Multi-scale fracture development compact oil and gas reservoir production dynamic rapid prediction method

Abstract

The invention discloses a multi-scale fracture development compact oil and gas reservoir production dynamic rapid prediction method, and belongs to the technical field of oil exploitation, and the method comprises the following steps: 1, obtaining a multi-scale fracture distribution data sample and a production simulation data sample, and constructing a sample library; 2, establishing a crack network diagram to represent data; 3, calculating a vertex connection strength matrix of the crack network based on the graph representation data; 4, establishing a depth map and time sequence mixed learning model for rapid prediction of the production dynamic state of the compact oil and gas reservoir; 5, training the depth map and time sequence mixed learning model, and verifying the performance of the model after the training is completed; and step 6, carrying out compact oil and gas reservoir production dynamic rapid prediction based on the trained model. According to the method, the production dynamic prediction efficiency of the compact oil and gas reservoir can be improved, and a basis is provided for history fitting, production optimization and the like of the compact oil and gas reservoir.

Description

Technical Field

[0001] The invention belongs to the technical field of oil production, and in particular relates to a method for quickly predicting the production dynamics of a dense oil and gas reservoir with multi-scale fracture development. Background Art

[0002] Tight oil and gas reservoirs have low permeability and well-developed natural fractures. Effective development requires hydraulic fracturing to transform the reservoir to form a complex fracture network. Therefore, the distribution of fractures greatly affects the production dynamics of tight oil and gas reservoirs.

[0003] The prediction of oil and gas reservoir production dynamics is to predict the production dynamics data such as the output and bottom hole pressure of production wells based on known oil and gas reservoir information such as geological parameters and development measures. The oil and gas reservoir production dynamics prediction technology can analyze the production changes of oil and gas reservoirs and predict the future production performance of oil and gas reservoirs, so as to timely optimize the production management of oil and gas reservoirs. Typical oil and gas reservoir production dynamics prediction methods include reservoir engineering methods and numerical simulation methods. Among them, the reservoir engineering method uses empirical formulas or charts to predict the production dynamics of oil and gas reservoirs, such as decline curve analysis or analytical methods. Its considerations are relatively simple, the scope of application is limited, and the prediction accuracy is low; the numerical simulation method is currently the most commonly used method for predicting oil and gas reservoir production dynamics. It simulates and predicts oil and gas reservoir production dynamics by solving the complex partial differential equations of oil and gas reservoir fluid flow through numerical calculations. It has high calculation accuracy and can fully consider all aspects of oil and gas reservoir information, but it has the disadvantages of low calculation efficiency and long solution time. For tight oil and gas reservoirs with complex multi-scale fracture distribution, typical numerical simulation methods include dual medium model, discrete fracture model and embedded discrete fracture model, etc. It is usually necessary to consider the fracture distribution and establish the flow equations of fluid in the matrix and fractures and the fluid exchange equations between the fractures and the matrix. The complex partial differential equations are numerically solved by finite difference to perform production dynamic calculations, which has low computational efficiency and large workload.

[0004] Considering that the numerical simulation method is labor-intensive and computationally inefficient for predicting the production dynamics of oil and gas reservoirs, in order to overcome this problem, many alternative models of numerical simulation methods have been constructed using deep learning methods to predict the production dynamics of oil and gas reservoirs. However, the current deep learning-based prediction model for oil and gas reservoir production dynamics mainly targets the continuous geological attributes of conventional oil and gas reservoirs, such as permeability, porosity and other parameters, and uses deep convolutional neural networks for feature learning and extraction, which is difficult to effectively apply to tight oil and gas reservoirs involving complex multi-scale discrete fracture network distribution. Fractures are the main flow channels of tight oil and gas reservoir fluids, and the distribution of fractures plays a leading role in the prediction of tight oil and gas reservoir production dynamics. However, fractures are discretely distributed in the reservoir space, with multi-scale, irregular and other characteristics. Conventional deep convolutional neural networks are difficult to effectively learn data with discrete characteristics in space. Therefore, the present invention provides a method for fast prediction of tight oil and gas reservoir production dynamics based on a deep learning model, and in particular, a graph representation method and a deep graph learning method suitable for complex discrete fracture networks are established, which greatly improves the computational efficiency of tight oil and gas reservoir production dynamics prediction while ensuring accuracy. Summary of the invention

[0005] To solve the above problems, the present invention proposes a method for rapid prediction of production dynamics of tight oil and gas reservoirs with multi-scale fracture distribution. On the basis of quantitatively characterizing the connectivity relationship of the fracture network, a deep graph learning module is established in combination with an improved Transformer model to perform feature learning and feature extraction on discrete fractures, and the extracted fracture network features are passed to a deep time series learning module to realize the prediction of production dynamic indicators of production wells in tight oil and gas reservoirs. The deep graph learning and deep time series learning modules are combined to construct an end-to-end deep graph and time series hybrid learning model to realize rapid prediction of production dynamics of tight oil and gas reservoirs.

