A carbonate reservoir plane prediction method, system, device and storage medium
By combining deep learning algorithms with geological knowledge and expert analysis, and using the patch labeling method to integrate multiple sensitive seismic attributes, the problem of multiple solutions and interpretation difficulties in the prediction of carbonate fracture-vuggy reservoirs has been solved, achieving high-precision planar prediction and seismic facies probability prediction of carbonate reservoirs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-06-25
- Publication Date
- 2026-03-24
AI Technical Summary
Existing technologies suffer from multiple solutions and interpretation difficulties in predicting fractured-vuggy carbonate reservoirs, especially the fusion of multiple sensitive seismic attributes, which makes it difficult to accurately reflect reservoir characteristics.
By combining deep learning algorithms with geological knowledge and expert analysis, and using patch tagging, we can perform planar prediction of carbonate reservoirs and improve interpretation accuracy by utilizing various sensitive seismic attributes such as coherence, likelihood, amplitude variation rate, and gradient structure tensor.
It improves the interpretation accuracy of planar prediction of carbonate reservoirs to the pixel level, solves the problem of inconsistent threshold selection for various dimensionless seismic attributes, and provides prediction probabilities of seismic facies and clear delineation of planar boundaries.
Smart Images
Figure CN115527107B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of carbonate rock seismic reservoir prediction technology, and relates to deep learning algorithms in artificial intelligence, particularly a method, system, device and storage medium for planar prediction of carbonate rock reservoirs. Background Technology
[0002] Seismic attributes typically reflect changes in subsurface geological bodies, such as fault development and lithological variations. Therefore, seismic attribute analysis plays a crucial role in seismic reservoir prediction. However, seismic attributes are diverse, and different attributes are often only sensitive to specific geological features or anomalies. They are also frequently influenced by geological conditions unrelated to the research objective, requiring comprehensive analysis to eliminate false positives. Thus, using single-attribute analysis for reservoir prediction often leads to multiple interpretations. While multi-attribute fusion or comprehensive analysis introduces a significant workload for attribute interpretation, it is indeed an effective means of reducing the ambiguity of reservoir prediction. Currently, there are numerous methods for multi-attribute fusion in seismic reservoir prediction, which can be broadly categorized into two main directions: one is fusion of planar attributes after sensitive attribute calculation, and the other is fusion during the seismic attribute calculation process using a specific vertical time window (or depth). The method we have developed belongs to the first category. Commonly used multi-attribute fusion methods in seismic reservoir prediction include RGB (or RGBA) fusion, cluster analysis fusion, multiple linear regression fusion, and BP artificial neural network fusion. Each multi-attribute fusion method has its own effectiveness and limitations. RGB attribute fusion technology enhances the visual appeal of geological features through the fusion of three primary colors, but it requires the use of three or four attributes for fusion, and the resulting attributes often exhibit disorder, making attribute interpretation difficult. Cluster analysis seismic attribute fusion classifies attributes by calculating the correlation between them, which can uncover distribution patterns in different attribute data. However, cluster analysis seismic attribute fusion usually involves data calculations over a vertical time window and is rarely used for planar attribute fusion. Furthermore, the results of cluster analysis are related to the number of categories and are not easily interpreted, so its application is limited. Multiple linear regression for seismic attribute fusion is a simple method that can overcome the one-sidedness of single attribute prediction. The most widely used method is well attribute-based multiple linear regression fusion, which uses existing well logging data to assign different weights to each seismic attribute. The overall fusion calculation approach is relatively reasonable, but the complex relationship between seismic attributes and reservoirs is often not accurately represented by linear mapping. Moreover, when there are abnormal geological conditions in a sensitive attribute, the final fusion result will inevitably be affected, meaning that abnormal feature information cannot be eliminated.
