Prediction Method, Equipment and Medium for Dynamic Distribution of Physical Fields in Multi-Stage Fractured Horizontal Wells

Abstract

The present application discloses a method, device and medium for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well, which relates to the technical field of oil and gas reservoir development. The method includes: obtaining the reservoir geological parameters and multi-stage fracture information of a target oil and gas reservoir, using the reservoir geological parameters and multi-stage fracture information as inputs, using a first prediction model to determine the physical field map at a first prediction time step, using a second prediction model to determine the physical field map at a second prediction time step, using the physical field map at the first prediction time step and the physical field map at the second prediction time step as inputs, using a third prediction model to determine the physical field map at each prediction time step from the third prediction time step to the last prediction time step, and forming the prediction result of the dynamic distribution of the physical field of the multi-stage fractured horizontal well with the physical field maps at all prediction time steps. The present application can efficiently and accurately predict the dynamic distribution of the physical field of a multi-stage fractured horizontal well.

Description

Technical Field

[0001] The present application relates to the technical field of oil and gas reservoir development, and particularly to a method, device and medium for predicting the dynamic distribution of physical fields in a multi-stage fractured horizontal well. Background Art

[0002] To improve the development efficiency of oil and gas reservoirs, horizontal wells and multi-stage fracturing technologies are widely used, which can effectively increase oil and gas production. Under the condition of multi-stage fractured horizontal wells, the dynamic distribution of multiple physical fields in the oil and gas reservoir is complex and is affected by multiple factors such as reservoir characteristics, fracturing operations and production dynamics. Accurately predicting the dynamic distribution of these physical fields can reveal the hydrodynamic characteristics of the fluid in the reservoir, optimize the development plan, master the development dynamics, and provide scientific support for the efficient development and refined management of oil and gas.

[0003] Currently, the method for predicting the dynamic distribution of physical fields in a multi-stage fractured horizontal well of an oil and gas reservoir is mainly the numerical simulation method. Although the numerical simulation method has high prediction accuracy and strong flexibility and can simulate more complex scenarios, the accuracy of numerical simulation highly depends on the accuracy of input parameter information, and at the same time, it requires a large amount of computing resources and the support of professional software.

[0004] In contrast, the application of machine learning methods in the field of oil and gas reservoirs is still in its infancy. Currently, common machine learning methods mainly use historical physical fields and other influencing factors to predict the physical fields for a period of time in the future. This method makes up for the deficiency of the numerical simulation method relying on high-quality input parameter information to a certain extent and has obvious advantages in terms of computing efficiency. However, this method requires obtaining historical physical fields, and the computing efficiency still needs to be improved. Summary of the Invention

[0005] The purpose of the present application is to provide a method, device and medium for predicting the dynamic distribution of physical fields in a multi-stage fractured horizontal well, which can efficiently and accurately predict the dynamic distribution of physical fields in a multi-stage fractured horizontal well.

[0006] To achieve the above purpose, the present application provides the following solutions.

[0007] In a first aspect, the present application provides a method for predicting the dynamic distribution of physical fields in a multi-stage fractured horizontal well, where the method for predicting the dynamic distribution of physical fields in a multi-stage fractured horizontal well includes:

[0008] Obtain the reservoir geological parameters and multi-stage fracture information of the target oil and gas reservoir; the target oil and gas reservoir is developed by a multi-stage fractured horizontal well;

[0009] Taking the reservoir geological parameters and the multi-segment fracture information as inputs, using the first prediction model to determine the physical field map at the first prediction time step, and using the second prediction model to determine the physical field map at the second prediction time step; both the first prediction model and the second prediction model include a first feature extraction module, a second feature extraction module, and a prediction module. The first feature extraction module is used to extract the first features of the reservoir geological parameters, the second feature extraction module is used to extract the second features of the multi-segment fracture information, and the prediction module is used to predict the physical field map based on the first features and the second features;

[0010] Taking the physical field map at the first prediction time step and the physical field map at the second prediction time step as inputs, using the third prediction model to determine the physical field map at each prediction time step from the third prediction time step to the last prediction time step; the third prediction model adopts a model capable of processing time series data;

[0011] Combining the physical field maps at all prediction time steps to form the prediction result of the dynamic distribution of the physical field of the multi-stage fractured horizontal well.

[0012] In a second aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. The processor executes the computer program to implement the above-mentioned method for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well.

[0013] In a third aspect, the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned method for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well is implemented.

[0014] According to the specific embodiments provided by the present application, the present application has the following technical effects:

[0015] The present application provides a method, device, and medium for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well. The reservoir geological parameters and multi-stage fracture information of the target oil and gas reservoir are obtained. Using the reservoir geological parameters and multi-stage fracture information as inputs, the physical field map at the first prediction time step is determined using the first prediction model, and the physical field map at the second prediction time step is determined using the second prediction model. Using the physical field map at the first prediction time step and the physical field map at the second prediction time step as inputs, the physical field map at each prediction time step from the third prediction time step to the last prediction time step is determined using the third prediction model. The physical field maps at all prediction time steps are combined to form the prediction result of the dynamic distribution of the physical field of the multi-stage fractured horizontal well. By designing the first prediction model, the second prediction model, and the third prediction model, the present application can directly predict the prediction result of the dynamic distribution of the physical field of the multi-stage fractured horizontal well based on static data such as reservoir geological parameters and multi-stage fracture information, thus eliminating the need for historical physical fields. Since obtaining historical physical fields either through actual production or through numerical simulation requires time, the present application directly uses static data for prediction without the need for historical physical fields, which can improve the prediction efficiency. At the same time, by comprehensively using static data such as reservoir geological parameters and multi-stage fracture information for prediction, the prediction accuracy can be improved. Therefore, the dynamic distribution of the physical field of the multi-stage fractured horizontal well can be predicted efficiently and accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0017] Figure 1 It is an application environment diagram of a method for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well provided in Embodiment 1 of the present application.

[0018] Figure 2 It is a flowchart of a method for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well provided in Embodiment 1 of the present application.

[0019] Figure 3 It is a schematic diagram of the fracture design scheme of a single horizontal well provided in Embodiment 1 of the present application.

[0020] Figure 4 It is a schematic diagram of the visualization result of reservoir geological parameters when taking 100 horizontal wells as an example provided in Embodiment 1 of the present application.

[0021] Figure 5 It is a schematic diagram of the fracture image of a single horizontal well provided in Embodiment 1 of the present application.

[0022] Figure 6 Schematic diagram of the dynamic distribution of gas saturation field in a single horizontal well provided in Embodiment 1 of the present application.

[0023] Figure 7 Schematic diagram of the dynamic distribution of fracture pressure field in a single horizontal well provided in Embodiment 1 of the present application.

[0024] Figure 8 Schematic diagram of the dynamic distribution of matrix pressure field in a single horizontal well provided in Embodiment 1 of the present application.

[0025] Figure 9 Schematic diagram of the overall architecture of the spatio-temporal prediction model framework based on multi-source feature fusion using deep learning provided in Embodiment 1 of the present application.

[0026] Figure 10 Schematic diagram of the structure of the pre-trained model with multi-source feature fusion provided in Embodiment 1 of the present application.

[0027] Figure 11 Schematic diagram of the structure of ResNet (Residual Network) integrated with SAM (Spatial Attention Mechanism) provided in Embodiment 1 of the present application.