[0006] The technical solution of the present invention is as follows: A method for quickly predicting the production performance of a tight oil and gas reservoir with multi-scale fracture development comprises the following steps: Step 1: Obtain multi-scale fracture distribution data samples and production simulation data samples, and build a sample library; Step 2, establishing a crack network diagram to represent the data; Step 3, calculating the vertex connectivity strength matrix of the fracture network based on the graph representation data; Step 4: Establish a hybrid learning model of depth map and time series for rapid prediction of tight oil and gas reservoir production dynamics; Step 5: Train the depth map and time series hybrid learning model, and verify the performance of the model after the training is completed; Step 6: Rapidly predict the production dynamics of tight oil and gas reservoirs based on the trained model.

[0007] Furthermore, the specific process of step 1 is as follows: Step 1.1, stochastically model the multi-scale fracture distribution of tight oil and gas reservoirs, and establish a multi-scale fracture distribution data sample containing several groups of different fracture distributions; Step 1.2, using an embedded discrete fracture numerical simulation method to perform numerical simulation calculations on different fracture distributions, and obtain corresponding production simulation data samples, where the production simulation data samples are production dynamic data of production wells, including daily oil production, daily gas production, daily water production, and bottom hole pressure; Step 1.3: The fracture distribution data samples and the corresponding production simulation data samples are combined into a sample library, and the sample library is divided into a training set and a test set in proportion.

[0008] Furthermore, in step 2, a group of fractures are distributed as a fracture network, including multiple discrete fractures, and the parameters of each fracture include length, position, direction, aperture, porosity and permeability; a graphical representation of the fracture network is established based on the discrete fracture parameters, that is, , Represents the crack network diagram to represent the data, represents the set of vertices of the graph, Represents the edge set of the graph; the two endpoints of each single fracture are used as the vertices of the graph, and the position coordinates of the fracture endpoints are used as vertex features; the fractures are used as the edges of the graph, and the length, direction, aperture, porosity and permeability of the fractures are used as edge features; intersecting fractures are re-divided from the intersection.

[0009] Furthermore, in step 3, the connectivity strength of the fracture vertices is calculated based on the graph representation data of the fracture network, and a vertex connectivity strength matrix of the fracture network is constructed; the calculation formula for the connectivity strength between any two vertices in the matrix is: (1); In the formula, Indicates The vertices and The connectivity strength between vertices corresponds to the first Line Elements of a column; Indicates The vertices and The connection distance between vertices; Indicates the total number of vertices.

[0010] Furthermore, the specific process of step 4 is as follows: Step 4.1, establish a deep graph learning module for feature learning and extraction of graph representation data of the fracture network; Step 4.2: Establish a deep time series learning module for production dynamic data prediction of production wells; Step 4.3, combine the deep graph learning module and the deep time series learning module to obtain an end-to-end deep graph and time series hybrid learning model; the input data of the deep graph and time series hybrid learning model is the graph representation data of the fracture network and the vertex connectivity strength matrix of the fracture network, and the output is production dynamic prediction data.

[0011] Furthermore, in step 4.1, the deep graph learning module includes an embedding layer, an information propagation aggregation model, an improved Transformer model, and a global pooling layer; the specific working process is as follows: Step 4.1.1: Represent the data in the fracture network diagram Input embedding layer, which uses two independent fully connected layers to perform feature mapping on vertex features and edge features of graph data respectively; Step 4.1.2: Use the information propagation aggregation model to fuse the edge features after embedding layer mapping into the vertex features; specifically: the information propagation aggregation model fuses the first The vertices and The edges between vertices The feature is propagated to the edge corresponding to Vertices and Vertices , use vector concatenation and merging operations to merge edge features into vertex feature vectors, and use a layer of fully connected neural network to perform a mapping calculation on the merged vectors at the vertex to obtain the updated vertex features; Step 4.1.3: Based on the improved Transformer model, perform multi-level update and feature transformation on the vertex features after integrating the edge features; the improved Transformer model embeds the connectivity strength matrix into the learning and prediction process of the original Transformer model, and uses the vertex connectivity strength matrix The calculation process of the crack network structure encoding information participating in an attention operation is as follows: (2); In the formula, Indicates Updated vertex features output by the layer encoder; Indicates The dimension of the vertex features after the layer encoder updates; is the softmax function; , and Represents the query matrix, key matrix, and value matrix in the attention operation; For the The dimension of the vertex features after the layer encoder updates; is a learnable weight matrix; is the transpose symbol; Step 4.1.4: Use the global pooling layer to pool all vertex features into a feature vector as the output of the deep graph learning module. The feature vector is the high-level feature of the extracted crack network. .