[0003] For the prediction of fractured-vuggy carbonate reservoirs, how to use the results of drilling, geological understanding, and comprehensive analysis and judgment by experts to constrain the fusion of multiple sensitive plane seismic attributes across the entire work area is an urgent problem to be solved. Summary of the Invention
[0004] To address the problem of inaccurate prediction of fractured-vuggy carbonate reservoirs in existing technologies, this invention proposes a planar prediction method, system, equipment, and storage medium for carbonate reservoirs. This method improves the interpretation accuracy of predictions based on the comprehensive analysis of multiple sensitive seismic attributes in the prediction of fractured-vuggy carbonate reservoirs.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] A planar prediction method for carbonate reservoirs includes the following steps:
[0007] Obtain the plane attribute data of the target layer in the carbonate rock seismic reservoir area, and select a variety of sensitive seismic attributes that are consistent with the geological understanding of the carbonate rock seismic reservoir area;
[0008] The earthquake attributes are arranged in corresponding order and merged into a whole prediction dataset;
[0009] The entire prediction dataset is used as input to perform prediction calculations through a deep learning algorithm prediction model, resulting in the final seismic facies prediction results and prediction probabilities for the entire working area of the carbonate rock seismic reservoir.
[0010] As a further improvement of the present invention, the method for establishing a deep learning algorithm prediction model includes the following steps:
[0011] Obtain the planar attribute data of the target layer in the work area, select a variety of sensitive seismic attributes that are consistent with the geological understanding of the work area, and output attribute data with consistent planar range according to the data storage format of planar seismic attributes.
[0012] Determine the number of reservoir seismic facies classifications within the target research area and establish a corresponding relationship with their assigned values;
[0013] Select the local area for creating patch labels, record the Line and Trace numbers of any two opposite corners of the rectangular area, select the reference attribute for creating patch labels, and determine the threshold range between different seismic phases.
[0014] Calculate patch labels for different local regions separately, and then merge multiple patch labels into a single training label dataset;
[0015] The training labeled dataset is used as input, and the model is trained using artificial intelligence deep learning algorithms to obtain a deep learning algorithm prediction model.
[0016] As a further improvement of the present invention, multiple target seismic properties include coherence, likelihood, amplitude variation rate, gradient structure tensor, and inverted wave impedance.
[0017] As a further improvement of the present invention, the number of reservoir seismic facies classifications within the target research area is determined, and a corresponding relationship is established between these classifications and assigned values. Specifically, this includes:
[0018] Based on the actual reservoir development characteristics of the work area, seismic facies were divided into three main categories: caverns, fracture zones, and matrix, corresponding to labels 2, 1, and 0, respectively.
[0019] As a further improvement of the present invention, the selection of the local area of the patch label needs to be based on the drilling information or geological understanding of the work area and seismic interpretation. The Line and Trace numbers at any diagonal position of the rectangular range are recorded. The grid is generated step by step according to the grid increment corresponding to the Line and Trace numbers. The reference attributes and threshold ranges of different seismic phases within each rectangular patch range are comprehensively analyzed. Then, the seismic attribute values within the local rectangular area corresponding to the patch label are filtered, and the calculated patch label is automatically generated.
[0020] As a further improvement of the present invention, after each rectangular patch label is calculated, they are concatenated in the same format to obtain a complete training dataset.
[0021] The earthquake attributes that need to be fused are formatted and multiple patch labels are merged into a single training label dataset according to the same format.
[0022] As a further improvement to the present invention, the specific steps for obtaining the deep learning algorithm prediction model are as follows:
[0023] Corresponding to the training and prediction steps of deep learning in artificial intelligence, the MLP algorithm is used to train the model using a training dataset to obtain data-driven model parameters, and then to obtain the prediction model of the deep learning algorithm.
[0024] A planar prediction system for carbonate reservoirs includes:
[0025] The data acquisition unit is used to acquire the plane attribute data of the target layer in the carbonate rock seismic reservoir area and select a variety of sensitive seismic attributes that are consistent with the geological understanding of the carbonate rock seismic reservoir area.