[0028] Figure 12 Schematic diagram of the structure of the ConvLSTM (Convolutional Long Short-Term Memory) model that fuses multi-scales provided in Embodiment 1 of the present application.

[0029] Figure 13 Schematic diagram of the structure of the ConvLSTM unit provided in Embodiment 1 of the present application.

[0030] Figure 14 Schematic diagram of the prediction result of the dynamic distribution of gas saturation field in a single horizontal well provided in Embodiment 1 of the present application.

[0031] Figure 15 Schematic diagram of the prediction result of the dynamic distribution of fracture pressure field in a single horizontal well provided in Embodiment 1 of the present application.

[0032] Figure 16 Schematic diagram of the prediction result of the dynamic distribution of matrix pressure field in a single horizontal well provided in Embodiment 1 of the present application. Detailed implementation manners

[0033] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0034] Embodiment 1.

[0035] The method for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well provided by the embodiment of the present application can be applied to an application environment as Figure 1 shown. Among them, the terminal communicates with the server through the network. The data storage system can store the data that the server needs to process. The data storage system can be set separately, integrated on the server, or placed on the cloud or other servers. The terminal can send a request for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well to be processed to the server. After receiving the request for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well to be processed, for the request for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well to be processed, the server obtains the reservoir geological parameters and multi-stage fracture information of the target oil and gas reservoir; using the reservoir geological parameters and multi-stage fracture information as inputs, uses the first prediction model to determine the physical field map at the first prediction time step, and uses the second prediction model to determine the physical field map at the second prediction time step; using the physical field map at the first prediction time step and the physical field map at the second prediction time step as inputs, uses the third prediction model to determine the physical field map at each prediction time step from the third prediction time step to the last prediction time step; and composes the physical field maps at all prediction time steps into the prediction result of the dynamic distribution of the physical field of the multi-stage fractured horizontal well. The server can feedback the prediction result of the dynamic distribution of the physical field of the multi-stage fractured horizontal well for the request for predicting the dynamic distribution of the physical field of the multi-stage fractured horizontal well to the terminal.

[0036] In addition, in some embodiments, the method for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well can also be implemented by the server or the terminal alone. For example, the terminal can directly process the request for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well to be processed, or the server can obtain the request for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well to be processed from the data storage system and process the request for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well to be processed.

[0037] In an exemplary embodiment, as Figure 2 shown, a method for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well is provided. This method is executed by a computer device, and can be specifically executed alone by a computer device such as a terminal or a server, or jointly executed by a terminal and a server. In the embodiment of the present application, this method is applied to Figure 1Taking the server in [description] as an example, the following steps are included.

[0038] Step S1: Obtain the reservoir geological parameters and multi - stage fracture information of the target oil and gas reservoir; the target oil and gas reservoir is developed by using multi - stage fractured horizontal wells.

[0039] Step S2: Using the reservoir geological parameters and the multi - stage fracture information as inputs, determine the physical field map at the first prediction time step by using the first prediction model, and determine the physical field map at the second prediction time step by using the second prediction model; both the first prediction model and the second prediction model include a first feature extraction module, a second feature extraction module, and a prediction module. The first feature extraction module is used to extract the first features of the reservoir geological parameters, the second feature extraction module is used to extract the second features of the multi - stage fracture information, and the prediction module is used to predict the physical field map based on the first features and the second features.

[0040] Step S3: Using the physical field map at the first prediction time step and the physical field map at the second prediction time step as inputs, determine the physical field map at each prediction time step from the third prediction time step to the last prediction time step by using the third prediction model; the third prediction model adopts a model capable of processing time - series data.

[0041] Step S4: Combine the physical field maps at all prediction time steps to form the prediction result of the dynamic distribution of the physical field of the multi - stage fractured horizontal well.

[0042] Implementing the above steps S1 to S4, in this embodiment, the static data such as reservoir geological parameters and multi - stage fracture information are directly used to predict the dynamic distribution of the physical field of the multi - stage fractured horizontal well in the oil and gas reservoir. The dynamic distribution of the physical field includes the physical field maps at multiple prediction time steps, without the need for historical physical field data, which can improve the prediction accuracy and efficiency, and is applicable to oil and gas reservoirs that have not been actually developed, and can evaluate whether it is necessary to develop the oil and gas reservoirs that have not been actually developed.

[0043] With the intensification of the contradiction between the supply and demand of conventional oil and gas resources, coalbed methane has become the focus of unconventional energy development due to its advantages such as rich reserves, wide distribution, cleanliness and environmental protection. However, coalbed methane is mainly stored in the micropores, fractures and matrix of coal seams, and its unique occurrence mode and low permeability bring development difficulties. To improve the development efficiency, horizontal wells and multi-stage fracturing technologies are widely used. Horizontal wells increase the contact area with the coal seam, while multi-stage fracturing significantly improves the permeability of the coal seam, effectively increasing the production of coalbed methane. Under the condition of multi-stage fracturing horizontal wells, the dynamic distribution of multiple physical fields (such as gas saturation field, fracture pressure field and matrix pressure field) in the coalbed methane reservoir is complex, affected by multiple factors such as reservoir characteristics, fracturing operations and production dynamics. Accurately predicting the dynamic distribution of these physical fields can reveal the hydrodynamic characteristics in the reservoir, optimize the development plan, master the development dynamics, and provide scientific support for the efficient development and refined management of coalbed methane.

[0044] At present, there are two common methods to predict the dynamic distribution of physical fields in coalbed methane reservoirs, which are as follows.

[0045] (1) Numerical simulation method: The numerical simulation method constructs a mathematical model of multiple physical fields in the coalbed methane reservoir to simulate the dynamic processes such as gas migration, fracture propagation and matrix interaction in the reservoir, and reveals its spatio-temporal distribution law. Its core principles include describing the migration behavior of coalbed methane in fractures and matrix based on Darcy's law and diffusion theory, using a dual-porosity model to reflect the interaction between microscopic pores and macroscopic fractures, and capturing the dynamic changes of pressure field, seepage field and stress field through a multi-field coupling model. The technical means mainly involve mathematical modeling, numerical discretization and computational solution: constructing a seepage-stress coupling equation through fracture-matrix coupling modeling, discretizing partial differential equations using the finite difference or finite volume method, and iteratively solving them in combination with multi-scale algorithms to output dynamic results such as pressure field and gas saturation field. The specific implementation steps include first collecting geological and production data, clarifying boundary conditions and initial conditions, secondly performing grid division and discretization processing, inputting parameters to complete numerical solution, and finally verifying and optimizing the model in combination with measured data, analyzing the results to reveal reservoir dynamic characteristics and support the optimization of development plans.

[0046] (2) Machine learning method: The machine learning method realizes the efficient prediction of dynamic distribution characteristics by learning the complex mapping relationship of multi-physical fields in coalbed methane reservoirs from historical data. Its core principle is to use data-driven modeling, combine deep learning methods and ensemble learning methods to capture the non-linear relationship between reservoir characteristics and development conditions, and focus on constructing mapping models that describe the static characteristics of geological parameters and fracture parameters, as well as the dynamic characteristics reflecting the changes in time-series pressure and production pressure difference. The technical means mainly include four links: data preprocessing, feature extraction, model training, and evaluation and optimization. In the data preprocessing stage, outlier cleaning and numerical normalization are carried out to ensure data consistency. In the feature extraction stage, the core information affecting the dynamic distribution of multi-physical fields is mined through screening key parameters and dimensionality reduction optimization. In the model training stage, long short-term memory networks are used to process time-series features to construct a dynamic prediction model. In the evaluation and optimization stage, the model performance is tested through cross-validation, and the prediction accuracy is improved by adjusting hyperparameters. The specific implementation steps include data collection and collation, feature construction and screening, model establishment and training, prediction and analysis of physical field distribution results, and evaluation and improvement of model performance. Finally, the machine learning method can accurately predict the dynamic distribution characteristics of physical fields such as gas saturation field and pressure field, provide a scientific basis for optimizing the development plan of coalbed methane reservoirs, and effectively support the refined management of reservoirs.