[0012] Furthermore, in step 4.2, the deep time series learning module uses the original Transformer model to perform production dynamic time series prediction; the specific working process is: Step 4.2.1: High-level features of the fracture network Copy to time steps as input vectors, and added with the time sequence coding information as the input of the encoder; the time sequence coding information is obtained by encoding the time sequence of the production dynamic data; Step 4.2.2: The multi-head self-attention mechanism inside the encoder performs feature transformation on the time series; Step 4.2.3: Use the fully connected layer as the output layer to output the production dynamic prediction value corresponding to each time step , ,in, To forecast the number of indicators for production; Indicates the number of producing wells; Indicates the number of production dynamic indicator types; The number of forecast time steps representing the production dynamics.

[0013] Furthermore, in step 5, the depth map and time series hybrid learning model is trained in combination with the training set divided in step 1, the model parameters are optimized to minimize the loss function, and the mean absolute error is used as the loss function of the model. The calculation formula is as follows: (3); In the formula, is the mean absolute error; Indicates the number of samples; and Respectively represent The true value of the sample and the model prediction result; After training is completed, the performance of the model is verified using the test set divided in step 1. The verification indicator is the determination coefficient, and the formula is as follows: (4); In the formula, is the coefficient of determination; express The average of the true values ​​of the samples.

[0014] Furthermore, in step 6, the depth map and time series hybrid learning model that has been trained and tested with good performance is output, and the fracture distribution parameters of the tight oil and gas reservoir and the vertex connectivity strength matrix corresponding to the calculated fracture distribution are used as model input to quickly predict the corresponding production dynamic data.

[0015] Beneficial technical effects brought by the present invention: Based on the discrete connected structure of the fracture network and the improved Transformer model, the present invention establishes a depth map and time series hybrid learning model for the production dynamic prediction task of tight oil and gas reservoirs with complex fracture network distribution. The model can quickly predict the production dynamics of tight oil and gas reservoir production wells according to the fracture distribution. The rapid prediction model constructed by the present invention can predict the production dynamics of tight oil and gas reservoirs, and can replace the reservoir numerical simulation method in production optimization, automatic history matching and other reservoir engineering tasks involving production dynamic prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The present invention is a flowchart of a method for rapidly predicting production dynamics of tight oil and gas reservoirs with multi-scale fracture distribution.

[0017] Figure 2 Schematic diagram of a tight oil reservoir block containing multi-scale fracture distribution in an embodiment of the present invention.

[0018] Figure 3 Schematic diagrams of three groups of different crack distributions randomly selected in the embodiments of the present invention; wherein (a), (b), and (c) represent schematic diagrams of the first group of crack distributions, the second group of crack distributions, and the third group of crack distributions, respectively.

[0019] Figure 4 It is a daily water production curve diagram of production well 1 corresponding to three groups of different fracture distributions in an embodiment of the present invention.

[0020] Figure 5 It is a daily oil production curve diagram of production well 1 corresponding to three groups of different fracture distributions in an embodiment of the present invention.

[0021] Figure 6 This is a graph showing how the loss function changes with iteration during the training of the depth map and time series hybrid learning model in an embodiment of the present invention.

[0022] Figure 7 This is a crack network distribution diagram of a test sample in an embodiment of the present invention.

[0023] Figure 8 It is a comparison chart of the daily water production model prediction results and the numerical simulation results of the production well 1 in the embodiment of the present invention.

[0024] Fig. 9It is a comparison chart of the daily oil production model prediction result and the numerical simulation result of the production well 1 in the embodiment of the present invention. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: The present invention models the multi-scale complex discrete fracture network widely distributed in tight oil and gas reservoirs as graph structure data; establishes a quantitative method to calculate the connectivity relationship of the fracture network; combines the fracture network connectivity relationship with the Transfomer neural network model to establish a deep graph learning module and a deep time series learning module to respectively perform multi-scale discrete fracture network feature learning and tight oil and gas reservoir production dynamic data prediction; integrates the deep graph learning module and the deep time series learning module to form a method for rapid production dynamic prediction of tight oil and gas reservoirs.