[0026] A data processing unit is used to arrange and merge the earthquake attributes in a corresponding order into a whole prediction dataset;
[0027] The prediction calculation unit is used to take the overall prediction dataset as input and perform prediction calculations through a deep learning algorithm prediction model to obtain the final seismic facies prediction results and prediction probabilities for the entire working area of the carbonate rock seismic reservoir.
[0028] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the carbonate reservoir planar prediction method.
[0029] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the carbonate reservoir planar prediction method.
[0030] The beneficial effects of this invention are reflected in:
[0031] This invention aims to improve the interpretation accuracy of comprehensive analysis and prediction of multiple sensitive seismic attributes in the prediction of fractured-vuggy carbonate reservoirs. It incorporates geological understanding and expert judgment as constraint labels into the fusion calculation of seismic attributes, achieving the effect of removing redundant information and eliminating anomalous features caused by non-research objectives. The final fusion yields a planar prediction result for carbonate reservoirs, providing the probability of different seismic facies predictions. This invention effectively incorporates the results of geological understanding and expert analysis into the constraint calculation of multi-attribute fusion in the form of data labels; it dramatically improves the interpretation accuracy of various sensitive seismic attributes to the pixel level (area size); it proposes a patch label method to solve the problem of batch label production, breaking away from the conventional approach of considering label data production from the direction of drilled curves, and instead producing labels from planar attribute data interpreted from area cells. This solves the dual problems of insufficient data sample types and insufficient sample data volume, enabling the application of deep learning algorithms in the highly heterogeneous carbonate rock field; it effectively solves the problem of inconsistent selection of thresholds for multiple dimensionless seismic attributes, predicting in the form of seismic facies, and also effectively solves the problem of planar boundary characterization. Attached Figure Description
[0032] Figure 1 The following are layer-by-layer plane diagrams of three commonly used post-stack seismic attributes in a carbonate rock research area, where (a) is the gradient structure tensor attribute, (b) is the coherence attribute, and (c) is the amplitude variation rate attribute.
[0033] Figure 2 Target area A patch label diagram: (a) Baseline reference attributes (b) Pixel-level patch label;
[0034] Figure 3 This is a schematic diagram of the patch labels for the target area B selected within the research area, where (a) is the baseline reference attribute and (b) is the corresponding pixel-level patch label.
[0035] Figure 4 Schematic diagram of outlier region inspection and analysis of amplitude change rate attribute;
[0036] Figure 5 Seismic profile of arbitrary line in anomaly region;
[0037] Figure 6 A plan view of three different patch labels;
[0038] Figure 7 A schematic diagram of the storage format of the merged training dataset;
[0039] Figure 8 Schematic diagram of MLP deep learning (fully connected neural network);
[0040] Figure 9 Comparison of the baseline reference attributes with labels A and B with the fused results: (a) Plan view of the baseline reference attributes after setting a certain threshold range; (b) Plan view of the seismic facies prediction results after multi-attribute fusion.
[0041] Figure 10 This is a schematic diagram of the preferred embodiment of the carbonate reservoir planar prediction method of the present invention;
[0042] Figure 11 This is a schematic diagram of the planar prediction system for carbonate reservoirs according to a preferred embodiment of the present invention;
[0043] Figure 12 This is a schematic diagram of the electronic device structure according to a preferred embodiment of the present invention. Detailed Implementation
[0044] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.
[0045] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.
[0046] BP artificial neural network attribute fusion currently mainly belongs to the second type of fusion calculation method. Compared with multiple linear regression, BP artificial neural network can represent nonlinear mapping. However, the current application usually uses shallow models with only one hidden layer of nodes. The network structure is simple and the nonlinear representation ability is poor. More importantly, it requires a large amount of training label data. Even using the longitudinal sample points of the drilled curve as training labels is still difficult to meet the needs of seismic attribute fusion. Therefore, it is even more difficult to use BP artificial neural network for the first type of planar attribute fusion, so there is little research in this direction.