[0047] However, due to its dependence on physical models, the numerical simulation method may be restricted by assumptions and simplifications, parameter uncertainties, and high computational costs, and has limited ability to capture the dynamic changes of physical fields. Although the machine learning method has high efficiency in dealing with non-linear relationships, it has a strong dependence on high-quality data, and there is relatively little applied research in the field of multi-stage fractured horizontal wells in coalbed methane reservoirs at present, making it difficult to directly use static data to accurately predict the dynamic changes of physical fields in coal reservoirs.

[0048] At the same time, other types of oil and gas reservoirs also have similar problems. That is to say, the current methods for predicting the dynamic distribution of physical fields in multi-stage fractured horizontal wells in oil and gas reservoirs are mainly numerical simulation methods. Although the numerical simulation method has high prediction accuracy and strong flexibility and can simulate more complex scenarios, the accuracy of the numerical simulation method highly depends on the accuracy of input parameter information, and at the same time, it requires a large amount of computing resources and the support of professional software. In contrast, the application of machine learning methods in the field of oil and gas reservoirs is still in its infancy. Currently, common machine learning methods mainly use historical physical fields and other influencing factors to predict the physical fields for a period of time in the future. This method to a certain extent makes up for the deficiency of the numerical simulation method's dependence on high-quality parameters and has obvious advantages in terms of computational efficiency. However, the research and application of directly using reservoir geological parameters and multi-stage fracture information to predict the dynamic distribution of physical fields in multi-stage fractured horizontal wells in oil and gas reservoirs are almost blank.

[0049] This embodiment proposes a deep learning method for predicting the dynamic distribution of the physical fields in a multi-stage fractured horizontal well in an oil and gas reservoir. For a coalbed methane reservoir, it is to predict the distribution of the gas saturation field, fracture pressure field, and matrix pressure field within the reservoir over time, aiming to get rid of the dependence on numerical simulation methods, fill the current research gap, and fully consider the superposition effect between multiple fractures on the physical field distribution in addition to reservoir geological parameters, so as to achieve a rapid and accurate prediction of the dynamic distribution of the physical fields in a multi-stage fractured horizontal well in an oil and gas reservoir within the latest 10 years (which can also be replaced by other prediction time periods), and present the prediction results in the form of a visualized field map, providing strong support for the refined management and efficient development of the oil and gas reservoir.

[0050] In step S1, when the target oil and gas reservoir is a coalbed methane reservoir, the reservoir geological parameters include: reservoir porosity, reservoir permeability, coal seam thickness, Young's modulus, gas-water relative permeability, irreducible water saturation, relative permeability curve parameters, reservoir pressure, ratio of desorption pressure to reservoir pressure, Langmuir volume, Langmuir pressure, and fracture permeability. The multi-fracture information includes: the proportion of the well-controlled area in the horizontal well, and the number, spacing, and length of the fracture reconstruction areas in the well-controlled area. The fracture reconstruction area includes several fractures. The physical field maps include: gas saturation field map, fracture pressure field map, and matrix pressure field map.

[0051] When the target oil and gas reservoir is a shale gas reservoir, the reservoir geological parameters include: reservoir porosity, reservoir permeability, shale thickness, organic matter content, thermal maturity, Young's modulus, Poisson's ratio, Langmuir volume, Langmuir pressure, fracture permeability, and reservoir pressure. The multi-fracture information includes: the proportion of the well-controlled area in the horizontal well, and the number, spacing, and length of the fracture reconstruction areas in the well-controlled area. The fracture reconstruction area includes several fractures. The physical field maps include: gas saturation field map, fracture pressure field map, and matrix pressure field map.

[0052] When the target oil and gas reservoir is a tight oil reservoir, the reservoir geological parameters include: reservoir porosity, reservoir permeability, reservoir thickness, mineral composition, Young's modulus, Poisson's ratio, crude oil viscosity, relative permeability curve parameters, saturation pressure, capillary pressure, and fracture permeability. The multi-fracture information includes: the proportion of the well-controlled area in the horizontal well, and the number, spacing, and length of the fracture reconstruction areas in the well-controlled area. The fracture reconstruction area includes several fractures. The physical field maps include: oil saturation field map and dissolved gas / free gas distribution field map.

[0053] In step S2, using the reservoir geological parameters and multi-segment fracture information as inputs, the physical field map at the first prediction time step is determined using the first prediction model, and the physical field map at the second prediction time step is determined using the second prediction model. Specifically, it includes: performing data normalization processing on the reservoir geological parameters to obtain the normalized parameters; converting the multi-segment fracture information into a fracture image, where the multi-segment fracture information is marked in the fracture image; using the normalized parameters and the fracture image as inputs to determine the physical field map at the first prediction time step using the first prediction model; using the normalized parameters and the fracture image as inputs to determine the physical field map at the second prediction time step using the second prediction model.

[0054] Among them, converting the multi-segment fracture information into a fracture image specifically includes: constructing a blank image, the size and shape of the blank image are determined by scaling down the size and shape of the horizontal well proportionally, drawing the well control area in the blank image based on the proportion of the well control area to the horizontal well, the shape of the well control area can be a rectangle, and the position can be determined based on the actual situation, drawing multiple rectangles in the well control area of the blank image based on the number, spacing, and length of the fracture transformation areas in the well control area, each rectangle represents a fracture transformation area, and the width of the rectangle can be determined based on the actual situation.

[0055] In step S2, both the first prediction model and the second prediction model include a first feature extraction module, a second feature extraction module, and a prediction module. The output ends of the first feature extraction module and the second feature extraction module are both connected to the input end of the prediction module. The first feature extraction module is used to extract the first features of the reservoir geological parameters, the second feature extraction module is used to extract the second features of the multi-segment fracture information, and the prediction module is used to predict the physical field map based on the first features and the second features.

[0056] Among them, the first feature extraction module uses a fully connected network, and the fully connected network includes a first fully connected layer, a first ReLU activation function layer, a second fully connected layer, a second ReLU activation function layer, and a third fully connected layer connected in sequence.

[0057] Among them, the second feature extraction module uses a convolutional neural network, and the convolutional neural network includes a second convolutional layer, a third ReLU activation function layer, a third convolutional layer, a fourth ReLU activation function layer, and a fourth convolutional layer connected in sequence.

[0058] Among them, the prediction module includes a channel splicing layer, a residual network integrating a spatial attention mechanism, and a first convolutional layer connected in sequence. The output ends of the fully connected network and the convolutional neural network are both connected to the input end of the channel splicing layer. Specifically, the output ends of the third fully connected layer and the fourth convolutional layer are both connected to the input end of the channel splicing layer. The residual network integrating the spatial attention mechanism includes a number of residual blocks connected in sequence. The residual block includes a fifth convolutional layer, a first batch normalization layer, a sixth convolutional layer, a second batch normalization layer, a fifth ReLU activation function layer, a spatial attention mechanism layer, and a sixth ReLU activation function layer. The input end of the fifth convolutional layer is also connected to the input end of the sixth ReLU activation function layer.