[0026] like Figure 1 As shown, a method for rapid prediction of production performance of tight oil and gas reservoirs with multi-scale fracture distribution specifically includes the following steps: Step 1: Obtain multi-scale fracture distribution data samples and production simulation data samples and build a sample library; the specific process is as follows: Step 1.1, stochastic modeling of the multi-scale fracture distribution of tight oil and gas reservoirs, and establishing a multi-scale fracture distribution data sample containing several groups of different fracture distributions; the description parameters of the fracture distribution include the location, direction, length, opening, porosity, permeability and other statistical information of the fracture; wherein, the location of the fracture can be obtained by a method in uniform distribution, Poisson process, Cluster process, Cox process or Markov process; the direction of the fracture can be obtained by a method in uniform distribution, Fisher distribution, von-Mises distribution or normal distribution; the fracture length distribution can be obtained by a method in uniform distribution, fractal simulation, power law distribution; the opening and porosity of the fracture need to be obtained through core analysis data and well logging analysis data; the permeability of the fracture can be calculated according to the fracture length combined with the cubic law. The present invention uses a discrete fracture model to explicitly characterize the fracture distribution, and the fracture is stochastically statistically modeled according to the fracture group, and a multi-scale fracture distribution data sample containing 1000 to 2000 groups of different fracture distributions is constructed.

[0027] Step 1.2, using the embedded discrete fracture numerical simulation method to perform numerical simulation calculations on different fracture distributions, and obtain corresponding production simulation data samples, where the production simulation data samples are production dynamic data of production wells, including daily oil production OPR, daily gas production GPR, daily water production WPR, and bottom hole pressure BHP; Step 1.3: The fracture distribution data samples and the corresponding production simulation data samples are combined into a sample library, and the sample library is divided into a training set and a test set in proportion.

[0028] Step 2, establishing a crack network diagram to represent the data; A group of fractures is distributed as a fracture network, which contains multiple discrete fractures. Each fracture is described by parameters such as length, position, direction, aperture, porosity and permeability. A graphical representation of the fracture network is established based on the discrete fracture parameters, i.e. , Represents the crack network diagram to represent the data, represents the set of vertices of the graph, Represents the edge set of the graph, which is composed of each fracture. The two endpoints of each single fracture are used as the vertices of the graph, and the position coordinates of the fracture endpoints are used as vertex features, also known as vertex attributes; fractures are used as edges of the graph, and the length, direction, aperture, porosity and permeability of the fractures are used as edge features, also known as edge attributes; intersecting fractures are re-divided from the intersection.

[0029] Step 3, calculating the vertex connectivity strength matrix of the fracture network based on the graph representation data; Based on the graph representation data of the fracture network obtained in step 2, the connectivity strength of the fracture vertices is calculated, and the vertex connectivity strength matrix of the fracture network is obtained. The connection distance of the fracture vertices is quantitatively calculated through the shortest connection path between the fracture vertices and the length of the fracture edge. The connectivity strength of the fracture vertices is calculated based on the connection distance, and the discrete structural relationship of the fracture network is quantitatively modeled. Specifically, the connectivity strength between any two vertices is quantitatively calculated through the connection distance between the vertices. In the fracture network, the vertices as the endpoints of the fractures are only connected to each other through the fracture edges. Therefore, the distance between the vertices can be accurately calculated by the length of the fracture edges connecting the vertices. There may be one or more connection paths between the two vertices, and the connection distance is calculated using the shortest path. The larger the distance between the fracture vertices, the weaker the connectivity relationship between the two vertices, that is, the smaller the connectivity strength index; conversely, the smaller the distance between the vertices, the stronger the connectivity relationship between the vertices, that is, the larger the connectivity strength index. After calculating the connection distances corresponding to all the vertices of the fracture network, a normalized calculation is performed to obtain the vertex connectivity strength matrix of the fracture network. ; The calculation formula for the connectivity strength between any two vertices is: (1); In the formula, Indicates The vertices and The connectivity strength between vertices corresponds to the first Line Elements of a column; Indicates The vertices and The connection distance between vertices; Indicates the total number of vertices.

[0030] The connectivity strength calculated by this method can achieve the maximum connectivity strength between vertices with the smallest connection distance and the minimum connectivity strength between vertices with the largest connection distance. For two crack vertices that do not have a connection relationship, their connectivity strength is set to 0.