[0047] To address the issue of poor nonlinear representation capabilities due to the simplistic structure of backpropagation (BP) artificial neural networks, the introduction of deep learning algorithms from current artificial intelligence (AI) can be considered. However, a common problem with deep learning algorithms in seismic reservoir prediction is the scarcity and difficulty in creating labeled training data samples, as well as the limited variety of sample types. This results in insufficient generalization ability of the prediction model, which is a major obstacle to the application of AI in seismic reservoir prediction. This is especially true for highly heterogeneous carbonate fracture-vuggy reservoirs. Currently, carbonate fracture-vuggy reservoirs are mainly composed of three types: caves, pores, and fractures. Apart from elastic parameters such as wave impedance, many other seismic attributes face debate regarding threshold selection. The delineation of carbonate fracture-vuggy reservoir boundaries is heavily influenced by human interpretation, lacking an ideal boundary delineation method. This is a problem that the industry has been working to solve.
[0048] To address the above problems, this invention provides a planar prediction method for carbonate reservoirs, comprising the following steps:
[0049] Obtain the plane attribute data of the target layer in the carbonate rock seismic reservoir area, and select a variety of sensitive seismic attributes that are consistent with the geological understanding of the carbonate rock seismic reservoir area;
[0050] The earthquake attributes are arranged in corresponding order and merged into a whole prediction dataset;
[0051] The entire prediction dataset is used as input to perform prediction calculations through a deep learning algorithm prediction model, resulting in the final seismic facies prediction results and prediction probabilities for the entire working area of the carbonate rock seismic reservoir.
[0052] Preferably, the method for establishing the deep learning algorithm prediction model includes the following steps:
[0053] Obtain the planar attribute data of the target layer in the work area, select a variety of sensitive seismic attributes that are consistent with the geological understanding of the work area, and output attribute data with consistent planar range according to the data storage format of planar seismic attributes.
[0054] Determine the number of reservoir seismic facies classifications within the target research area and establish a corresponding relationship with their assigned values;
[0055] Select the local area for creating patch labels, record the Line and Trace numbers of any two opposite corners of the rectangular area, select the reference attribute for creating patch labels, and determine the threshold range between different seismic phases.
[0056] Calculate patch labels for different local regions separately, and then merge multiple patch labels into a single training label dataset;
[0057] The training labeled dataset is used as input, and the model is trained using artificial intelligence deep learning algorithms to obtain a deep learning algorithm prediction model.
[0058] The above method is described in detail below. This invention proposes a planar multi-attribute fusion method for predicting carbonate rock seismic reservoirs based on artificial intelligence deep learning algorithms.
[0059] To fully explain the entire technical solution, the following two definitions should be clarified first:
[0060] Definition 1: Planar seismic attributes are data pairs stored within the target study area in XY Line Trace Z format. X and Y represent the x and y coordinates, Line is the main survey line number, Trace is the connecting survey line number, and Z value is the seismic attribute value at the corresponding location. Within the same study area, there is a one-to-one correspondence between XY and Line Trace data pairs, both of which can pinpoint specific planar locations.
[0061] Definition 2: Patch label, a training sample dataset created within a small rectangular area of the target research area, is assigned a value of 0, 1, 2, 3, etc. according to different numbers of seismic facies classifications. Different assignments correspond to different seismic facies or reservoir types. It is the result of datafication of the results of human comprehensive interpretation.
[0062] To achieve the above-mentioned objectives, the present invention provides the following technical solution:
[0063] Step 1: Organize the planar attribute data of the target layer in the research area, select the most sensitive seismic attributes that are consistent with the geological understanding of the research area, and output the attribute data with consistent planar range according to the data storage format in Definition 1.
[0064] Step 2: Determine the number of reservoir seismic facies classifications within the target study area and establish a correspondence with the numbers 0, 1, 2, etc.
[0065] Step 3: Select the local area for creating patch labels, record the Line and Trace numbers of any two opposite corners of the rectangular area, select the reference attribute for creating patch labels, and determine the threshold range between different seismic phases.
