generated according to the initial known data: assuming that the initial known position set is K, the phase type values of all positions in the space represented by the data volume are assigned as 0, and the initial known positions are assigned as the corresponding phase type discrete values, i.e.

[0098] ③The iteration is calculated to obtain all samples in the sample chain. The entire sample chain is sequentially generated according to the initial sample D1. The sample represented by the ith+1 node is obtained by assigning the phase type of the pth position in the sample space represented by the ith node as i the sample generated subsequently, i.e.

[0099] In the present case, there are both one-dimensional single-well phase distribution data and two-dimensional planar phase distribution data. Therefore, two sample chains D1 and D2 can be established. In the two-dimensional sample chain D2, the known phase distribution data of the first sample is the phase attribute of the single well in the layer; in the one-dimensional sample chain D1, the known phase distribution of the first sample is the known phase data of each layer on the well. Figure 9

[0100] B Sequential neural network structure establishment

[0101] The sequential neural network is essentially a composite model composed of multiple neural network models, each sub-network corresponds to a sample chain, and the structure of each sub-network is determined, so the entire sequential neural network is determined. The one-dimensional data sample chain is learned by using a DNN, and the two-dimensional or three-dimensional data sample chain is trained by using a two-dimensional convolution or three-dimensional convolution neural network, respectively. The convolution kernel is a non-regular shape convolution kernel. The structure determination of the DNN is relatively simple, which can be determined by simple trial training according to the length and number of the sample chain; while the structure determination of the CNN is relatively complex, the key of which lies in determining the structure shape of the non-regular convolution kernel. The present application determines the structure shape of the non-regular convolution kernel by using the range of the variation function, so the structure shape of the non-regular convolution kernel can be determined by analyzing the range direction and length of the variation function of the data volume. Two ranges in two perpendicular directions need to be determined for a two-dimensional volume, and a range in the vertical direction needs to be determined for a three-dimensional volume. The specific steps are as follows: Figure 10

[0102] ①Determine the structure shape of the regular convolution kernel. For a planar sample chain, the structure shape of the regular convolution kernel needs to be determined. Several directions are selected at certain intervals to calculate the experimental variation function of these directions, the direction with the largest range is selected as the direction of the main range, the length of the convolution kernel is twice the range of the direction plus one pixel distance, and the direction perpendicular to the direction is the secondary range direction, and the width of the convolution kernel is twice the range of the direction plus one pixel distance.

[0103] ​​​② Determine the irregular convolution kernel structure shape using the nearest neighbor principle with regular convolution kernels. Overlap the center of the regular convolution kernel obtained in the previous step with the pixels of the sample, calculate the distance between each of the remaining convolution kernel units and the pixels in the sample, and select the nearest pixel as the corresponding pixel of that convolution kernel unit. Repeat the above steps to find the corresponding pixel for each convolution kernel unit. Some convolution kernel units may correspond to the same pixel. After this operation, the actual structure shape of the irregular convolution kernel can be obtained.

[0104] In this case, a one-dimensional single-well chain and a two-dimensional planar chain have been established, therefore two neural network models need to be built for training. For the one-dimensional single-well chain, after parameter tuning of the neural network, a three-layer DNN model with a specification of 22×11×6 was obtained. For the two-dimensional planar sample chain, variogram analysis revealed that the principal direction is due east. In this direction, the irregular convolution kernel is a special case, meaning the structural shape becomes regular, ultimately resulting in a model as shown below. Figure 11 The CNN model shown.

[0105] Parallel Training and Combinatorial Inference of C Sequential Neural Networks

[0106] During training, each sample chain is independent of the others, so the various sub-network models in the sequential neural network can be trained in parallel. When making predictions and inferences, the weighted summation of the predictions and inferences of each sub-network is the final modeling result of the sequential neural network.

[0107] During training, the samples used are extracted from the sample chain, so the total number of samples is equal to the length N of the sample chain. The sample D represented by the i-th node position on the random path... i The input data for the algorithm is , and the label value of the node sample is . This forms a complete training sample. The entire training process is as follows: 1) Extract the sample D1 represented by the first node p1 on the random path, and the specific phase type of the position represented by that node in the data volume. 2) Sample The data is fed into the sub-network corresponding to the sample chain for training; 3) Repeat the first two steps in random path order until all nodes on the random path have been traversed, which completes one training iteration of the entire sample chain. Then, the sample chain is trained multiple times according to the set number of iterations (epochs). The network performance is judged by the degree of fit to the sample chain (loss function value). During training, when the loss rate tends to stabilize, it indicates that the network parameters of the sample chain have converged to the optimal level, and the training process can be terminated.