[0059] In step S3, the third prediction model can adopt any model capable of processing time series data. In this embodiment, the third prediction model adopts a convolutional long short-term memory network integrating multi-scales.

[0060] Among them, the convolutional long short-term memory network integrating multi-scales includes a plurality of convolutional blocks connected in sequence. The number of convolutional blocks is the same as the number of prediction time steps. The convolutional block includes three convolutional units connected in sequence. The convolutional kernel sizes of the three convolutional units are different, and the convolutional kernel sizes of the three convolutional units are 9×9, 7×7, and 3×3 respectively.

[0061] It should be noted that the number of the first prediction model, the second prediction model, and the third prediction model is the same as the number of physical fields. For the prediction of each physical field, a set of the first prediction model, the second prediction model, and the third prediction model is required. For example, for a coalbed methane reservoir, three sets of the first prediction model, the second prediction model, and the third prediction model are required.

[0062] To overcome the limitations of existing research methods, this embodiment proposes a deep learning method based on static data-driven for predicting the dynamic distribution of physical fields of multi-stage fractured horizontal wells in oil and gas reservoirs. This method takes static data such as reservoir geological parameters and multi-stage fracture information as input, constructs a pre-training model integrating multi-source features (i.e., the first prediction model and the second prediction model) and a ConvLSTM model integrating multi-scales (i.e., the third prediction model), and can efficiently and accurately predict the dynamic distribution of physical fields of multi-stage fractured horizontal wells in oil and gas reservoirs without relying on historical physical field distribution data. Experiments are carried out on the prediction effect of the coalbed methane reservoir. The experimental results show that this method can accurately predict the distributions of the gas saturation field, the fracture pressure field, and the matrix pressure field in the latest 10 years, and the determination coefficients R 2 are 0.947, 0.925, and 0.923 respectively. When predicting the distributions of the gas saturation field, the fracture pressure field, and the matrix pressure field in the latest 5 years, R 2Further improved to 0.953, 0.944, and 0.946, fully demonstrating that the method proposed in this embodiment has strong generalization ability and good applicability. The innovation of this embodiment lies in that it not only gets rid of the dependence on traditional numerical simulation methods, greatly saving calculation time, but also fully considers the influence of reservoir geological parameters and the superposition effect of multi-stage fractures on the physical field distribution of coal reservoirs, improving the prediction accuracy. By directly using static data, it can quickly predict the dynamic distribution of the physical field inside coal reservoirs, providing more convenient technical support for efficient development.

[0063] Next, taking the coalbed methane reservoir as an example, the training processes of the first prediction model, the second prediction model, and the third prediction model of this embodiment are introduced in detail as follows.

[0064] This embodiment proposes a deep learning method driven by static data, mainly including a pre-training model for multi-source feature fusion and a ConvLSTM model for fusing multi-scales, which are used to predict the dynamic distribution of the physical field of multi-stage fractured horizontal wells in coalbed methane reservoirs. Among them, the pre-training model for multi-source feature fusion is constructed based on a fully connected network (FCN), a convolutional neural network (CNN), a residual network, and a spatial attention mechanism, which can effectively map static reservoir geological parameters and multi-stage fracture information into a form similar to a physical field map. The ConvLSTM model for fusing multi-scales uses the output results of the pre-training model to achieve accurate prediction of the physical field map in subsequent prediction time steps, mainly including four steps.

[0065] Step 1: Obtain reservoir geological parameters, multi-stage fracture information, and coal reservoir physical field distribution data.

[0066] (1) Collect reservoir geological parameters.

[0067] Reservoir geological parameters include reservoir porosity, reservoir permeability, coal seam thickness, Young's modulus, gas-water relative permeability (including the maximum gas-phase relative permeability and the maximum water-phase relative permeability), irreducible water saturation, relative permeability curve parameters, reservoir pressure, the ratio of desorption pressure to reservoir pressure, Langmuir volume, Langmuir pressure, and fracture permeability, etc. The value ranges of specific parameters are shown in Table 1. According to the reservoir geological parameters of different horizontal wells, a reservoir geological parameter plan is formulated, and the reservoir geological parameter plan includes the values of each parameter in the reservoir geological parameters.

[0068] Table 1 Detailed information of reservoir geological parameters

[0069]

[0070] (2) Design multi-stage fracture information.

[0071] In this embodiment, the Latin hypercube sampling method is used to design the positions of multiple fractures in a horizontal well, forming a fracture design scheme. The fracture design scheme includes the proportion of the well control area in the horizontal well, the number of fracture treatment zones, the spacing between fracture treatment zones, and the length of the fracture treatment zone. The detailed design scope includes: the number of fracture treatment zones is between 5 and 15, the spacing between adjacent fracture treatment zones is between 60 m and 100 m, and the length of the fracture treatment zone does not exceed the width of the well control area (representing its width along the vertical direction). The fracture design scheme for a single horizontal well is as Figure 3 shown, Figure 3 and the fracture area in

[0072] is the fracture treatment zone.

[0073] Set the number of grids along the horizontal direction of the multi-stage fractured horizontal well to 256, and the number of grids along the vertical direction to 128. The distance of a single grid is 5 m both in the horizontal and vertical directions. The well control area represents its length along the horizontal direction and its width along the vertical direction. The fracture treatment zone represents the spacing between fracture treatment zones along the horizontal direction and the length of the fracture treatment zone along the vertical direction.

[0074] (4) Extract the physical field distribution data of the coal reservoir.

[0075] By combining the designed reservoir geological parameter scheme and the fracture design scheme, use numerical simulation software to simulate the coalbed methane reservoir. First, write the information of the reservoir geological parameter scheme and the fracture design scheme into the coalbed methane reservoir simulation file in a predetermined format. Subsequently, call the numerical simulator to batch execute the coalbed methane reservoir simulation process, and extract the required physical field distribution data of the coal reservoir from the simulation results, including the gas saturation field distribution data, the fracture pressure field distribution data, and the matrix pressure field distribution data.

[0076] Step 2, data preprocessing, preprocess the reservoir geological parameter scheme, the fracture design scheme, and the physical field distribution data of the coal reservoir respectively.

[0077] For the reservoir geological parameter scheme, perform data normalization processing on the reservoir geological parameter scheme so that the values of all data are between 0 and 1, as Figure 4 shown, Figure 4 where A - O in Figure 4 is the schematic diagram of 15 reservoir geological parameters, and parameters 1 - 15 in

[0078] For the fracture design scheme, in this embodiment, it is found that since the number of fracture transformation zones for each horizontal well is not exactly the same, there is an inconsistency in the data length of multi-segment fracture information. To facilitate subsequent model processing, in combination with the actual meaning of multi-segment fracture information, including the proportion of the well-controlled area in the horizontal well, the number of fracture transformation zones, the distance between fracture transformation zones, and the length of fracture transformation zones, the multi-segment fracture information is converted into an image format to obtain a fracture image, as Figure 5 shown.