[0066] Step 4: Calculate the patch labels for different local regions separately, and merge multiple patch labels into a whole training label dataset.
[0067] Step 5: Arrange and merge the selected seismic attributes within the work area into a unified prediction dataset in the corresponding order.
[0068] Step 6: Use the training label dataset as input and train the model using artificial intelligence deep learning algorithms to obtain the prediction model parameters.
[0069] Step 7: Then use the prediction dataset as input to perform prediction calculations to obtain the final seismic facies prediction results and prediction probabilities for the entire work area.
[0070] In step 1, based on the detailed seismic horizon interpretation of the research area, various target seismic attributes are extracted. For carbonate fracture-vuggy reservoirs, coherence, likelihood, amplitude variation rate, gradient structure tensor, inversion wave impedance, etc. can be selected. The specific attributes to be included in the fusion calculation are analyzed and selected and output in a unified data format.
[0071] In step 2, seismic facies are classified based on the actual reservoir development characteristics of the work area. For simplification, they can usually be divided into three main categories: caverns (Class I reservoirs), fracture zones (Class II reservoirs), and matrix (non-reservoir), corresponding to labels 2, 1, and 0, respectively. Of course, more seismic facies types can be further subdivided, and corresponding additions will be made accordingly.
[0072] In step 3, the selection of local areas for patch labels needs to be based on the drilling information or geological understanding of the work area and seismic interpretation. The Line and Trace numbers at any diagonal position of the rectangular range are recorded, and the reference attributes and threshold ranges of different seismic phases within each rectangular patch range are comprehensively analyzed. These data will be used for the automatic generation and calculation of patch labels.
[0073] In step 4, after calculating the label of each rectangular patch, they are concatenated in the same format to obtain a complete training dataset. Taking three attributes as an example, the data label can be arranged in 8 columns: X YLine Trace Z1 Z2 Z3 L. Z1, Z2, and Z3 represent the three seismic attribute values to be fused (at the Line and Trace plane positions), and L is the data label value (0, 1, 2, etc.) at that pixel position.
[0074] In step 5, the seismic attributes to be fused are formatted, similar to the format described in the previous step, except that a label column is missing. Taking three seismic attributes as an example, the prediction dataset format is arranged in 7 columns: XY, LineTrace, Z1, Z2, and Z3.
[0075] Steps 6 and 7 correspond to the training and prediction steps of deep learning in artificial intelligence, respectively. They employ detailed algorithms and steps of MLP (also known as multilayer perceptron or fully connected neural network), a relatively mature computational method in deep learning. This technology involves complex computational steps and content, including neural network structure, neuron selection, initial weight settings, backpropagation, and loss functions, which will not be detailed here. In short, training with a labeled dataset yields data-driven model parameters, which are then used as input for prediction calculations to obtain the final seismic facies prediction results and probabilities within the research area.
[0076] 1. This invention incorporates geological understanding and expert comprehensive analysis results into the constraint calculation of multi-attribute fusion in the form of data tags;
[0077] 2. This invention has revolutionized the interpretation accuracy of various sensitive seismic attributes to the pixel level (area size of the work area);
[0078] 3. This invention innovatively proposes a patch label method, which solves the problem of batch training label production in batches. It breaks away from the conventional approach of thinking about label data production from the direction of drilled curves, and instead produces labels from planar attribute data interpreted by surface elements. This solves the dual problems of lack of data sample types and insufficient sample data volume, enabling deep learning algorithms to be applied in the field of carbonate rocks with extremely high heterogeneity.
[0079] 4. This invention can effectively solve the problem of inconsistent selection of threshold values for various dimensionless seismic attributes, and can predict seismic phases. It also effectively solves the problem of delineating planar boundaries.