[0108] During prediction and inference, each sub-network model first makes predictions independently based on the sample data. Then, the prediction results of each sub-network are weighted and summed according to certain weight values ​​to obtain the output result of the sequential neural network, which is the combined inference result of the sub-networks.Figure 14 ). Assume each sub-network is M i , the weight is w i , then

[0109]

[0110] The whole prediction process is: 1) According to the initial data, determine the input data D1 of each sub-network, and get the combined inference facies type value of each sub-network at the first node p1 position on the random path through the initial input data , thus generating the next input data 2) Repeat the above steps until the last node on the random path, through which the facies type of the position represented by the last node on the random path can be predicted The final prediction result of the model is

[0111] In this case, the loss rate change of the training process of the two sample chains is shown in Figure 12 and Figure 13 . Since there are only two models in this case, the final combined inference method is:

[0112] M(D) = a - Ml(Dl) + (1 - a) - M2(D2) (2)

[0113] Here, the value of a is 0.5. Figure 15 The simulation results obtained by the final inference of this method are shown, compared with the original volume data and the facies proportion obtained by the sequential indicator simulation of Petrel. From the facies content, this method is closer to the distribution of the original volume.

[0114] (3) Sequential neural network facies modeling result visualization

[0115] The open source graphics interface used for visualization is VTK, which encapsulates the interface of OpenGL twice, forming a computer graphics rendering development tool with high abstraction and easy use. When rendering the model, define the rendering data body of the facies model according to the topological relationship between the point, line and surface and each mesh block, define the rendering color of each mesh block according to the attribute value of each mesh, and then render the whole model according to the established rendering pipeline. The final simulation results obtained by the combined inference of the sequential neural network and the simulation results obtained by the sequential indicator simulation of the commercial modeling software Petrel are shown in Figure 16 , 17 , and 18.

[0116] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

[0117] It should be noted that, for each of the above-mentioned method embodiments, in order to simplify the description, each is described as a combination of a series of actions, but those skilled in the art should understand that the present application is not limited by the order of the actions described, because according to the present application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and units involved are not necessarily essential to the present application.

[0118] In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0119] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, can include the processes of the above-mentioned embodiments. The storage medium can be a magnetic disc, an optical disc, a ROM, a RAM, etc.

[0120] The above only discloses preferred embodiments of the present application, and of course cannot limit the scope of protection of the present application. Therefore, equivalent changes made in accordance with the claims of the present application are still within the scope of the present application.

Claims

1. A reservoir facies modeling method based on sequential neural networks, which, based on petrographic data of oil and gas reservoirs and their corresponding logging curves, achieves rapid, reliable, and intelligent facies modeling of complex oil and gas reservoir facies systems, characterized in that... Includes the following steps: Sequential neural network phase modeling techniques and methods; The sequential neural network phase modeling technique includes a sequential neural network sample chain establishment sub-step, a sequential neural network structure establishment sub-step, and a parallel training and combined inference sub-step for sequential neural network phase modeling. Let the data volume be If there are M positions and c phase types, then Representing the discrete value of the phase type at the j-th position in the scale space, the sequential neural network sample chain establishment sub-step further includes: Generating a random path for the sample chain: A series of random numbers is generated using a pre-defined seed number. Each random number represents a position in space. Based on the positions of the initially known data, the random numbers corresponding to those positions are removed from this set. Finally, the random numbers in the set are sorted to obtain a random path. (N<M); in This represents the position of the i-th node on the path; Initialize the sample chain and determine the first sample in the sample chain: Initialize the sample chain ； in, Let represent the sample represented by the i-th node on the random path. This represents the phase type value at the j-th position in the sample represented by the i-th node on the random path; Generate initial samples based on the initial known data. Let the initial known location set be K. Then, the phase type value of all locations in the space represented by the data volume is assigned to 0, and the initial known locations are assigned the corresponding discrete phase type values, i.e. ； Iterative calculation yields all samples in the sample chain: based on the initial samples. The entire sample chain is generated sequentially, where the sample represented by the (i+1)th node is the sample from the sample space represented by the ith node. The phase type assignment for each position is... The subsequently generated sample, i.e.: ； It also includes the step of establishing a phase mode data plate for a sequential neural network, wherein establishing the phase mode data plate for a sequential neural network includes: Statistical learning algorithms and prior geological knowledge are used to establish the correspondence between logging data and facies types, thereby creating a single-well facies model quantization. A structured grid is used to discretize the planar facies distribution map created by geologists, and facies values ​​are assigned to each grid according to the principle of maximizing area, thus creating a two-dimensional planar facies model quantization. The number of each facies type is counted, which serves as the basis for evaluating the reliability of the final results. The establishment of the phase pattern data plate of the sequential neural network includes the following sub-steps: Single-well facies model template established; Planar phase mode quantization plate establishment; Statistical characteristics analysis and preprocessing for each volume version.

2. The reservoir facies modeling method based on a sequential neural network according to claim 1, characterized in that, The sequential neural network structure establishment sub-steps include: The learning of one-dimensional data sample chains is trained using deep neural networks, while the learning of two-dimensional or three-dimensional data sample chains is trained using two-dimensional convolutional neural networks or three-dimensional convolutional neural networks, respectively. All convolutional kernels used are irregularly shaped convolutional kernels.