[0079] For the coal reservoir physical field distribution data, since the production cycle of multi-segment fractured horizontal wells is set to 20 years and the time step is monthly in the numerical simulation process, a total of 240 time steps of coal reservoir physical field maps are generated, including gas saturation field maps, fracture pressure field maps, and matrix pressure field maps. However, since the physical field of the coalbed methane reservoir changes little in a short time, to reduce the data storage pressure without losing key information, downsampling is performed on all the extracted physical field maps. Specifically, one time step is retained every 12 time steps, and finally the coal reservoir physical field distribution data is reduced to 20 key time steps. This strategy not only effectively preserves the key physical field map data during production but also significantly reduces the cache burden, while improving the data processing efficiency, providing guarantee for the efficient development of subsequent analysis and model training. The visualization results of the gas saturation field distribution data, fracture pressure field distribution data, and matrix pressure field distribution data of a single horizontal well are as Figure 6 , Figure 7 and Figure 8 shown.

[0080] Step 3: Construct a spatio-temporal prediction model framework based on multi-source feature fusion of deep learning.

[0081] In order to efficiently fuse reservoir geological parameters and multi-segment fracture information, accurately capture the complex relationships of spatio-temporal features, and thus achieve high-precision prediction of the dynamic distribution of the coal reservoir physical field, this embodiment constructs a spatio-temporal prediction model framework based on multi-source feature fusion of deep learning. The overall architecture schematic diagram of this model framework is as Figure 9 shown.

[0082] The spatio-temporal prediction model framework based on multi-source feature fusion in deep learning consists of a pre-trained model for multi-source feature fusion and a ConvLSTM model that fuses multi-scales. Its design purpose is to solve the problem that traditional methods cannot directly use reservoir geological parameters and multi-stage fracture information to predict the dynamic physical field map of multi-stage fractured horizontal wells in coalbed methane reservoirs. For this reason, in this embodiment, a pre-trained model for multi-source feature fusion is used to process reservoir geological parameters and multi-stage fracture information, and map them into the form of physical field maps of the first two time steps. Subsequently, through the ConvLSTM model that fuses multi-scales, the dynamic prediction of subsequent physical field maps is realized, so as to effectively capture the spatio-temporal evolution characteristics of multi-stage fractured horizontal wells in coalbed methane reservoirs.

[0083] As for why the physical field map of each time step is not gradually predicted through the pre-trained model of multi-source feature fusion, the reason is that although the pre-trained model of multi-source feature fusion can generate the physical field map of a single time step, gradually predicting the physical field map of each time step will consume a large amount of computing resources and time, and at the same time, it fails to make full use of the time correlation between physical field maps. In contrast, the reason for choosing to pre-train the physical field maps of two time steps instead of one time step is that the physical field map of a single time step can only reflect the physical field distribution at a specific moment, and it is difficult to reveal the dynamic change trend and time series characteristics. The physical field maps of two time steps can capture the spatio-temporal laws of the dynamic evolution of the physical field more comprehensively. In addition, through the ConvLSTM model that fuses multi-scales, the spatio-temporal evolution characteristics of the physical field maps of the first two time steps can be efficiently captured. Using its time memory mechanism and spatial convolution operation, the ConvLSTM model that fuses multi-scales can effectively reduce the computational complexity and realize the efficient prediction of the physical field maps of subsequent time steps, thus significantly improving the prediction efficiency and prediction accuracy.

[0084] The main method used in this embodiment is implemented based on PyTorch in Python. The specific process of model construction is as follows.

[0085] (1)Construction and training of the pre-trained model for multi-source feature fusion.

[0086] The pre-trained model for multi-source feature fusion is constructed by combining a fully connected network, a convolutional neural network, a residual network, and a spatial attention mechanism. The FCN is designed to extract the global features of reservoir geological parameters, the CNN is designed to extract the local features of multi-stage fracture information. The ResNet can not only efficiently extract the deep correlation features between reservoir geological parameters and multi-stage fracture information by introducing a skip connection mechanism, but also effectively alleviate the problem of gradient disappearance. The SAM can dynamically adjust the attention area according to the spatial distribution of data, so as to accurately capture the local correlation and global distribution characteristics of the spatial position and geometric shape of fractures. The structural schematic diagram of the pre-trained model for multi-source feature fusion is as Figure 10 shown.

[0087] The implementation process of the pre-trained model with multi-source feature fusion is as follows: First, a fully connected layer is used to perform non-linear mapping on one-dimensional static reservoir geological parameters, converting them into a shape consistent with the target field Figure 1 to obtain global features; at the same time, two-dimensional static multi-segment fracture information is used with multiple layers of convolution to extract high-level spatial features to obtain local features. Subsequently, the extracted global features and local features are concatenated in the channel dimension to form a fusion field map. The fusion field map is optimized through the integrated SAM ResNet (whose structural schematic diagram is as Figure 11 shown) to enhance the network's attention to key features and effectively capture the complex interaction relationships between multi-dimensional features. Finally, a single-channel physical field map prediction result is generated through the convolutional layer. Through two pre-trained models with multi-source feature fusion, the physical field map prediction results for the first two time steps are finally obtained.

[0088] ResNet solves the problems of gradient disappearance and model degradation in the training of deep networks by introducing residual connections. Figure 11 The residual module in it reflects this feature: it adds a "skip connection" to the main path, directly adding the input to the output. This structure can make information flow more smoothly in the network and reduce information loss caused by increasing depth. The calculation formula is as follows:

[0089] ;

[0090] Among them, is the output; represents the residual part, usually a combination of several operations such as convolution and activation, is the input, is the weight of the residual part.

[0091] SAM is a technology used to enhance the model's ability to focus on spatial features. It generates a weight distribution map by analyzing the spatial distribution pattern of the input feature map, enabling the model to pay more attention to key regions or important features. In Figure 11 it, SAM uses convolution operations to process the original feature map. The generated spatial attention weight map is multiplied element-wise with the original feature map, so that the features in the key fracture well control area are amplified while the rest of the area is suppressed. This mechanism can improve the model's performance in the prediction of complex dynamic physical field maps, especially the ability to capture local features and key regions. The calculation formula is as follows:

[0092] ;

[0093] ;

[0094] Among them, is the spatial attention weight map; is the Sigmoid activation function that normalizes the weights to the range [0, 1]; is a two-dimensional convolution operation used to generate the spatial attention weight map; and are the global average pooling and max pooling operations on the input feature map respectively, and the results are concatenated along the channel dimension. is the input feature map; is the output feature map enhanced by spatial attention; denotes element-wise multiplication.

[0095] The training process of the pre-trained model for multi-source feature fusion is as follows.

[0096] (1.1) Data division: The dataset is divided into a training set and a test set in the ratio of 8:2. The dataset includes the first samples and the first labels. The first samples are the reservoir geological parameter schemes and the fracture design schemes, and the first labels are the physical field maps at the first time step and the physical field maps at the second time step of the coal reservoir physical field distribution data.

[0097] (1.2) Model initialization: The Adam (Adaptive Moment Estimation) optimizer is used, and the learning rate is set to 0.0005. At the same time, the mean squared error (MSE) is used as the loss function to minimize the difference between the predicted value and the true value.

[0098] (1.3) Data loading: The training set data is divided into a training set and a validation set, where 10% of the training set data is used as the validation set, and a data loader is set to load the data in batches of size 4.

[0099] (1.4) Model training process: The total number of epochs is set to 75. For each batch of data, the model performs forward propagation to calculate the predicted value, then calculates the error through the loss function, and updates the model weights using backpropagation. After each epoch, the loss value of the validation set is calculated and compared with the previously recorded optimal validation loss (i.e., the minimum value of the loss value of the validation set).