[0031] Step 4: Establish a depth map and time series hybrid learning model for rapid prediction of tight oil and gas reservoir production dynamics; the specific process is as follows: Step 4.1, establish a deep graph learning module suitable for feature learning of fracture network graph representation data. The main function of the deep graph learning module is to learn and extract features of fracture network graph representation data; The deep graph learning module uses an information propagation aggregation model to fuse the edge features of the graph representation data of the fracture network into the vertex features. It further performs multi-level updates and feature transformations on the vertex features of the graph based on the improved Transformer model. The connectivity strength matrix of the fracture vertices provides additional structural encoding information for the improved Transformer model, models the discrete structural information of the fracture network, and extracts high-level features of the discrete fracture network. The input data of the deep graph learning module is the fracture network graph representation data established in step 2 And the vertex connectivity strength matrix of the fracture network obtained in step 3 The deep graph learning module includes an embedding layer, an information propagation aggregation model, an improved Transformer model, and a global pooling layer; the specific working process is as follows: Step 4.1.1: Represent the data in the fracture network diagram Input embedding layer, which uses two independent fully connected layers to perform feature mapping on vertex features and edge features of graph data respectively; Step 4.1.2: Use the information propagation aggregation model to fuse the edge features of the graph representation data after the embedding layer mapping into the vertex features; specifically: the information propagation aggregation model fuses the first The vertices and The edges between vertices The feature is propagated to the edge corresponding to Vertices and Vertices , using vector concatenation and merging operations, the edge features are merged into the vertex feature vector, and a fully connected neural network is used to perform a mapping calculation on the merged vectors at the vertex to obtain the updated vertex features; this operation can achieve the fusion of the edge features of the crack into the vertex features, and only the vertex features are updated in subsequent learning, avoiding the need to update and learn the edge and vertex features independently; Step 4.1.3, based on the improved Transformer model, the vertex features after the edge features are integrated are updated and transformed at multiple levels. The vertex connectivity strength matrix of the crack network provides the improved Transformer model with the spatial structure encoding information between the vertices of the crack network, and models the discrete structure information of the crack network. The improved Transformer model uses the encoder as the basic unit. An improved Transformer model consists of The encoders are connected in series. An encoder consists of multi-head attention operations, normalization operations, linear operations, and addition operations. In order to embed the vertex connectivity strength matrix into the learning and prediction process of the original Transformer model, the vertex connectivity strength matrix is ​​used. The calculation process of the crack network structure encoding information participating in an attention operation is as follows: (2); In the formula, Indicates Updated vertex features output by the layer encoder; Indicates The dimension of the vertex features after the layer encoder updates; is the softmax function; , and represents the query matrix (vector), key matrix (vector), and value matrix (vector) in the attention operation, which can be obtained by The vertex features of the layer encoder input (i.e. The updated vertex features output by the layer encoder) calculate, For the The dimension of the vertex features after the layer encoder updates, , , , , and is a learnable weight matrix; is a learnable weight matrix used to adaptively adjust the vertex connectivity strength matrix How important is it in the computation of attention? is the transpose symbol.

[0032] Formula (2) represents a single attention calculation, and multi-head attention is repeated calculation The output of the encoder is obtained through the concatenation operation. Formula (2) integrates the vertex connectivity strength matrix of the crack network based on the basic attention calculation operation, which is an improvement on the Transformer encoder attention calculation method. is the basic attention calculation, which allocates and calculates attention to all crack vertex features and has global characteristics; is the vertex connectivity strength matrix of the fracture network. The connectivity strength of unconnected vertices is 0. It is a representation of the discrete spatial structure of the fracture network and has local characteristics. Therefore, formula (2) integrates the global characteristics with the local characteristics, that is, it retains the original Transformer model's autonomous learning of the relationship between fracture vertices and embeds the inherent structural relationship of fracture vertices, which enables the improved Transformer model to actively learn and update the fracture network vertex features.

[0033] Step 4.1.4. Finally, use the global pooling layer to pool all vertex features into a feature vector as the output of the deep graph learning module. The feature vector is the high-level feature of the extracted crack network. .

[0034] Step 4.2: Establish a deep time series learning module for production dynamic data prediction of production wells; The input of the deep time series learning module is the high-level features of the fracture network extracted in step 4.1. The basic unit of this module is the encoder, which combines the time series coding to maintain the time sequence information of the production dynamic data and predicts the production dynamic data of the production well corresponding to each time step; specifically, the deep time series learning module uses the original Transformer model for production dynamic time series prediction. The encoder is used as the basic unit. An encoder consists of multi-head attention operation, normalization operation, linear operation, addition operation and time series coding. The encoder predicts the production dynamic data by time series changes of the extracted fracture features through the multi-head self-attention mechanism, and the time series coding information is obtained by the time series coding of the production dynamic data.