3. The reservoir facies modeling method based on a sequential neural network according to claim 1, characterized in that, The sequential neural network structure establishment sub-step also includes: Determine the structure and shape of the regular convolution kernel: For a planar sample chain, it is necessary to determine the structure and shape of the regular convolution kernel. Set a certain angle as the interval, select several directions, calculate the experimental variation function of these directions, and select the direction with the largest range as the main range direction. Twice the range of the main range direction plus one pixel distance is the length of the convolution kernel. The direction perpendicular to this direction is the secondary range direction, and twice the range of the secondary range direction plus one pixel distance is the width of the convolution kernel. The irregular convolution kernel structure shape is determined by using regular convolution kernels based on the nearest distance principle: the center of the regular convolution kernel obtained in the step of determining the regular convolution kernel structure shape is overlapped with the pixel of the sample, the distance between the remaining convolution kernel units and the pixels in the sample is calculated respectively, and the nearest pixel is selected as the corresponding pixel of the convolution kernel unit; By repeating the above steps, the corresponding pixel for each convolution kernel unit can be found. Some convolution kernel units may correspond to the same pixel. After repeating the process, the actual structural shape of the irregular convolution kernel can be obtained.

4. The reservoir facies modeling method based on a sequential neural network according to claim 1, characterized in that, The samples used in the parallel training and combinatorial inference sub-steps of the sequential neural network are extracted from the sample chain. The total number of samples is the length N of the sample chain, and the sample represented by the i-th node position on the random path. For the input data used in training, the label value of this node sample is... .

5. The reservoir facies modeling method based on a sequential neural network according to claim 1, characterized in that, The parallel training and combined inference sub-steps of the sequential neural network phase modeling include: Step 1: Extract the first node on the random path The sample represented And the specific phase type of the location represented by that node in the data volume. ; Step 2: Place the sample ( , The data is fed into the sub-network corresponding to the sample chain for training. Repeat steps 1-2 in the order of the random path until all nodes on the random path have been traversed, which is to complete one training of the entire sample chain. Train the sample chain multiple times according to the set number of iterations. The performance of the network is judged by the degree of fit to the sample chain, i.e., the loss function value. During training, when the loss rate tends to stabilize, it means that the network parameters of the sample chain have converged to the optimal level, and the parallel training process is terminated. The combined inference involves each sub-network model making independent predictions based on sample data, and then weighting and summing the predictions of each sub-network according to certain weight values ​​to obtain the output of the sequential neural network, which is the combined inference result of the sub-networks. This includes the following steps: Determine the input data for each sub-network based on the initial data. And obtain the first node on the random path using this initial input data. The combination of subnetworks at each location infers the phase type value. Generate the next input data: ； Repeat the process of generating the next set of input data until the last node on the random path. The phase type of the position represented by the last node on the random path can then be predicted. The model's final prediction result is ； The inference results of combining the subnetworks are as follows: ； The models are as follows: The weight is .

6. The reservoir facies modeling method based on a sequential neural network according to claim 1, characterized in that, The sub-step for establishing the single-well facies model template also includes: Identify the characteristic data corresponding to the phase pattern type, which is usually logging data; Using a small sample set, a better solution is selected from commonly used statistical learning algorithms and the structure of the algorithm is established, usually the support vector machine (SVM) algorithm or the multilayer perceptron (MLP) algorithm; After establishing the structure, core analysis samples and corresponding feature data values ​​are used as training samples. The trained algorithm is used to infer the overall lithofacies distribution model and establish a single-well lithofacies model template. The specific sub-step for establishing the planar phase mode scale is as follows: By analyzing factors including regional sedimentary environment, stratigraphic position, tectonic features, and mechanical properties, and combining them with experience, a planar facies distribution map is established. The planar region is discretized to form a grid surface. The attribute value of each grid is assigned according to the phase region to which it belongs. The grid containing the boundary line is assigned the value according to the principle of the largest area, that is, the attribute of the grid is the phase type with the largest area.

7. The reservoir facies modeling method based on a sequential neural network according to claim 1, characterized in that, The statistical feature analysis and preprocessing for each volume version are as follows: Statistical analysis was performed on the number of each lithofacies type in the entire model to obtain the distribution probability of each type, which served as a basis for the final evaluation of the model's reliability.

8. The reservoir facies modeling method based on a sequential neural network according to claim 1, characterized in that, It also includes a step of visualizing the results of sequential neural network phase modeling, wherein the visualization of the results of sequential neural network phase modeling includes: The rendering data volume of the model is defined by using the topological relationships between points, lines, surfaces and mesh blocks in the model data volume obtained by simulation. Specifically, the vertices of each mesh block are first stored in two directions along the plane, and then the vertices of each plane are stored along the depth direction. Finally, the topological relationships of each vertex are stored in the way that four points enclose a surface, thus forming a complete rendering data volume. By utilizing rendering data volumes and open-source graphics rendering application interfaces, the final visual graphics can be obtained by following a fixed rendering pipeline.

Images