[0100] (1.5) Early stopping mechanism: To prevent overfitting, training is terminated early when there is no improvement in the loss value of the validation set for 10 consecutive epochs.

[0101] (1.6) Model saving: When the loss value of the validation set reaches the lowest value, the parameter file of the model is saved for subsequent model prediction and optimization.

[0102] (2)Construction and training of the multi-scale integrated ConvLSTM model.

[0103] The multi-scale integrated ConvLSTM model extracts multi-scale features more comprehensively through multi-layer stacking and convolutional operations with different receptive fields, capturing the complex spatial distribution law of the reservoir dynamic physical field. In addition, the joint modeling of multi-scale features and time series dynamics helps to more precisely simulate the fracture pressure propagation and reservoir change law. Overall, the multi-scale integrated ConvLSTM model has significant advantages in capturing the features of complex dynamic physical fields, enhancing feature expression, and dealing with spatio-temporal interaction laws. The multi-scale integrated ConvLSTM model uses the physical field maps of two time steps obtained from the pre-trained model as inputs to predict the physical field maps of subsequent multiple time steps, thus achieving accurate and rapid prediction of the reservoir dynamic physical field maps. The schematic diagram of the structure of the multi-scale integrated ConvLSTM model is as Figure 12 shown. It contains 3 layers of ConvLSTM units (3 layers of ConvLSTM units are convolutional blocks, and 1 layer of ConvLSTM unit is a convolutional unit). It extracts large, medium, and small scale spatio-temporal features layer by layer through convolutional kernels of different sizes (9×9, 7×7, 3×3). Each layer of ConvLSTM unit combines the convolutional operation and the time modeling ability of LSTM to capture rich multi-scale information, fully meeting the requirements of complex spatial distribution and dynamic changes in the prediction of reservoir dynamic physical field maps.

[0104] The core of the multi-scale integrated ConvLSTM model is the ConvLSTM unit, and its basic structure is as Figure 13 shown. Among them, the data of each time step is processed by convolution, and the gating mechanism is used to control the flow and storage of information, and finally generates a high-quality spatio-temporal feature representation, laying a foundation for the subsequent prediction of the physical field map.

[0105] The specific calculation process of the ConvLSTM unit is shown in the following formula. From top to bottom, they are the calculation of the forget gate, input gate, candidate memory unit, memory unit update, output gate, and hidden state:

[0106] ;

[0107] ;

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] Among them, , , , , , , are, respectively, the forget gate, input gate, candidate memory cell, memory cell, output gate, hidden state, and input data of the ConvLSTM at time t; is the Sigmoid activation function, with a range between 0 and 1; is the hidden state of the ConvLSTM at time t - 1; is the memory cell of the ConvLSTM at time t - 1; , , , are, respectively, the convolutional kernel weights of the forget gate, input gate, candidate memory cell, and output gate; , , , are, respectively, the bias terms of the forget gate, input gate, candidate memory cell, and output gate; tanh is the hyperbolic tangent function, with a range between -1 and 1; represents the convolution operation; represents the matrix dot product; represents the matrix addition.

[0113] The training process of the multi-scale fused ConvLSTM model is as follows.

[0114] (2.1) Data partitioning: The dataset is partitioned into a training set and a test set in a ratio of 8:2. The dataset includes the second sample and the second label. The second sample is the prediction data obtained from the pre-trained model, that is, the physical field maps of the first two time steps, and the second label is the physical field map of the subsequent time step.

[0115] (2.2)Data Preparation: Data preparation is a crucial step in the entire training process. By calling the create_data_loader function, the training sample data X_train, training label data y_train, test sample data X_test, and test label data y_test are encapsulated into data loaders train_loader and test_loader. The batch size batch_size = 4 is set for each loader, and the data shuffling function is enabled. At this step, the input data is preprocessed. The unsqueeze(1) function is used to expand the 4D data into 5D data to conform to the input format required by the ConvLSTM model, i.e., [batch_index (batch index), channel_index (channel index), time_index (time index), height_index (height index), width_index (width index)]. The data loader can load data in batches and provide random shuffling during training to increase the diversity of training.

[0116] (2.3)Model Initialization: After data preparation is completed, the next step is model initialization. First, an instance of the ConvLSTM model is created. The input dimension input_dim = 2 is set, the hidden layer dimensions are [64, 64, 8], the sizes of the convolutional kernels are [(9, 9), (7, 7), (3, 3)], and a three-layer network structure is set. The mean squared error loss (MSELoss) is used as the loss function, which is suitable for regression tasks. To optimize the model parameters, the Adam optimizer is adopted, and the learning rate is set to 0.0005. With such a configuration, the necessary computational graph and optimizer are prepared for the training process so that the model can update the weights through gradient descent and learn the patterns in the data.

[0117] (2.4)Model Training: Model training is the core of the whole process. Set the total number of epochs to 25. In each epoch, first check if there is an available GPU through the torch.device function. If the GPU is available, both the model and the loss function will be transferred to the GPU to accelerate the calculation. At the beginning of each epoch, set the model to training mode model.train(), and initialize running_loss to accumulate the loss of each epoch. Next, train with the data in the train_loader. For each mini-batch, the data will be transferred to the GPU and passed through the model for forward propagation to output a series of predicted values for each time step. Then, calculate the mean squared error loss between the predicted output of the last time step and the true target value, and execute the backward propagation loss.backward() function to calculate the gradients. Then, use the optimizer.step() function to update the model parameters. After processing each mini-batch, the loss value will be accumulated, and at the end of each epoch, calculate the average loss value of that epoch, and print out the current training status and timestamp.

[0118] (2.5)Clear Cache: During the training process, since the model and data are calculated on the GPU, it may occupy a large amount of video memory. To prevent video memory overflow, after each epoch during the training process, call the torch.cuda.empty_cache() function to clear the unused GPU memory. This can not only effectively release the video memory but also help avoid errors caused by insufficient memory during training, ensuring that the model can smoothly perform multiple rounds of training. GPU memory management is crucial in deep learning training, especially when training large models. Reasonable memory cleaning can ensure the stability and efficiency of training.

[0119] (2.6)Model Saving: After all training cycles are completed, the parameters of the model are saved at the specified path. Through the torch.save(model.state_dict(), "file name") function, save the trained model weights in a.pth file, so that these weights can be loaded in subsequent inferences or continued training. Model saving is a routine operation in deep learning projects, which can prevent losses caused by interrupted training in the middle and also facilitate the use of the model in future practical applications.

[0120] Step Four, Model Optimization and Prediction.

[0121] The model optimization process uses random search and Bayesian optimization methods, combined with an early stopping strategy, to tune the hyperparameters of the multi-scale fused ConvLSTM model. First, the random search method is used to preliminarily explore the reasonable range of hyperparameters and quickly obtain an approximately optimal region. Then, based on the results of random search, the probability model in the Bayesian optimization method is used to further finely tune the key hyperparameters. Finally, the performance of the model is evaluated based on the loss of the validation set, and the early stopping strategy is used to avoid overfitting. In the above optimization process, the validation set is divided from the training set, and the ratio of the validation set to the training set is 1:4.

