Reservoir production prediction method, device and product based on adaptive graph convolutional network

Abstract

An embodiment of the present application provides a reservoir production prediction method, device, and product based on an adaptive graph convolutional network. The method includes: First, obtain the first dynamic data of injection wells and the second dynamic data of production wells. Subsequently, input these dynamic data into a pre-trained prediction model together. The feature extraction network part of the model is built with an adjacency matrix that can characterize the spatial influence relationship between injection wells and production wells, so as to obtain a predicted oil production that takes into account both the influence of injection wells and the factors of production wells. This method not only solves the problem of limited prediction accuracy in existing reservoir production prediction methods due to only focusing on the own attributes of production wells, but also significantly improves the accuracy of reservoir production prediction by comprehensively considering the dynamic data of injection wells and production wells and their interaction relationships.

Description

Technical Field

[0001] This application relates to the field of oil extraction, and particularly to a method, device, and product for predicting reservoir oil production based on an adaptive graph convolutional network. Background Art

[0002] In the field of oil extraction, water flooding of reservoirs is a commonly used method to increase the production of oil wells. This method injects water or other displacement fluids into the reservoir, and uses the hydraulic pressure to push the crude oil towards the production wells, thereby increasing the production of oil wells.

[0003] However, during the implementation of water flooding of reservoirs, in order to more effectively manage resources and optimize the extraction strategy, it is necessary to predict the reservoir oil production.

[0004] Existing methods for predicting reservoir oil production, although they can provide references for production decisions to a certain extent, often only consider the attributes of the oil production wells themselves, which have certain limitations and result in low accuracy of production prediction. Summary of the Invention

[0005] Embodiments of this application provide a method, device, and product for predicting reservoir oil production based on an adaptive graph convolutional network, so as to solve the technical problem that existing reservoir oil production prediction only considers the attributes of the oil production wells themselves, resulting in a decrease in the accuracy of production prediction.

[0006] In a first aspect, embodiments of this application provide a method for predicting reservoir oil production based on an adaptive graph convolutional network, including:

[0007] Obtain the first dynamic data of the water injection wells and the second dynamic data of the oil production wells, where the first dynamic data includes the water injection rate and / or bottom hole flowing pressure of the water injection wells, and the second dynamic data includes at least one of the oil production rate, water production rate, liquid production rate, and bottom hole flowing pressure of the oil production wells;

[0008] Input the first dynamic data and the second dynamic data into a pre-trained prediction model to obtain the predicted oil production output by the prediction model, where the feature extraction network in the prediction model includes an adaptive graph convolutional network, and the adjacency matrix in the adaptive graph convolutional network is used to represent the spatial influence relationship between the water injection wells and the oil production wells.

[0009] In a possible implementation manner, the prediction model includes an encoding network, the feature extraction network, and a decoding network, where the feature extraction network includes multiple cascaded convolutional layers, and each convolutional layer includes a time series processing module and the adaptive graph convolutional network;

[0010] Inputting the first dynamic data and the second dynamic data into a pre-trained prediction model to obtain the predicted oil production output by the prediction model includes:

[0011] Input the first dynamic data and the second dynamic data into the encoding network to obtain initial features, and use the initial features as the input vectors of the first convolutional layer in the multiple cascaded convolutional layers. For any convolutional layer, pass the input vector of the convolutional layer through the temporal processing module of the convolutional layer to generate first features, and then pass the first features through the adaptive graph convolutional network of the convolutional layer to generate second features, and use the second features as the input vectors of the next convolutional layer until the first features generated by each of the multiple cascaded convolutional layers are obtained;

[0012] Input the first features generated by each of the multiple cascaded convolutional layers into the decoding network to obtain the predicted oil production.

[0013] In a possible implementation, the temporal processing module includes a first temporal convolutional network and a second temporal convolutional network. The step of passing the input vector of the convolutional layer through the temporal processing module of the convolutional layer to generate first features, and then passing the first features through the adaptive graph convolutional network of the convolutional layer to generate second features includes:

[0014] Input the input vector of the convolutional layer into the first temporal convolutional network to obtain first sub-features;

[0015] Input the input vector of the convolutional layer into the second temporal convolutional network to obtain second sub-features;

[0016] Multiply the first sub-features and the second sub-features to obtain the first features;

[0017] Generate the second features based on the first features and the adjacency matrix in the adaptive graph convolutional network.

[0018] In a possible implementation, the step of inputting the first features generated by each of the multiple cascaded convolutional layers into the decoding network to obtain the predicted oil production includes:

[0019] Accumulate the first features generated by each of the multiple cascaded convolutional layers to obtain a fused feature;

[0020] Decode the fused feature through one or more decoding layers of the decoding network to obtain the predicted oil production.

[0021] In a possible implementation manner, the adjacency matrix is obtained by training the prediction model based on an initial adjacency matrix, and the initial adjacency matrix is determined based on the first static data of the water injection well and the second static data of the oil production well. The first static data includes the permeability, reservoir thickness, and position coordinates of the reservoir where the water injection well is located, and the second static data includes the permeability, reservoir thickness, and position coordinates of the reservoir where the oil production well is located.

[0022] In a second aspect, an embodiment of the present application provides a method for training an oil production prediction model, including:

[0023] Obtain sample data, where the sample data includes the first dynamic sample data of the water injection well, the second dynamic sample data of the oil production well, and an oil production label. Among them, the first dynamic sample data includes the water injection rate and / or bottom hole flowing pressure of the water injection well, and the second dynamic sample data includes at least one of the oil production rate, water production rate, liquid production rate, and bottom hole flowing pressure of the oil production well;

[0024] Input the first dynamic sample data and the second dynamic sample data into the oil production prediction model to obtain the predicted oil production output by the oil production prediction model. Among them, the feature extraction network in the oil production prediction model includes an adaptive graph convolutional network, and the initial adjacency matrix in the adaptive graph convolutional network is used to characterize the spatial influence relationship between the water injection well and the oil production well;

[0025] Adjust the parameters of the oil production prediction model and the initial adjacency matrix based on the predicted oil production and the oil production label.