[0035] The Transformer model of the deep temporal learning module consists of The encoders are connected in series, and the specific working process is as follows: Step 4.2.1: High-level features of the fracture network Copy to time steps as the input vector and add it with the temporal coding information as the input of the encoder; Step 4.2.2: The multi-head self-attention mechanism inside the encoder performs feature transformation on the time series; Step 4.2.3: Finally, use the fully connected layer as the output layer to output the production dynamic prediction value corresponding to each time step. , ,in, To forecast the number of indicators for production; Indicates the number of producing wells; Indicates the number of production dynamic indicator types; The number of forecast time steps representing the production dynamics.

[0036] Step 4.3, combine the deep graph learning module and the deep time series learning module to obtain an end-to-end deep graph and time series hybrid learning model; the input data of this model is the graph representation data of the fracture network and the vertex connectivity strength matrix of the fracture network, and the output is production dynamic prediction data.

[0037] Step 5: Use the divided training set to train the depth map and time series hybrid learning model. After the training is completed, use the test set to verify the performance of the model; Combine the training set divided in step 1 to train the depth map and time series hybrid learning model, optimize the model parameters to minimize the loss function, and use the mean absolute error (MAE) as the loss function of the model. The calculation formula of MAE is as follows: (3); In the formula, is the mean absolute error; Indicates the number of samples; and Respectively represent The true value of the sample and the model prediction result.

[0038] After training is completed, the performance of the model is verified using the test set divided in step 1. The verification indicator is the determination coefficient, and the formula is as follows: (4); In the formula, is the coefficient of determination; express The closer the coefficient of determination is to 1, the closer the model's prediction result is to the numerical simulation result, and the more reliable the model performance is.

[0039] Step 6: Rapidly predict the production dynamics of tight oil and gas reservoirs based on the trained model; The output of the training-completed and tested deep graph and time series hybrid learning model with good performance uses the fracture distribution parameters of tight oil and gas reservoirs and the vertex connectivity strength matrix corresponding to the calculated fracture distribution as model input to quickly predict the corresponding production dynamic data.

[0040] In order to demonstrate the feasibility and superiority of the present invention, the following examples are given.

[0041] Figure 2 It is a tight oil reservoir block with multi-scale fracture distribution, including four production wells (numbered as production wells 1 to production wells 4) and one water injection well (numbered as water injection well 1). Figure 2 The random simulation constructed a multi-scale fracture distribution data sample containing 1200 groups of different fracture distributions; three groups of different fracture distributions were randomly selected as follows: Figure 3 As shown, Figure 3 In the figure, (a), (b), and (c) represent the first group of fracture distribution, the second group of fracture distribution, and the third group of fracture distribution, respectively. The embedded discrete fracture numerical simulation method is used to perform numerical simulation calculations on different fracture distributions, and 1200 groups of production dynamic data samples corresponding to different fracture distributions are obtained. In this embodiment, the production dynamic data samples include the daily oil production and daily water production of four production wells. Figure 4 and Figure 5 The following are the comparison results of daily water production and daily oil production of production well 1 corresponding to three groups of different fracture distributions. Figure 4 and Figure 5 It can be seen that different fracture distributions will lead to different production performances, that is, the daily water production and daily oil production of the production wells are different.

[0042] All fracture distributions and corresponding production data are combined into a sample library, which is divided into a 5:1 ratio, with 1,000 samples as training sets and the other 200 samples as test sets. According to the process in step 3, the graph representation data of the discrete fracture network is established, and the vertex connectivity strength matrix of the fracture network is calculated based on the graph representation data of the fractures; Then, a deep graph learning module is established to extract the features of the current crack distribution. First, after the embedding layer, the initial feature transformation is performed, and the vertex attribute features and edge attribute features are mapped into vertex embedding features and edge embedding features through two different one-layer fully connected neural networks. The number of neurons in the fully connected layer is set to [128×1], the activation function is set to the Sigmoid function, and the embedding feature dimensions of vertices and edges are uniformly set to [128×1]; secondly, the edge features are fused with the vertex features, and the edge features mapped by the embedding layer are fused into the vertex features using the information propagation aggregation model, and the features of each edge are propagated to the two vertices corresponding to each edge, and the edge features are merged into the vertex feature vector using the vector concatenation and merge operation. A fully connected neural network is used to perform a mapping calculation on the merged vectors at the vertices. The number of neurons in the fully connected layer is set to [128×1], and the activation function is set to the ReLU function to obtain updated vertex features with a dimension of [128×1]. Then, the vertex features fused with the edge features and the obtained vertex connectivity strength matrix are input into the encoder calculation process of the Transformer. In this step, a 2-layer encoder is used to update the vertex features, and the vertex feature dimension is kept at [128×1] during the update process. Finally, a vertex-based global pooling layer is used to perform a maximum pooling operation on all vertex features to obtain high-level features of the crack network.