[0122] Random search is an optimization method that randomly samples parameter combinations within a preset hyperparameter range and evaluates their performance. Bayesian optimization models the distribution of the objective function based on a probability model (such as a Gaussian process), and efficiently locates the optimal hyperparameter combination by balancing "exploration" (trying new parameters) and "exploitation" (optimizing existing good parameters). The early stopping strategy monitors the performance change of the validation set and stops training early when the validation loss fails to decrease significantly within 10 training epochs set, to avoid overfitting problems.

[0123] Combining random search, Bayesian optimization, and the early stopping strategy to tune the hyperparameters of the multi-scale fused ConvLSTM model can give full play to the efficiency of random search in quickly exploring the hyperparameter space, leverage the ability of Bayesian optimization to finely tune key parameters, and at the same time avoid overfitting and improve the tuning efficiency through the early stopping strategy. This method has higher efficiency and adaptability compared to single optimization methods in deep learning tasks with high complexity and high computational costs.

[0124] The hyperparameters to be optimized in this embodiment include the learning rate, training batch size, number of hidden units, and convolutional kernel size. The set hyperparameter ranges are shown in Table 2.

[0125] Table 2 Hyperparameter Setting Ranges

[0126]

[0127] In the prediction and evaluation process of the model, first, the trained model is used to predict the training set and the test set. Then, a custom evaluation function is used to calculate and return the model evaluation metrics of the training set and the test set, including the correlation coefficient (R 2 )), mean squared error (abbreviated as MSE), root mean squared error (Root Mean Squared Error, RMSE), and mean absolute error (Mean Absolute Error, MAE). The calculation formulas are as follows. Finally, the evaluation results of the training set and the test set are output, and some predicted images are displayed to help further analyze the model performance.

[0128] ;

[0129] ;

[0130] ;

[0131] ;

[0132] Among them, is the physical field map at the th predicted time step; is the physical field map at the th actual time step in the multi-stage fractured horizontal well; is the mean value of the actual physical field maps for all time steps; is the number of time steps of the physical field map in the coalbed methane horizontal well. Among these evaluation metrics, the closer the value of R 2 is to 1, the smaller the values of RMSE, MSE, and MAE, and the better the prediction performance of the model.

[0133] In this embodiment, a dataset consisting of 1000 multi-stage fractured horizontal wells is used to verify the above method.

[0134] In step one, 1000 sets of reservoir geological parameters and multi-stage fracture information are collected, and a reservoir geological parameter scheme and a fracture design scheme are respectively formed. Then, the numerical simulation software MATLAB and CMG are used to simulate the 20-year production process of the multi-stage fractured horizontal well. Finally, the coal reservoir physical field distribution data including gas saturation field data, fracture pressure field data, and matrix pressure field data are extracted from it.

[0135] In step two, the reservoir geological parameter scheme, the fracture design scheme, and the coal reservoir physical field distribution data are respectively preprocessed, and finally a high-quality dataset containing 1000 horizontal wells is constructed, with rich data structure and multi-dimensional characteristics. Specifically, this dataset includes: 15 one-dimensional static reservoir geological parameters, image data of two-dimensional multi-stage fracture distributions, and multi-physical field dynamic data corresponding to 20 time steps over a 20-year time span. These multi-physical field dynamic data cover the gas saturation field, the fracture pressure field, and the matrix pressure field, providing comprehensive and multi-level support for subsequent research and analysis.

[0136] In step three, a dataset containing 1000 horizontal wells is randomly divided into a training set and a test set in a ratio of 8:2. During the random division process, it is ensured that the static reservoir geological parameters and multi-segment fracture information of each group of samples can correspond one-to-one with the multi-physical-field dynamic data. Subsequently, the training set data is input into the pre-trained model with multi-source feature fusion. Through the training and prediction of the model, the gas saturation field maps, fracture pressure field maps, and matrix pressure field maps for the first 2 time steps are generated in sequence. During this process, due to the prediction characteristics of different field maps, 3 independent pre-trained models are respectively trained, which are used to predict the gas saturation field map, fracture pressure field map, and matrix pressure field map. Then, taking the field maps generated by the pre-trained model as input, a ConvLSTM model integrating multi-scales is further used to predict the physical field maps for subsequent multiple time steps. Similarly, due to the different prediction requirements for different physical field maps, 3 independent ConvLSTM models are finally trained to predict the gas saturation field map, fracture pressure field map, and matrix pressure field map respectively. This method of hierarchical training and multi-model collaboration provides higher accuracy and flexibility for the time-series prediction of complex physical fields.

[0137] In step four, the hyperparameter optimization results of the ConvLSTM model integrating multi-scales are shown in Table 3.

[0138] Table 3 Hyperparameter Optimization Results

[0139]

[0140] The prediction results of the pre-trained model with multi-source feature fusion are shown in Table 4.

[0141] Table 4 Prediction Results of the Pre-trained Model with Multi-source Feature Fusion

[0142]

[0143] The prediction results of the ConvLSTM model integrating multi-scales are shown in Tables 5, 6, 7, Figure 14 , Figure 15 and Figure 16 as shown.

[0144] Table 5 Gas Saturation Field Prediction Results

[0145]

[0146] Table 6 Fracture Pressure Field Prediction Results

[0147]

[0148] Table 7 Matrix Pressure Field Prediction Results

[0149]

[0150] The above experiments prove that the method proposed in this embodiment performs well in predicting the multi-physical field dynamic distribution of multi-stage fractured horizontal wells in coalbed methane reservoirs by using static reservoir geological parameters and multi-stage fracture information. It not only has high accuracy but also faster prediction speed. The above experiments are not a limitation to this embodiment. This embodiment is not limited to the above samples and can also be applied to other problems with similar static input and dynamic time-series output patterns.

[0151] Aiming at the research gap in the prediction of the dynamic physical field map of coalbed methane reservoirs, this embodiment proposes a time-series prediction model framework based on deep learning multi-source feature fusion. By constructing a multi-source fusion pre-training model and a ConvLSTM model that fuses multi-scales, it realizes the efficient fusion of static reservoir geological parameters and multi-stage fracture information and the rapid and accurate prediction of the dynamic physical field map of coalbed methane reservoirs. By introducing residual networks, spatial attention mechanisms, and ConvLSTM, the model deeply explores the complex correlations between multi-source data, accurately depicts the characteristics of reservoir changes, gets rid of the dependence on numerical simulation methods, and provides a new path for directly predicting the dynamic physical field map based on static data. This method not only promotes the theoretical innovation of the dynamic behavior modeling of coalbed methane reservoirs but also has the potential to be extended to the prediction of unconventional resources such as shale gas and tight oil and gas reservoirs. In the future, with the optimization of algorithms and the expansion of applications, this method is expected to play an important role in the prediction and decision support of complex dynamic systems.

[0152] Existing methods for predicting the multi-physical field dynamic distribution of coalbed methane reservoirs, such as traditional numerical simulation methods and some machine learning methods, all have obvious deficiencies. Traditional numerical simulation methods highly rely on geological modeling, require a large amount of computing resources and professional software support, have poor applicability, mostly focus on the initial production stage, and only provide regular curves without visual field maps. Most machine learning methods are for predicting production or single characteristic curves and generally rely on historical data. The method for predicting the multi-physical field dynamic distribution of multi-stage fractured horizontal wells in coalbed methane reservoirs proposed in this embodiment has many significant advantages. In terms of data utilization, it gets rid of the dependence on historical data and can make predictions only relying on static data such as reservoir geological parameters and multi-stage fracture information, greatly broadening the application scenarios, especially being able to work effectively in the case of scarce historical data. In terms of result presentation, it shows the prediction results with intuitive visual field maps, which is more helpful for researchers to understand the internal dynamic changes of the reservoir compared with traditional methods, thus being able to more efficiently provide strong support for refined gas reservoir management and development decisions, significantly improving the efficiency and quality of coalbed methane reservoir development research, and reducing the potential risks and costs caused by inaccurate predictions.