[0026] In a possible implementation manner, it further includes:

[0027] Obtain the first static sample data of the water injection well and the second static sample data of the oil production well. The first static sample data includes the permeability, reservoir thickness, and position coordinates of the reservoir where the water injection well is located, and the second static sample data includes the permeability, reservoir thickness, and position coordinates of the reservoir where the oil production well is located;

[0028] Determine the initial adjacency matrix according to the first static sample data and the second static sample data.

[0029] In a third aspect, an embodiment of the present application provides a reservoir production prediction device based on an adaptive graph convolutional network, including:

[0030] An acquisition module for acquiring first dynamic data of an injection well and second dynamic data of a production well, where the first dynamic data includes the injection rate and / or bottom-hole flowing pressure of the injection well, and the second dynamic data includes at least one of the oil production rate, water production rate, liquid production rate, and bottom-hole flowing pressure of the production well;

[0031] An input module for inputting the first dynamic data and the second dynamic data into a pre-trained prediction model to obtain a predicted oil production output by the prediction model, where the feature extraction network in the prediction model includes an adaptive graph convolutional network, and the adjacency matrix in the adaptive graph convolutional network is used to characterize the spatial influence relationship between the injection well and the production well.

[0032] In a possible implementation manner, the input module is further configured to input the first dynamic data and the second dynamic data into the encoding network to obtain initial features, use the initial features as the input vector of the first convolutional layer in the multiple cascaded convolutional layers, and for any convolutional layer, generate first features by passing the input vector of the convolutional layer through the temporal processing module of the convolutional layer, and then generate second features by passing the first features through the adaptive graph convolutional network of the convolutional layer, and use the second features as the input vector of the next convolutional layer until the first features generated by each of the multiple cascaded convolutional layers are obtained;

[0033] The input module is specifically configured to input the first features generated by each of the multiple cascaded convolutional layers into the decoding network to obtain the predicted oil production.

[0034] In a possible implementation manner, the input module is further configured to input the input vector of the convolutional layer into the first temporal convolutional network to obtain first sub-features;

[0035] The input module is further configured to input the input vector of the convolutional layer into the second temporal convolutional network to obtain second sub-features;

[0036] The device further includes: a product module;

[0037] The product module is configured to multiply the first sub-features and the second sub-features to obtain the first features;

[0038] The device further includes: a generation module;

[0039] The generation module is configured to generate the second features based on the first features and the adjacency matrix in the adaptive graph convolutional network.

[0040] In a possible implementation manner, the device further includes: an accumulation module;

[0041] The accumulation module is configured to accumulate the first features generated by the multiple cascaded convolutional layers respectively to obtain a fused feature;

[0042] The apparatus further includes: a decoding module;

[0043] The decoding module is configured to decode the fused feature through one or more decoding layers of the decoding network to obtain the predicted oil production.

[0044] In a fourth aspect, an embodiment of the present application provides a training apparatus for an oil production prediction model, including:

[0045] An acquisition module, configured to acquire sample data, where the sample data includes first dynamic sample data of an injection well, second dynamic sample data of a production well, and an oil production label, where the first dynamic sample data includes the injection rate and / or bottom hole flowing pressure of the injection well, and the second dynamic sample data includes at least one of the oil production rate, water production rate, liquid production rate, and bottom hole flowing pressure of the production well;

[0046] An input module, configured to input the first dynamic sample data and the second dynamic sample data into the oil production prediction model to obtain the predicted oil production output by the oil production prediction model, where the feature extraction network in the oil production prediction model includes an adaptive graph convolutional network, and an initial adjacency matrix in the adaptive graph convolutional network is used to represent the spatial influence relationship between the injection well and the production well;

[0047] An adjustment module, configured to adjust the parameters of the oil production prediction model and the initial adjacency matrix based on the predicted oil production and the oil production label.

[0048] In a possible implementation manner, the acquisition module is further configured to acquire first static sample data of the injection well and second static sample data of the production well, where the first static sample data includes the permeability, reservoir thickness, and location coordinates of the reservoir where the injection well is located, and the second static sample data includes the permeability, reservoir thickness, and location coordinates of the reservoir where the production well is located;

[0049] The apparatus further includes: a determination module;

[0050] The determination module is configured to determine the initial adjacency matrix according to the first static sample data and the second static sample data.

[0051] In a fifth aspect, an embodiment of the present application provides a reservoir production prediction device based on an adaptive graph convolutional network, including: a memory, a processor;

[0052] The memory stores computer-executable instructions;

[0053] The processor executes the computer-executable instructions stored in the memory, so that the processor executes the first aspect and / or various possible implementation manners of the first aspect as described above.

[0054] In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed by a processor, they are used to implement the first aspect and / or various possible implementation manners of the first aspect as described above.

[0055] In a seventh aspect, an embodiment of the present application provides a computer program product, including a computer program, and when the computer program is executed by a processor, it implements the first aspect and / or various possible implementation manners of the first aspect as described above.

[0056] The reservoir production prediction method, device and product based on the adaptive graph convolutional network provided by the embodiments of the present application first obtain the first dynamic data of injection wells and the second dynamic data of production wells. Subsequently, these dynamic data are jointly input into a pre-trained prediction model, and the feature extraction network part of the model is built-in with an adjacency matrix that can represent the spatial influence relationship between injection wells and production wells, so as to obtain a predicted oil production that takes into account both the influence of injection wells and the factors of production wells. This method not only solves the problem of limited prediction accuracy in the existing reservoir production prediction methods due to only focusing on the own attributes of production wells, but also significantly improves the accuracy of reservoir production prediction by comprehensively considering the dynamic data of injection wells and production wells and their interaction relationship. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The drawings here are incorporated into the specification and form a part of this specification, showing embodiments consistent with the present application, and are used together with the specification to explain the principles of the present application.