[0080] This method is reflected in the following aspects: (1) It proposes a patch label production method for local target research areas with multiple planar attributes. This method can produce sample label data for training deep learning models from the local comprehensive analysis results, and effectively incorporate the human comprehensive analysis results into the fusion constraint calculation; (2) It improves the planar interpretation accuracy of seismic attributes to the pixel level (calculation and prediction are performed using the work area surface element as the pixel), which is a revolutionary improvement in the interpretation accuracy of multiple input seismic attributes; (3) The patch label production can select any selected sensitive seismic attribute that conforms to the geological understanding of the work area as the benchmark reference. Different patch labels can select different local thresholds, and can even be manually set one by one surface element pixel, which is very flexible; (4) The prediction result is a seismic phase planar distribution map, the reservoir boundary can be clearly seen, and the prediction results of different seismic phases can also give the prediction probability, which solves the problem of unclear threshold selection basis for many dimensionless seismic attributes.
[0081] like Figure 3As shown, another object of the present invention is to provide a planar prediction system for carbonate reservoirs, comprising:
[0082] The data acquisition unit is used to acquire the plane attribute data of the target layer in the carbonate rock seismic reservoir area and select a variety of sensitive seismic attributes that are consistent with the geological understanding of the carbonate rock seismic reservoir area.
[0083] A data processing unit is used to arrange and merge the earthquake attributes in a corresponding order into a whole prediction dataset;
[0084] The prediction calculation unit is used to take the overall prediction dataset as input and perform prediction calculations through a deep learning algorithm prediction model to obtain the final seismic facies prediction results and prediction probabilities for the entire working area of the carbonate rock seismic reservoir.
[0085] like Figure 4 As shown, a third objective of the present invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the carbonate reservoir planar prediction method.
[0086] A fourth objective of this invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the carbonate reservoir planar prediction method.
[0087] The following detailed description uses a specific embodiment to illustrate the technical solution of the method embodiment shown in the accompanying drawings, but this is not intended to limit the scope of the invention. The description is based on the accompanying drawings.
[0088] Figure 1 The following are layer-by-layer plane diagrams of three commonly used post-stack seismic attributes in a carbonate rock research area, where (a) is the gradient structure tensor attribute, (b) is the coherence attribute, and (c) is the amplitude variation rate attribute.
[0089] This embodiment utilizes seismic data from a specific work area. First, three commonly used seismic attributes for carbonate reservoir research were extracted from the 3D post-stack seismic data, such as... Figure 1 The gradient structure tensor, coherence, and amplitude change rate attributes shown in this embodiment are only selected for fusion. The method of this invention is not limited to three attributes and can theoretically achieve the fusion of any number of attributes. However, considering the actual situation and workload, it is usually recommended to select three to five sensitive attributes for fusion prediction.
[0090] Figure 2This is a schematic diagram of patch labels for target area A within the research area, where (a) is the baseline reference attribute and (b) is the corresponding pixel-level patch label. Target research area A is a specific drilled well location. The amplitude variation rate seismic attribute is selected as the baseline reference attribute. By determining the color scale range (i.e., the threshold range between different seismic phases), corresponding patch labels can be generated. Figure 2 (b), where red represents Class I reservoirs and is assigned a value of 2, yellow represents Class II reservoirs and is assigned a value of 1, and light blue represents the matrix and is assigned a value of 0.
[0091] Figure 3 This is a schematic diagram of the patch labels for the target area B selected within the research area, where (a) represents the baseline reference attribute and (b) represents the corresponding pixel-level patch label. Area B was created using patch labels from other drilled wells. The baseline reference attributes can be different. Even if the baseline reference attributes are the same, the threshold ranges for different seismic facies may differ from those in area A, depending on the actual drilling conditions. This allows for high flexibility.