[0043] A deep time series learning module is constructed to predict production dynamic data. The deep time series module consists of two layers of encoders and one layer of fully connected layer distributed on the time step. The input is the extracted high-level features of the fracture network, which are copied on the time step as the input parameters of each time step through the vector copy operation. The feature transformation on the time series is performed through a two-layer Transformer encoder. Finally, a fully connected layer wrapped by a time distribution layer is used as the output layer to output the production dynamic prediction value corresponding to each time step.

[0044] The deep graph learning module and the deep time series learning module are combined to construct a deep graph and time series hybrid learning model. The input of the model is the graph representation data of the fracture network and the vertex connectivity strength matrix, and the output is the production dynamic data. The mean absolute error (MAE) is used as the loss function of the model. The model is trained in combination with the divided training set, and the weight parameters of the model are optimized to minimize the loss function. Figure 6 It shows the change of mean absolute error loss function with iterations during model training. It can be seen that the loss function has converged after 80 iterations.

[0045] After training is completed, the performance of the model is verified using the divided test set, and the determination coefficient of the model on the test set is calculated. Figures 7 to 9 The prediction effect of the trained model on a test sample is shown. The test sample is selected based on the median value of the determination coefficient of the prediction results of 200 test samples, so it is representative. Figure 7 In order to test the fracture network distribution of the test samples, the production dynamic data of the test samples were predicted, and the production dynamic data (including daily water production and daily oil production) of one of the production wells were selected for display. Figure 8 and Fig. 9 The comparison between the model prediction results and the numerical simulation results for daily water production and daily oil production corresponding to production well 1 shows that the model prediction results are consistent with the numerical simulation results, especially the water breakthrough time and the turning point of oil production decline can be accurately predicted.

[0046] The output of the training-completed and tested deep graph and time series hybrid learning model with good performance uses the fracture distribution parameters of tight oil and gas reservoirs and the vertex connectivity strength corresponding to the calculated fracture distribution as model input to quickly predict the corresponding production dynamic data.

[0047] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for rapid prediction of production performance of tight oil and gas reservoirs with multi-scale fracture development, characterized in that: The steps include: Step 1: Obtain multi-scale fracture distribution data samples and production simulation data samples, and build a sample library; Step 2, establishing a crack network diagram to represent the data; Step 3, calculating the vertex connectivity strength matrix of the fracture network based on the graph representation data; Step 4: Establish a hybrid learning model of depth map and time series for rapid prediction of tight oil and gas reservoir production dynamics; Step 5: Train the depth map and time series hybrid learning model, and verify the performance of the model after the training is completed; Step 6: Rapidly predict the production dynamics of tight oil and gas reservoirs based on the trained model.

2. The method for rapid prediction of production performance of tight oil and gas reservoirs with multi-scale fracture development according to claim 1 is characterized in that: The specific process of step 1 is as follows: Step 1.1, stochastically model the multi-scale fracture distribution of tight oil and gas reservoirs, and establish a multi-scale fracture distribution data sample containing several groups of different fracture distributions; Step 1.2, using an embedded discrete fracture numerical simulation method to perform numerical simulation calculations on different fracture distributions, and obtain corresponding production simulation data samples, where the production simulation data samples are production dynamic data of production wells, including daily oil production, daily gas production, daily water production, and bottom hole pressure; Step 1.3: The fracture distribution data samples and the corresponding production simulation data samples are combined into a sample library, and the sample library is divided into a training set and a test set in proportion.

3. The method for rapid prediction of production performance of tight oil and gas reservoirs with multi-scale fracture development according to claim 2 is characterized in that: In step 2, a group of fractures are distributed as a fracture network, including multiple discrete fractures, and the parameters of each fracture include length, position, direction, aperture, porosity and permeability; a graphical representation of the fracture network is established based on the discrete fracture parameters, that is, , Represents the crack network diagram to represent the data, represents the set of vertices of the graph, Represents the edge set of the graph; the two endpoints of each single fracture are used as the vertices of the graph, and the position coordinates of the fracture endpoints are used as vertex features; the fractures are used as the edges of the graph, and the length, direction, aperture, porosity and permeability of the fractures are used as edge features; intersecting fractures are re-divided from the intersection.