[0058] Figure 1 Flow schematic of the reservoir production prediction method based on the adaptive graph convolutional network provided by the present application Figure 1 ;

[0059] Figure 2 Flow schematic of the reservoir production prediction method based on the adaptive graph convolutional network provided by the present application Figure 2 ;

[0060] Figure 3 Flow schematic of the reservoir production prediction method based on the adaptive graph convolutional network provided by the present application Figure 3 ;

[0061] Figure 4 Flow schematic of the training method of the oil production prediction model provided by the present applicationFigure 4 ;

[0062] Figure 5 Schematic flow chart of the training method for the oil production prediction model provided by this application Figure 5 ;

[0063] Figure 6 Schematic diagram of the prediction structure of the reservoir production prediction method based on the adaptive graph convolution network provided by this application;

[0064] Figure 7 Schematic diagram of the structure of the reservoir production prediction device based on the adaptive graph convolution network provided by this application;

[0065] Figure 8 Schematic diagram of the structure of the training device for the oil production prediction model provided by this application;

[0066] Figure 9 Schematic diagram of the structure of the reservoir production prediction device based on the adaptive graph convolution network provided by this application.

[0067] Through the above-mentioned drawings, specific embodiments of this application have been shown, and there will be more detailed descriptions hereinafter. These drawings and text descriptions are not intended to limit the scope of the concept of this application in any way, but to illustrate the concept of this application to those skilled in the art by referring to specific embodiments. Detailed Description of the Embodiments

[0068] Here, the exemplary embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. On the contrary, they are merely examples of devices and methods consistent with some aspects of this application as detailed in the appended claims.

[0069] First, the nouns involved in this application are explained.

[0070] (1) The Adaptive Graph Convolution Network (AGCN) is a neural network model specifically used to process graph-structured data. Its core lies in being able to perform convolution operations adaptively according to the topological structure and node features of the graph, thereby effectively extracting the feature information in the graph data.

[0071] (2) Temporal Convolutional Network (TCN) is a network architecture that transforms the structure of a convolutional neural network to handle sequence modeling tasks. The role of the Temporal Convolutional Network is to extract the temporal features of time series data. The characteristic of the Temporal Convolutional Network is that it ensures that the prediction at each time point depends only on the current and previous time points through causal convolution. Dilated convolution is used to expand the receptive field and capture features with long time spans without increasing computational complexity.

[0072] With the continuous progress of oil extraction technology, water flooding in oil reservoirs has become a widely used method to increase oil well production. The core of this method lies in precisely injecting water or other displacement fluids into the oil reservoir, skillfully using the liquid pressure as a driving force to effectively drive the crude oil towards the production well, thereby significantly increasing the oil well production.

[0073] However, in the actual operation of water flooding in oil reservoirs, in order to achieve refined management of oil resources and optimization of extraction strategies, oil reservoir production prediction is required.

[0074] Existing oil reservoir production prediction methods, although they can provide references for production decisions to a certain extent, often only consider the attributes of the production wells themselves, have certain limitations, and result in low accuracy of production prediction.

[0075] To address the above problems, the oil reservoir production prediction method based on an adaptive graph convolutional network provided by this application first obtains the first dynamic data of injection wells and the second dynamic data of production wells. Subsequently, these dynamic data are jointly input into a pre-trained prediction model. The feature extraction network part of this model is built with an adjacency matrix that can represent the spatial influence relationship between injection wells and production wells, so as to obtain a predicted oil production that takes into account both the influence of injection wells and the factors of production wells. This method not only solves the problem of limited prediction accuracy caused by only focusing on the attributes of production wells themselves in existing oil reservoir production prediction methods, but also significantly improves the accuracy of oil reservoir production prediction by comprehensively considering the dynamic data of injection wells and production wells and their interaction relationships.

[0076] The following uses specific embodiments to elaborate in detail on the technical solutions of this application and how the technical solutions of this application solve the above technical problems. These specific embodiments below can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of this application will be described below in conjunction with the accompanying drawings.

[0077] Figure 1 Flow schematic of the oil reservoir production prediction method based on an adaptive graph convolutional network provided by this application Figure 1 , asFigure 1 As shown, the method includes:

[0078] S101. Obtain the first dynamic data of the water injection well and the second dynamic data of the oil production well. Among them, the first dynamic data includes the water injection rate and / or the bottom-hole flowing pressure of the water injection well, and the second dynamic data includes at least one of the oil production rate, water production rate, liquid production rate, and bottom-hole flowing pressure of the oil production well.

[0079] Among them, the water injection rate of the water injection well refers to the amount of water injected into the formation by the water injection well per unit time. For example, the water injection rate of water injection well A is 100 cubic meters of water injected per day, which means that this well will inject 100 cubic meters of water into the formation within 24 hours.

[0080] The bottom-hole flowing pressure of the water injection well refers to the fluid pressure at the bottom of the water injection well. For example, the bottom-hole flowing pressure of water injection well A is 20 MPa, which means that the fluid injected by the water injection well exerts a pressure of 20 MPa on the formation.

[0081] The bottom-hole flowing pressure of the water injection well is used to characterize the influence of the water injection fluid of the water injection well on the formation pressure. The greater the bottom-hole flowing pressure of the water injection well, the greater the pressure exerted by the water injection fluid on the formation. Conversely, the smaller the bottom-hole flowing pressure of the water injection well, the smaller the pressure exerted by the water injection fluid on the formation.

[0082] The oil production rate of the oil production well refers to the amount of oil extracted from the formation by the oil production well per unit time. For example, the oil production rate of oil production well A is 5 cubic meters of oil produced per day, which means that this well will extract 5 cubic meters of oil from the formation within 24 hours.

[0083] The water production rate of the oil production well refers to the amount of water extracted from the formation by the oil production well per unit time. For example, the water production rate of oil production well A is 10 cubic meters of water produced per day, which means that this well will extract 10 cubic meters of water from the formation within 24 hours.

[0084] The liquid production rate of the oil production well refers to the total volume of the liquid (including crude oil and water) produced from the formation by the oil production well per unit time. For example, the liquid production rate of oil production well A is 7 cubic meters produced per day, which means that this well will extract 7 cubic meters of liquid from the formation within 24 hours.

[0085] The bottom-hole flowing pressure of the oil production well refers to the fluid pressure at the bottom of the oil production well. For example, the bottom-hole flowing pressure of oil production well A is 15 MPa, which means that the fluid at the bottom of the oil production well is under a pressure of 15 MPa.