[0092] Based on a comprehensive analysis of the actual drilling data in the work area, seismic facies are divided into three main categories: cavern reservoirs (Class I reservoirs), fractured zones (Class II reservoirs), and matrix (non-reservoir), corresponding to numerical labels of 2, 1, and 0, respectively. The amplitude variation rate is selected as the reference attribute for the patch label in this embodiment. The threshold ranges for the three seismic facies are locally adjusted at the drilled well locations, and the Line Trace number of the diagonal points of the selected local rectangular area is recorded, such as... Figure 2 and Figure 3 The diagrams shown illustrate the creation of patch labels for local areas A and B. The rectangular grid can be generated based on the Line Trace number. This embodiment uses automated programming; having the Line Trace number is equivalent to knowing the grid point numbers at the outermost four corners of the rectangular grid. The grid is generated incrementally according to the grid increment. Then, by setting threshold ranges between different seismic phases, the corresponding data label values are assigned. Finally, the seismic attribute values within the local rectangular area corresponding to the patch label are filtered out and written according to the patch label format mentioned earlier to complete the process.
[0093] Figure 4 For checking the plane anomaly region of amplitude change rate, the red broken line represents the location of an arbitrary line seismic profile, corresponding to... Figure 5 Arbitrary line seismic profile, in Figure 4 After setting the threshold ranges for different seismic phases according to the patch label analysis, the yellow areas (Class II reservoirs) in many other parts of the work area were not caused by reservoir development, such as... Figure 5As shown, there are no obvious reflection characteristics of carbonate reservoir development within the target layer below a depth of 6500, which is why multiple single seismic attributes cannot fully meet the requirements for seismic reservoir prediction. To separate the false from the true, this local area can be created as a matrix patch tag, such as... Figure 6 The rectangular label in the bottom right corner is for illustrative purposes. Figure 6 The display includes the two patch labels A and B mentioned above, as well as the matrix label C from the Southern Integrated Analysis.
[0094] Figure 7 This is the training dataset after merging the three patch labels (only a portion is shown). From left to right, the six columns represent Line Trace Z1 Z2 Z3 L, where Z1, Z2, and Z3 correspond to... Figure 1 The three attribute values are: L is the label column, and only Line Trace is used to locate the elements (equivalent to XY).
[0095] After generating three patch labels, they can be merged into a single training dataset (the number of patch labels is not limited to this; theoretically, the more labels and the more evenly distributed they are within the work area, the better the training and prediction results). The data format is as follows: Figure 7 As shown, the six columns from left to right are Line Trace Z1 Z2 Z3 L, where Z1, Z2, and Z3 correspond to... Figure 1 The dataset contains three attribute values, with L representing the label column. Here, only Line Trace is used to locate the elements (equivalent to XY data). After creating the training dataset, the corresponding prediction dataset for the three seismic attributes of the entire work area is created according to step 5 of this invention. The format corresponds to the training dataset, except that the label column is missing. What we want to predict is the label column, which is also the target of the next prediction.
[0096] After creating the training and prediction datasets, the basic data requirements for deep learning in artificial intelligence are met. This section mainly covers the specific steps of the deep learning algorithm, a relatively mature technology. This embodiment uses the publicly released Google Tensorflow.keras open-source framework for deep learning. After converting the training dataset to one-hot encoding, a seven-layer (including input and output layers, hence five hidden layers) deep learning model is trained. Each hidden layer has 256 neurons. The neurons are activated using the ReLU activation function (the sigmoid activation function can also be used; in this example, the ReLU function has better learning convergence, so it is chosen). The settings of other hyperparameters can be adjusted to some extent. Each step of the deep learning process will not be described in detail here. This invention is not limited to specific parameter and hyperparameter settings. After training, prediction calculations can be performed, such as... Figure 9 As shown in (b), when the prediction results are compared with the selected benchmark reference attributes, the prediction results are almost completely consistent with the label results at the two patch label positions A and B indicated by the arrows. The outliers in the remaining yellow areas show a significant improvement effect. The prediction probabilities of different seismic phases can be output through the softmax function in the open source framework.