4. The method for rapid prediction of production performance of tight oil and gas reservoirs with multi-scale fracture development according to claim 3 is characterized in that: In step 3, the connectivity strength of the fracture vertices is calculated based on the graph representation data of the fracture network, and a vertex connectivity strength matrix of the fracture network is constructed; the calculation formula for the connectivity strength between any two vertices in the matrix is: (1)； In the formula, Indicates The vertices and The connectivity strength between vertices corresponds to the first Line Elements of a column; Indicates The vertices and The connection distance between vertices; Indicates the total number of vertices.

5. The method for rapid prediction of production performance of tight oil and gas reservoirs with multi-scale fracture development according to claim 4, characterized in that: The specific process of step 4 is as follows: Step 4.1, establish a deep graph learning module for feature learning and extraction of graph representation data of the fracture network; Step 4.2: Establish a deep time series learning module for production dynamic data prediction of production wells; Step 4.3, combine the deep graph learning module and the deep time series learning module to obtain an end-to-end deep graph and time series hybrid learning model; the input data of the deep graph and time series hybrid learning model is the graph representation data of the fracture network and the vertex connectivity strength matrix of the fracture network, and the output is production dynamic prediction data.

6. The method for rapid prediction of production performance of tight oil and gas reservoirs with multi-scale fracture development according to claim 5, characterized in that: In step 4.1, the deep graph learning module includes an embedding layer, an information propagation aggregation model, an improved Transformer model, and a global pooling layer; the specific working process is as follows: Step 4.1.1: Represent the data in the fracture network diagram Input embedding layer, which uses two independent fully connected layers to perform feature mapping on vertex features and edge features of graph data respectively; Step 4.1.2: Use the information propagation aggregation model to fuse the edge features mapped by the embedding layer into the vertex features; Specifically: the information propagation aggregation model will The vertices and The edges between vertices The feature is propagated to the edge corresponding to Vertices and Vertices , use vector concatenation and merging operations to merge edge features into vertex feature vectors, and use a layer of fully connected neural network to perform a mapping calculation on the merged vectors at the vertex to obtain the updated vertex features; Step 4.1.3: Based on the improved Transformer model, perform multi-level update and feature transformation on the vertex features after integrating the edge features; The improved Transformer model embeds the connectivity strength matrix into the learning and prediction process of the original Transformer model, using the vertex connectivity strength matrix The calculation process of the crack network structure encoding information participating in an attention operation is as follows: (2)； In the formula, Indicates Updated vertex features output by the layer encoder; Indicates The dimension of the vertex features after the layer encoder updates; is the softmax function; , and Represents the query matrix, key matrix, and value matrix in the attention operation; For the The dimension of the vertex features after the layer encoder updates; is a learnable weight matrix; is the transpose symbol; Step 4.1.4: Use the global pooling layer to pool all vertex features into a feature vector as the output of the deep graph learning module. The feature vector is the high-level feature of the extracted crack network. .

7. The method for rapid prediction of production performance of tight oil and gas reservoirs with multi-scale fracture development according to claim 6, characterized in that: In step 4.2, the deep time series learning module uses the original Transformer model to perform production dynamic time series prediction; the specific working process is as follows: Step 4.2.1: High-level features of the fracture network Copy to time steps as input vectors, and added with the time sequence coding information as the input of the encoder; the time sequence coding information is obtained by encoding the time sequence of the production dynamic data; Step 4.2.2: The multi-head self-attention mechanism inside the encoder performs feature transformation on the time series; Step 4.2.3: Use the fully connected layer as the output layer to output the production dynamic prediction value corresponding to each time step , ,in, To forecast the number of indicators for production; Indicates the number of producing wells; Indicates the number of production dynamic indicator types; The number of forecast time steps representing the production dynamics.

8. The method for rapid prediction of production performance of tight oil and gas reservoirs with multi-scale fracture development according to claim 7, characterized in that: In step 5, the depth map and time series hybrid learning model is trained in combination with the training set divided in step 1, the model parameters are optimized to minimize the loss function, and the mean absolute error is used as the loss function of the model. The calculation formula is as follows: （3）； In the formula, is the mean absolute error; Indicates the number of samples; and Respectively represent The true value of the sample and the model prediction result; After training is completed, the performance of the model is verified using the test set divided in step 1. The verification indicator is the determination coefficient, and the formula is as follows: （4）； In the formula, is the coefficient of determination; express The average of the true values ​​of the samples.

9. The method for rapid prediction of production performance of tight oil and gas reservoirs with multi-scale fracture development according to claim 8, characterized in that: In step 6, the depth map and time series hybrid learning model that has been trained and tested with good performance is output, and the fracture distribution parameters of the tight oil and gas reservoir and the vertex connectivity strength matrix corresponding to the calculated fracture distribution are used as model input to quickly predict the corresponding production dynamic data.

Images