[0086] The purpose of obtaining the first dynamic data of the water injection well and the second dynamic data of the oil production well in this step is to obtain the overall acquisition situation of the oil production well and the water injection well over a past period of time.

[0087] It is understandable that during the exploitation of a water - drive reservoir, injection wells and production wells work in coordination. By continuously injecting water into the formation, injection wells effectively increase the reservoir pressure, forming a water - pressure driving force that causes crude oil to flow towards the production wells under pressure and ultimately be produced to the surface.

[0088] The first dynamic data of the injection well directly reflects the strength of the injection well's water - injection capacity, while the second dynamic data of the production well reveals the production capacity of the production well with the assistance of the injection well.

[0089] Therefore, by obtaining the first dynamic data of the injection well and the second dynamic data of the production well, it is possible to more accurately evaluate the overall production and collection situation of the production well and the injection well over a certain period of time.

[0090] This step can be, for example, to obtain the first dynamic data of the injection well and the second dynamic data of the production well over a certain period of time. For example, assuming the past 7 days as the time range, the system will collect the water - injection rate, bottom - hole flowing pressure of the injection well, and the oil - production rate, water - production rate, liquid - production rate, and bottom - hole flowing pressure of the production well within the past 7 days.

[0091] The method of obtaining the first dynamic data of the injection well and the second dynamic data of the production well in this step can be, for example, obtained from the data management library, or can also be obtained from the data management platform of the data recorders. This application does not make special restrictions on this.

[0092] S102: Input the first dynamic data and the second dynamic data into a pre - trained prediction model to obtain the predicted oil production output by the prediction model. Among them, the feature extraction network in the prediction model includes an adaptive graph convolutional network, and the adjacency matrix in the adaptive graph convolutional network is used to represent the spatial influence relationship between the injection well and the production well.

[0093] Among them, the purpose of this step is to obtain a predicted oil production that takes into account both the influence of the injection well and the factors of the production well.

[0094] It is understandable that since the feature extraction network part of the prediction model uses an adaptive graph convolutional network that can represent the spatial influence relationship between the injection well and the production well, and the core of the adaptive graph convolutional network is that it can effectively spread and update the feature information on the nodes based on the adjacency matrix. Through a unique point - edge adjacent convolution operation, this network can deeply fuse the correlation information between the first dynamic data of the injection well and the second dynamic data of the production well, thereby capturing the complex interaction between the two.

[0095] Therefore, when the system inputs the first dynamic data of the injection well and the second dynamic data of the production well into this pre-trained prediction model, the model can make full use of the spatial influence relationship between the injection well and the production well, as well as the water injection capacity and production capacity information contained in the dynamic data, to predict future oil production.

[0096] For example, assume that there are the respective dynamic data of 10 production wells and 5 injection wells in the past 10 days. Then, based on the above information, the system first inputs the dynamic data of these 10 days into the prediction model together. Subsequently, the prediction model can use its built-in adjacency matrix for characterizing the spatial influence relationship between the injection well and the production well to comprehensively consider and analyze the dynamic data. Finally, the prediction model outputs the predicted oil production for the 11th day.

[0097] Optionally, the present application provides a possible implementation process of the adjacency matrix, including: the adjacency matrix is obtained by training the prediction model based on an initial adjacency matrix, and the initial adjacency matrix is determined based on the first static data of the injection well and the second static data of the production well. The first static data includes the permeability, reservoir thickness, and location coordinates of the reservoir where the injection well is located, and the second static data includes the permeability, reservoir thickness, and location coordinates of the reservoir where the production well is located.

[0098] The reservoir production prediction method based on the adaptive graph convolutional network provided by the embodiments of the present application inputs the respective dynamic data of the injection well and the production well into a pre-trained prediction model, so as to obtain the predicted oil production that takes into account both the influence of the injection well and the factors of the production well. This method successfully solves the technical problem of insufficient prediction accuracy in the existing reservoir production prediction methods due to only considering the own attributes of the production well, and thus significantly improves the accuracy of reservoir production prediction, providing more reliable data support for the decision-making of oil exploitation.

[0099] Figure 2 Flow schematic of the reservoir production prediction method based on the adaptive graph convolutional network provided by the present application Figure 2 , as Figure 2 shown, on the basis of the Figure 1 embodiment, the reservoir production prediction method based on the adaptive graph convolutional network is described in detail. This method includes:

[0100] S201. Obtain the first dynamic data of the injection well and the second dynamic data of the production well, where the first dynamic data includes the water injection rate and / or bottom hole flowing pressure of the injection well, and the second dynamic data includes at least one of the oil production rate, water production rate, liquid production rate, and bottom hole flowing pressure of the production well.

[0101] Among them, the explanation of step S201 is similar to that of step S101 above and will not be elaborated here.

[0102] S202. Input the first dynamic data and the second dynamic data into the encoding network to obtain initial features, and use the initial features as the input vector of the first convolutional layer in a plurality of cascaded convolutional layers. For any convolutional layer, the input vector of the convolutional layer is used to generate first features through the temporal processing module of the convolutional layer, and then the first features are used to generate second features through the adaptive graph convolutional network of the convolutional layer. The second features are used as the input vector of the next convolutional layer until the first features generated by each of the plurality of cascaded convolutional layers are obtained. Among them, the prediction model includes an encoding network, a feature extraction network, and a decoding network. The feature extraction network includes a plurality of cascaded convolutional layers, and each convolutional layer includes a temporal processing module and an adaptive graph convolutional network.

[0103] Among them, the encoding network is built by cascading a plurality of one-dimensional convolutional layers.

[0104] The encoding process of the encoding network is a process of linearly transforming the feature vectors at each position of the input data. Through convolutional operations and feature extraction, the encoding network can extract the inherent features of the input data and convert the input data into more abstract feature vectors. These feature vectors are aggregated in the spatial and temporal dimensions to form the output feature map of the encoding network.

[0105] The processing process of the encoding network for the input data is as follows:

[0106] The first step is input data parsing: The encoding network receives a multi-dimensional input tensor, which includes multiple dimensions such as batch size, input height, input width, and input channel number.