[0097] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0098] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0099] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0100] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for planar prediction of carbonate reservoirs, characterized in that, Includes the following steps: Obtain the plane attribute data of the target layer in the carbonate rock seismic reservoir area, and select a variety of sensitive seismic attributes that are consistent with the geological understanding of the carbonate rock seismic reservoir area; The earthquake attributes are arranged in corresponding order and merged into a whole prediction dataset; The entire prediction dataset is used as input to perform prediction calculations through a deep learning algorithm prediction model, resulting in the final seismic facies prediction results and prediction probabilities for the entire work area of the carbonate rock seismic reservoir. The method for building a prediction model using deep learning algorithms includes the following steps: Obtain the planar attribute data of the target layer in the work area, select a variety of sensitive seismic attributes that are consistent with the geological understanding of the work area, and output attribute data with consistent planar range according to the data storage format of planar seismic attributes. Determine the number of reservoir seismic facies classifications within the target research area and establish a corresponding relationship with their assigned values; Select the local area for creating patch labels, record the Line and Trace numbers of any two opposite corners of the rectangular area, select the reference attribute for creating patch labels, and determine the threshold range between different seismic phases. Calculate patch labels for different local regions separately, and then merge multiple patch labels into a single training label dataset; After calculating the label of each rectangular patch, they are concatenated in the same format to obtain a complete training dataset; The earthquake attributes that need to be fused are formatted and multiple patch labels are merged into a whole training label dataset according to the same format. For the three data labels, the data format is arranged in 8 columns: XY Line Trace Z1 Z2 Z3 L. Z1, Z2, and Z3 represent the three seismic attribute values that need to be fused at the Line and Trace plane locations, respectively, and L is the data label value of the surface cell location. X and Y are the x-coordinate and y-coordinate, Line is the main survey line number, and Trace is the connecting survey line number. Within the same work area, there is a one-to-one correspondence between XY and Line / Trace data pairs, both locating specific plane locations. The training label dataset is used as input, and the model is trained using artificial intelligence deep learning algorithms to obtain a deep learning algorithm prediction model. Multiple target seismic properties include coherence, likelihood, amplitude variation rate, gradient structure tensor, and inverted wave impedance; The selection of local areas for patch labels needs to be based on existing drilling information or geological knowledge of the work area and seismic interpretation. The Line and Trace numbers at any diagonal position of the rectangular range are recorded. The grid is generated step by step according to the grid increment corresponding to the Line and Trace numbers. The reference attributes and threshold ranges of different seismic phases within each rectangular patch range are comprehensively analyzed. Then, the seismic attribute values within the local rectangular area corresponding to the patch label are filtered, and the calculated patch label is automatically generated.
2. The method according to claim 1, characterized in that: Determine the number of reservoir seismic facies classifications within the target study area and establish a corresponding relationship with their assigned values, specifically including: Based on the actual reservoir development characteristics of the work area, seismic facies were divided into three main categories: caverns, fracture zones, and matrix, corresponding to labels 2, 1, and 0, respectively.
3. The method according to claim 1, characterized in that: The specific steps to obtain the prediction model using the deep learning algorithm are as follows: Corresponding to the training and prediction steps of deep learning in artificial intelligence, the MLP algorithm is used to train the model using a training dataset to obtain data-driven model parameters, and then to obtain the prediction model of the deep learning algorithm.
4. A planar prediction system for carbonate reservoirs, based on the planar prediction method for carbonate reservoirs according to any one of claims 1-3, characterized in that, include: The data acquisition unit is used to acquire the plane attribute data of the target layer in the carbonate rock seismic reservoir area and select a variety of sensitive seismic attributes that are consistent with the geological understanding of the carbonate rock seismic reservoir area. A data processing unit is used to arrange and merge the earthquake attributes in a corresponding order into a whole prediction dataset; The prediction calculation unit is used to take the overall prediction dataset as input and perform prediction calculations through a deep learning algorithm prediction model to obtain the final seismic facies prediction results and prediction probabilities for the entire working area of the carbonate rock seismic reservoir.
5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the carbonate reservoir planar prediction method according to any one of claims 1-3.
6. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the planar prediction method for carbonate reservoirs according to any one of claims 1-3.