[0153] The present application also provides an application scenario, which applies the above-mentioned multi-stage fracturing horizontal well physical field dynamic distribution prediction method. Specifically, the multi-stage fracturing horizontal well physical field dynamic distribution prediction method provided in this embodiment can be applied in an oil and gas reservoir development scenario. The oil and gas reservoir development scenario includes a prediction stage and a development stage. The prediction stage is used to predict the physical field distribution, and the development stage is used to develop the oil and gas reservoir based on the physical field distribution. The multi-stage fracturing horizontal well physical field dynamic distribution prediction method provided in this embodiment belongs to the prediction stage.

[0154] Embodiment 2.

[0155] In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the multi-stage fracturing horizontal well physical field dynamic distribution prediction method in Embodiment 1 is implemented.

[0156] Embodiment 3.

[0157] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the multi-stage fracturing horizontal well physical field dynamic distribution prediction method in Embodiment 1 is implemented.

[0158] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0159] Specific examples are used in this article to elaborate on the principles and implementation manners of the present application. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A method for predicting the dynamic distribution of physical fields in multi-stage fractured horizontal wells, characterized in that, The method for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well includes: Obtaining the reservoir geological parameters and multi-stage fracture information of the target oil and gas reservoir; the target oil and gas reservoir is developed by a multi-stage fractured horizontal well; Using the reservoir geological parameters and the multi-stage fracture information as inputs, determining the physical field map at the first prediction time step by using the first prediction model, and determining the physical field map at the second prediction time step by using the second prediction model; both the first prediction model and the second prediction model include a first feature extraction module, a second feature extraction module, and a prediction module. The first feature extraction module is used to extract the first features of the reservoir geological parameters, the second feature extraction module is used to extract the second features of the multi-stage fracture information, and the prediction module is used to predict the physical field map based on the first features and the second features; Using the physical field map at the first prediction time step and the physical field map at the second prediction time step as inputs, determining the physical field map at each prediction time step from the third prediction time step to the last prediction time step by using the third prediction model; the third prediction model uses a model capable of processing time series data; Combining the physical field maps at all prediction time steps to form the prediction result of the dynamic distribution of the physical field of the multi-stage fractured horizontal well.

2. The method for predicting the dynamic distribution of physical fields in a multi-stage fractured horizontal well according to claim 1, wherein When the target oil and gas reservoir is a coalbed methane reservoir, the reservoir geological parameters include: reservoir porosity, reservoir permeability, coal seam thickness, Young's modulus, gas-water relative permeability, irreducible water saturation, relative permeability curve parameters, reservoir pressure, ratio of desorption pressure to reservoir pressure, Langmuir volume, Langmuir pressure, and fracture permeability. The multi-stage fracture information includes: the proportion of the well-controlled area in the horizontal well and the number, spacing, and length of the fracture transformation areas in the well-controlled area. The fracture transformation area includes several fractures. The physical field map includes: gas saturation field map, fracture pressure field map, and matrix pressure field map; When the target oil and gas reservoir is a shale gas reservoir, the reservoir geological parameters include: reservoir porosity, reservoir permeability, shale thickness, organic matter content, thermal maturity, Young's modulus, Poisson's ratio, Langmuir volume, Langmuir pressure, fracture permeability, and reservoir pressure. The multi-stage fracture information includes: the proportion of the well-controlled area in the horizontal well and the number, spacing, and length of the fracture transformation areas in the well-controlled area. The fracture transformation area includes several fractures. The physical field map includes: gas saturation field map, fracture pressure field map, and matrix pressure field map; When the target oil and gas reservoir is a tight oil reservoir, the reservoir geological parameters include: reservoir porosity, reservoir permeability, reservoir thickness, mineral composition, Young's modulus, Poisson's ratio, crude oil viscosity, relative permeability curve parameters, saturation pressure, capillary pressure, and fracture permeability. The multi-stage fracture information includes: the proportion of the well-controlled area in the horizontal well and the number, spacing, and length of the fracture transformation areas in the well-controlled area. The fracture transformation area includes several fractures. The physical field map includes: oil saturation field map and dissolved gas / free gas distribution field map.

3. The physical field dynamic distribution prediction method for multi-stage fractured horizontal wells according to claim 1, wherein Using the reservoir geological parameters and the multi-stage fracture information as inputs, determining the physical field map at the first prediction time step by using the first prediction model, and determining the physical field map at the second prediction time step by using the second prediction model, specifically including: Perform data normalization processing on the reservoir geological parameters to obtain the normalized parameters; Convert the multi-segment fracture information into a fracture image; the multi-segment fracture information is marked in the fracture image; Using the normalized parameters and the fracture image as inputs, determine the physical field map at the first prediction time step using the first prediction model; Using the normalized parameters and the fracture image as inputs, determine the physical field map at the second prediction time step using the second prediction model.

4. The prediction method for the dynamic distribution of the physical field of a multi-stage fractured horizontal well according to claim 1, wherein The first feature extraction module uses a fully connected network, the second feature extraction module uses a convolutional neural network, the prediction module includes a channel splicing layer, a residual network integrating a spatial attention mechanism, and a first convolutional layer connected in sequence. The output ends of the fully connected network and the convolutional neural network are both connected to the input end of the channel splicing layer.

5. The prediction method for the dynamic distribution of the physical field of a multi-stage fractured horizontal well according to claim 4, wherein The fully connected network includes a first fully connected layer, a first ReLU activation function layer, a second fully connected layer, a second ReLU activation function layer, and a third fully connected layer connected in sequence; The convolutional neural network includes a second convolutional layer, a third ReLU activation function layer, a third convolutional layer, a fourth ReLU activation function layer, and a fourth convolutional layer connected in sequence; The output ends of the third fully connected layer and the fourth convolutional layer are both connected to the input end of the channel splicing layer; The residual network integrating the spatial attention mechanism includes a plurality of residual blocks connected in sequence. The residual block includes a fifth convolutional layer, a first batch normalization layer, a sixth convolutional layer, a second batch normalization layer, a fifth ReLU activation function layer, a spatial attention mechanism layer, and a sixth ReLU activation function layer. The input end of the fifth convolutional layer is also connected to the input end of the sixth ReLU activation function layer.

6. The method for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well according to claim 1, wherein The third prediction model uses a convolutional long short-term memory network that fuses multiple scales.

7. The method for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well according to claim 6, wherein The convolutional long short-term memory network that fuses multiple scales includes a plurality of convolutional blocks connected in sequence. The number of convolutional blocks is the same as the number of prediction time steps. The convolutional block includes three convolutional units connected in sequence, and the convolutional kernel sizes of the three convolutional units are different.

8. The method for predicting the dynamic distribution of the physical field of a multi-stage fractured horizontal well according to claim 7, characterized in that The convolutional kernel sizes of the three convolutional units are 9×9, 7×7, and 3×3 respectively.

9. A computer device, comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-segment fractured horizontal well physical field dynamic distribution prediction method according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-segment fractured horizontal well physical field dynamic distribution prediction method according to any one of claims 1-8.

Images