[0107] The second step is linear transformation: The encoding network extracts features from each position of the input data through convolutional operations. The convolutional operation slides the convolutional kernel on the input tensor, performs weighted summation on the feature vectors at each position, and adds a bias term to obtain a new feature vector.

[0108] The third step is feature aggregation: After the convolutional operation, the feature vectors at each position are converted into new and more abstract feature vectors. These feature vectors are aggregated in the spatial and temporal dimensions to form the output feature map of the encoding network. The output feature map reflects the abstract representation of the input data in different feature dimensions.

[0109] The fourth step is feature vector output: The encoding network converts the output feature map into a series of feature vectors, which can be further processed and analyzed by the subsequent feature extraction network.

[0110] The input dimensions of the encoding network include: the input tensor, the convolutional kernel size, and the initial features of the input data. Among them, (1) the format of the input tensor is: , where represents the input tensor; represents the batch size, which is determined according to human experience and can be, for example, 1; represents the input height, and the input height is used to indicate the length of the historical time range of the dynamic data of each injection well and production well; represents the input width, and the input width N is used to indicate the total number of injection wells and production wells; represents the number of input channels, and the number of input channels is used to indicate the data types of the first dynamic data of the injection wells and the second dynamic data of the production wells. (2) The format of the convolutional kernel is: , where represents the convolutional kernel, represents the number of output channels, that is, the dimension of the feature vector. (3) The format of the initial features is , where represents the vector format of the initial features.

[0111] The purpose of this step of using the encoding network to process the first dynamic data and the second dynamic data is to convert the first dynamic data and the second dynamic data into their respective corresponding feature vectors. These feature vectors not only represent a higher-level and more abstract data representation form, but also can accurately capture the key information in the original dynamic data.

[0112] And the purpose of using the initial features as the input vector of the first convolutional layer in multiple cascaded convolutional layers in this step is to gradually extract and fuse the feature information of the data at different levels to form a more rich, comprehensive and abstract feature representation.

[0113] It can be understood that after the dynamic data of the injection well and the production well are converted into corresponding initial features through the encoding network, these initial features are fed into the feature extraction network for further feature extraction. The feature extraction network consists of multiple cascaded convolutional layers, and each convolutional layer contains a time series processing module and an adaptive graph convolutional network. First, the initial features are input into the first convolutional layer. The time series processing module of this layer can capture the dependencies and dynamic changes of the data in the time series dimension, thereby generating the first feature. Subsequently, the first feature is fed into the adaptive graph convolutional network, which can adaptively use the adjacency matrix to capture the correlations and feature interactions of the first feature in the spatial dimension, generating a more abstract and rich second feature. This process is continuously iterated in multiple cascaded convolutional layers. Each convolutional layer takes the second feature generated by the previous convolutional layer as input and outputs its own first feature. In this way, the feature extraction network can gradually extract and fuse the feature information of the data at different levels, forming a more rich, comprehensive, and abstract feature representation.

[0114] Optionally, the present application provides an implementation process of inputting the first dynamic data and the second dynamic data into the encoding network to obtain the initial features, which specifically includes:

[0115] In the first step, the first dynamic data of the injection well and the second dynamic data of the production well are merged to form a unified input data set.

[0116] In the second step, according to the merged input data set, a multi-dimensional input tensor is constructed. The format of the input tensor includes: batch size B, input height T, input width N, and the number of input channels .

[0117] In the third step, the convolutional kernel size is determined.

[0118] In the fourth step, the prepared convolutional kernel is used to perform a convolution operation on the input tensor, that is, the feature vectors at each position are weighted and summed, and a bias term is added to obtain a new feature vector.

[0119] In the fifth step, the new feature vector is output to the feature extraction network part of the prediction model so that the feature extraction network can further process and analyze it.

[0120] For example, assume that the batch size is 1, there is 1 injection well and 2 production wells, and the dynamic data of three days are used to predict the future. There are 2 types of data for both the injection well and the production well. The data of the injection well are injection rate and bottom hole flowing pressure, and the data of the production well are oil production rate and bottom hole flowing pressure. Then based on the above information, first, the batch size can be determined to be 1, the input height is 3, the input width N is 3, and the number of input channels is 2. Subsequently, the input tensor can be determined as , the convolutional kernel size is . Finally, the format of the initial features is determined as .

[0121] S203. Input the first features generated by each of the multiple cascaded convolutional layers into the decoding network to obtain the predicted oil production.

[0122] Among them, the purpose of this step is to enable the decoding network to make full use of the first features at different levels and output a more accurate oil production value.

[0123] It can be understood that each convolutional layer can learn different feature representations from the input data, and as the convolutional layer deepens, these features gradually become more abstract and complex at different levels. For example, shallow convolutional layers can usually capture the basic features and edge information in the input data, while deep convolutional layers can further extract more advanced and complex features.

[0124] Therefore, inputting the first features generated by each of the multiple cascaded convolutional layers into the decoding network can enable the decoding network to make full use of the feature information at these different levels, thereby outputting a more accurate predicted oil production.

[0125] Optionally, the present application provides a possible implementation method, which specifically includes:

[0126] In the first step, add up the first features generated by each of the multiple cascaded convolutional layers to obtain a fused feature.

[0127] It can be understood that each convolutional layer has its unique perspective and focus when extracting features. Shallow convolutional layers tend to capture the basic and local features in the input data, while deep convolutional layers can extract more abstract and global features. Therefore, adding up these features at different levels can enable the decoding network of the prediction model to utilize the feature information at these different levels simultaneously, thereby enhancing the expression ability and prediction ability of the prediction model.

[0128] Optionally, the present application provides a possible implementation method, which specifically includes: using the addition operation of matrices to add up the first features generated by each of the multiple cascaded convolutional layers to obtain a fused feature.

[0129] Among them, the addition algorithm operation of matrices means that two matrices with the same dimension produce another matrix with the same dimension, where each element is the sum of the elements (i, j) of the original two matrices.

[0130] First, initialize a zero matrix and use it as the initial value of the fused feature to ensure that the dimensions of the zero matrix are the same as those of each first feature.

[0131] Second, traverse each first feature matrix and add them to the fused feature matrix one by one using matrix addition operations until the fused feature is completely fused.

[0132] Second, decode the fused feature through one or more decoding layers of the decoding network to obtain the predicted oil production.

[0133] Among them, in each decoding layer, there is a ReLU activation function and a convolution operation. Among them, the ReLU activation function (Rectified Linear Unit) is a commonly used activation function in artificial neural networks. The calculation method of the ReLU activation function is:

[0134]

[0135] Among them, represents the fused feature. This means that if the input is greater than 0, the output is itself; if the input is less than or equal to 0, the output is 0.

[0136] The purpose of decoding the fused feature through one or more decoding layers of the decoding network is to gradually transform the high-level abstract information contained in the fused feature into specific and interpretable prediction results, that is, the predicted oil production.

[0137] It can be understood that the role of the decoding network is to restore the compact and high-dimensional feature representation obtained in the encoding stage back to the original and easy-to-understand dimensional space, and at the same time use these features to generate the final predicted value. This structured decoding process helps the prediction model to more accurately capture and utilize the information in the fused feature, thereby improving the prediction accuracy.

[0138] Optionally, in the case where the decoding network has two decoding layers, the present application gives a possible implementation method, including:

[0139] First, input the fused feature into the first decoding layer and use the ReLU activation function in the first decoding layer to perform non-linear activation processing on the fused feature to obtain the activated feature after non-linear transformation.

[0140] The purpose of this step is to increase the non-linear expression ability of the prediction model, so that the prediction model can learn more complex feature relationships.

[0141] In the second step, a convolution operation is adopted to perform convolution processing on the activated features obtained after the non-linear transformation in the first step, thereby obtaining the first decoded vector output by the first decoding layer.

[0142] In the third step, the first decoded vector output by the first decoding layer is input into the second decoding layer, and the ReLU activation function is again adopted to perform non-linear activation processing on the first decoded vector output by the first decoding layer, thereby obtaining the activated features after the non-linear transformation again.

[0143] Among them, the purpose of this step is to further enhance the non-linear expression ability of the model and make the decoded vector more in line with the requirements of the prediction target.

[0144] In the fourth step, a convolution operation is again adopted to perform convolution processing on the activated features after the non-linear transformation obtained in the third step, thereby obtaining the predicted oil production output by the first decoding layer.

[0145] The reservoir production prediction method based on the adaptive graph convolutional network provided by the embodiments of the present application first obtains the first dynamic data of injection wells and the second dynamic data of production wells. Subsequently, these dynamic data are jointly input into a pre-trained prediction model, which is composed of an encoding network, a feature extraction network, and a decoding network. In the feature extraction stage, initial features are generated by the encoding network and used as the input vectors of the first convolutional layer in multiple cascaded convolutional layers. Each convolutional layer is embedded with a temporal processing module and an adaptive graph convolutional network. The former is responsible for extracting the first features from the input vectors, and the latter further processes these features using graph convolution technology to generate the second features containing spatial influence relationships. This processing process is passed on step by step until each cascaded convolutional layer generates its corresponding first features. Finally, these first features are fed into the decoding network, and after decoding processing, the accurate predicted oil production is output. This method not only effectively solves the problem of limited prediction accuracy in existing reservoir production prediction due to only considering the attributes of production wells themselves, but also significantly improves the prediction accuracy and reliability by fusing the dynamic data of injection wells and production wells and their spatial influence relationships, thereby providing reliable data support for oil exploitation.

[0146] Figure 3 It is a schematic flow chart of the reservoir production prediction method based on the adaptive graph convolutional network provided by the present application Figure 3 , as Figure 3 shown. On the basis of the Figure 2 embodiment, the process of generating the first features by the temporal processing module of the convolutional layer for the input vectors of the convolutional layer and then generating the second features by the adaptive graph convolutional network of the convolutional layer is described in detail. The method includes:

[0147] S301. Input the input vector of the convolutional layer into the first temporal convolutional network to obtain the first sub-feature, where the temporal processing module includes the first temporal convolutional network and the second temporal convolutional network.

[0148] Among them, the calculation process of the first temporal convolutional network can be expressed by the following formula:

[0149]

[0150] Among them, represents the value of the input time series (encoded vector) at time t, represents the weight (numerical value) of the convolutional kernel, represents the size of the convolutional kernel, represents the output vector after the convolutional operation, represents the dilation coefficient.

[0151] The purpose of this step is to extract preliminary feature representations from the input vector of the convolutional layer.

[0152] It can be understood that since the input vectors of the convolutional layer are obtained through the encoding network, these input vectors already contain the encoding information of the original data (i.e., the first dynamic data of the water injection well and the second dynamic data of the oil production well). And the first temporal convolutional network is good at processing data with time or sequence characteristics, and it can effectively capture the temporal dependence relationships in the data.

[0153] Therefore, by inputting the input vector of the convolutional layer into the first temporal convolutional network, this network can make full use of its temporal characteristics to further mine and extract richer and more detailed feature information from the input vector.

[0154] Optionally, after inputting the input vector of the convolutional layer into the first temporal convolutional network to obtain the first sub-feature, in order to further improve the expression ability of the first sub-feature, it is also necessary to use the sigmoid activation function to perform a non-linear transformation on the first sub-feature, so as to obtain a more accurate first sub-feature.

[0155] Among them, the Sigmoid activation function is a mathematical function that maps any real number input to the interval (0, 1).

[0156] The calculation method of the Sigmoid activation function is:

[0157]

[0158] Among them, represents the output of the first temporal convolutional network, that is, the first sub-feature, Represents the output of the Sigmoid activation function. The sigmoid activation function has a smooth S-shaped curve. When the input value is small, the output value is close to 0; when the input value is large, the output value is close to 1.

[0159] It is understandable that in the feature extraction network of the prediction model, if only linear transformation is used, then no matter how many layers the feature extraction network has, its output can be simplified to a combination of linear functions, which limits the feature extraction network's ability to model complex nonlinear relationships. The sigmoid activation function can introduce nonlinear factors into the feature extraction network of the prediction model, which can enhance the expression ability of the first sub-feature.

[0160] Therefore, after the input vector of the convolution layer is input into the first temporal convolutional network to obtain the first sub-feature, the first sub-feature needs to be processed using the Sigmoid activation function to improve the accuracy of the first sub-feature.

[0161] S302: Input the input of the convolutional layer into a second temporal convolutional network to obtain a second sub-feature.

[0162] Among them, the calculation process of the second temporal convolutional network can be expressed by the following formula:

[0163]

[0164] in, Represents the input time series (encoded vector) at time The value of represents the weight (value) of the convolution kernel, represents the size of the convolution kernel, represents the output vector after the convolution operation, Represents the coefficient of expansion.

[0165] The purpose of this step is to extract different feature representations in the input vector through different temporal convolutional networks (i.e., the second temporal convolutional network), thereby enhancing the diversity and accuracy of the prediction model.

[0166] It is understandable that although the first temporal convolutional network can initially extract feature representations from the input vector, given the complexity and diversity of the input vector, a single network often cannot fully capture all key information. Therefore, by introducing the second temporal convolutional network, it is possible to go deep into different levels and capture more detailed and rich temporal dependencies and features, thereby mining more potential feature information from the input vector.

[0167] Optionally, after the input of the convolutional layer is input into the second temporal convolutional network to obtain the second sub-feature, in order to further improve the representation ability of the second sub-feature, it is also necessary to use the tanh activation function to perform a non-linear transformation on the first sub-feature, so as to obtain a more accurate first sub-feature.

[0168] Among them, the tanh activation function is a mathematical function that maps any real number input to the interval (-1, 1).

[0169] The calculation method of the tanh activation function is:

[0170]

[0171] Among them, represents the output of the second temporal convolutional network, that is, the second sub-feature, represents the output of the tanh activation function. The tanh activation function has a smooth curve shape. When the input value x is small, the output value is close to -1; when the input value x is large, the output value is close to 1; and when the input value x is 0, the output value is exactly 0.

[0172] It can be understood that, first of all, the tanh activation function can introduce stronger non-linearity, which can enable the feature extraction network of the prediction model to capture and represent more complex data relationships. Secondly, the output of the tanh activation function is symmetric about the origin, which helps the feature extraction network of the prediction model to converge faster during training and may reduce the problem of gradient disappearance or explosion, thereby improving the stability and performance of the prediction model. Therefore, after the input of the convolutional layer is input into the second temporal convolutional network to obtain the second sub-feature, the tanh activation function can be used to process the second sub-feature to enhance the representation ability of the second sub-feature, so that the second sub-feature is more accurate.

[0173] S303. Multiply the first sub-feature and the second sub-feature to obtain the first feature.

[0174] Among them, the purpose of this step is to fuse the information contained in the first sub-feature and the second sub-feature respectively and generate a new feature.

[0175] It can be understood that the first sub-feature and the second sub-feature respectively represent different aspects or attributes of the input vector. Therefore, by performing a multiplication operation on these two sub-features, a new feature representation is actually created. This new feature not only contains the information of the first sub-feature but also contains the information of the second sub-feature, realizing the fusion of the information of the two sub-features.

[0176] Optionally, the present application provides a possible implementation method, including: using the Hadamard product algorithm to multiply the first sub-feature and the second sub-feature to obtain the first feature.

[0177] Among them, the Hadamard product is a type of operation on matrices. It operates on matrices with the same shape and produces a third matrix of the same dimension. In mathematics, the Hadamard product is a binary operation that uses two matrices of the same dimension to produce another matrix of the same dimension, where each element (i, j) is the product of the elements (i, j) of the original two matrices.

[0178] By using the Hadamard product algorithm, the product of the first sub-feature and the second sub-feature can achieve a more refined and effective fusion during the fusion process, thereby contributing to obtaining a richer and more expressive first feature.

[0179] It can be understood that during the process of using the Hadamard product algorithm, first, it is determined that the first sub-feature and the second sub-feature have the same dimension; subsequently, the elements at the corresponding positions in the first sub-feature and the second sub-feature are multiplied element by element; then, the result after multiplication is used as the new feature matrix, that is, the first feature; finally, the first feature is input into the feature extraction network of the prediction model for further processing by the prediction model.

[0180] S304. Generate a second feature based on the first feature and the adjacency matrix in the adaptive graph convolutional network.

[0181] Among them, the purpose of this step is to utilize the adaptive characteristics of the adjacency matrix in the adaptive graph convolutional network to dynamically adapt to and process the first feature.

[0182] It can be understood that the adjacency matrix in the adaptive graph convolutional network not only represents the spatial influence relationship between the injection well and the production well, but also allows it to be adaptively adjusted according to the first feature. Therefore, by combining the first feature and the adjacency matrix, the adjacency matrix can capture the dependence relationships among the injection well, the production well, and the first feature, so that these relationships can be encoded into the second feature. In this way, the second feature not only contains the information of the first feature, but also incorporates the spatial influence relationship between the injection well and the production well.

[0183] The oil reservoir production prediction method based on the adaptive graph convolutional network provided by the embodiments of the present application, in the processing flow of the convolutional layer, first sends the input vector into the first temporal convolutional network and the second temporal convolutional network respectively. The first temporal convolutional network is responsible for extracting the preliminary temporal features of the input vector to generate the first sub-feature; while the second temporal convolutional network further analyzes the temporal information of the input vector to generate the second sub-feature. Subsequently, these two sub-features are subjected to a multiplication operation to fuse the temporal characteristics captured by each of them, thereby obtaining a richer and more comprehensive first feature. Immediately afterwards, based on this first feature and the adjacency matrix in the adaptive graph convolutional network, a graph convolution operation is performed, fully considering the spatial influence relationship between the injection well and the production well, and finally generating a second feature containing spatio-temporal information. This method not only effectively fuses temporal and spatial features, but also significantly improves the accuracy and efficiency of feature extraction, providing a more reliable and rich feature input for subsequent prediction tasks.

[0184] Figure 4 Flow schematic of the training method of the oil production prediction model provided by this application Figure 4 , such as Figure 4 shown, this method includes:

[0185] S401. Obtain sample data, where the sample data includes the first dynamic sample data of the injection well, the second dynamic sample data of the production well, and the oil production label. Among them, the first dynamic sample data includes the injection rate and / or bottom hole flowing pressure of the injection well, and the second dynamic sample data includes at least one of the oil production rate, water production rate, liquid production rate, and bottom hole flowing pressure of the production well.

[0186] Among them, the purpose of this step is to collect and prepare the sample data required for training and optimizing the oil production prediction model.

[0187] It can be understood that the first dynamic sample data of the injection well is used to describe the working state of the injection well in the actual production environment and reflect the pressure change of the underground fluid. The second dynamic sample data of the production well is used to characterize the production performance of the production well in the actual production environment and reveal the flow state of the underground fluid. The oil production label represents the specific value reached by the oil production in the actual production environment, and it serves as the target value of the oil production prediction model to guide the training and evaluation of the model.

[0188] Therefore, by obtaining the sample data, it shows that rich and comprehensive training samples can be provided for the oil production prediction model.

[0189] S402. Input the first dynamic sample data and the second dynamic sample data into the oil production prediction model to obtain the predicted oil production output by the oil production prediction model. Among them, the feature extraction network in the oil production prediction model includes an adaptive graph convolutional network, and the initial adjacency matrix in the adaptive graph convolutional network is used to characterize the spatial influence relationship between injection wells and production wells.

[0190] Among them, the purpose of this step is to utilize the information contained in the first dynamic sample data and the second dynamic sample data respectively to train and optimize the oil production prediction model, enabling it to learn the relationship between the dynamic characteristics of injection wells and production wells and oil production.

[0191] It can be understood that the first dynamic sample data and the second dynamic sample data not only reflect the working status and production performance of injection wells and production wells in the actual production environment, but also reveal the pressure changes and flow states of underground fluids.

[0192] The feature extraction network in the oil production prediction model has an initial adjacency matrix built in. This matrix can characterize the spatial influence relationship between injection wells and production wells and has the characteristics of self - adaptation, being able to dynamically adjust itself according to other data.

[0193] Therefore, by taking the first dynamic sample data and the second dynamic sample data as input factors and inputting them into the oil production prediction model, it means that the oil production prediction model can make full use of the self - adaptive characteristics of the feature extraction network and, combined with the first dynamic sample data and the second dynamic sample data, train and optimize the oil production prediction model, thereby generating a predicted oil production that is more in line with the actual production environment.

[0194] S403. Adjust the parameters and the initial adjacency matrix of the oil production prediction model based on the predicted oil production and the oil production label.

[0195] Among them, the purpose of this step is to evaluate the accuracy of the oil production prediction model by comparing the predicted oil production with the actual oil production label, and accordingly adjust the parameters and the initial adjacency matrix of the oil production prediction model.

[0196] Understandably, first, the predicted oil production output by the oil production prediction model is compared with the actual oil production label to evaluate the prediction accuracy of the oil production prediction model. This comparison process can reveal the biases and errors existing in the oil production prediction model during prediction. Subsequently, based on these biases and errors, the parameters and the initial adjacency matrix of the oil production prediction model are adjusted. The parameter adjustment includes, but is not limited to, increasing internal parameters such as the weights and biases of the model to optimize the prediction performance of the oil production prediction model. The adjustment of the initial adjacency matrix may involve changing the representation of the spatial influence relationship between wells to more accurately reflect the interaction between wells in the actual production environment. Through this series of adjustments and optimizations, the prediction accuracy of the oil production prediction model can be gradually improved to better meet the requirements of the actual production environment.

[0197] The training method of the oil production prediction model provided by the embodiment of the present application first obtains sample data, which includes the first dynamic sample data of injection wells, the second dynamic sample data of production wells, and the corresponding oil production labels. Then, these dynamic sample data are input into the oil production prediction model. The feature extraction network of this model uses an adaptive graph convolutional network, and its initial adjacency matrix is used to represent the spatial influence relationship between injection wells and production wells. Through the operation of the model, the predicted oil production can be obtained. To continuously improve the prediction accuracy of the model, it is also necessary to adjust and optimize the parameters and the initial adjacency matrix of the oil production prediction model based on the predicted oil production and the actual oil production label. This method effectively solves the technical problem that in the existing reservoir production prediction, only the attributes of the production wells themselves are considered, while the spatial influence relationship between injection wells and production wells is ignored, resulting in a decrease in the production prediction accuracy. In addition, by introducing the initial adjacency matrix of the adaptive graph convolutional network, this method can capture the dynamic changes in the reservoir more comprehensively, thereby improving the accuracy of oil production prediction.

[0198] Figure 5 It is a flow schematic of the training method of the oil production prediction model provided by the present application Figure 5 , such as Figure 5 shown. Based on the Figure 4 embodiment, the training method of the oil production prediction model is described in detail. This method includes:

[0199] S501. Obtain sample data, where the sample data includes the first dynamic sample data of injection wells, the second dynamic sample data of production wells, and the oil production labels. Among them, the first dynamic sample data includes the injection rate and / or bottom-hole flowing pressure of the injection wells, and the second dynamic sample data includes at least one of the oil production rate, water production rate, liquid production rate, and bottom-hole flowing pressure of the production wells.

[0200] Among them, the explanation of step S501 is similar to that of step S401 above, and will not be elaborated here.

[0201] S502. Input the first dynamic sample data and the second dynamic sample data into the oil production prediction model to obtain the predicted oil production output by the oil production prediction model. Among them, the feature extraction network in the oil production prediction model includes an adaptive graph convolutional network, and the initial adjacency matrix in the adaptive graph convolutional network is used to characterize the spatial influence relationship between injection wells and production wells.

[0202] Among them, the explanation of step S502 is similar to that of step S402 above, and will not be elaborated here.

[0203] S503. Obtain the first static sample data of the injection well and the second static sample data of the production well. The first static sample data includes the permeability, reservoir thickness, and location coordinates of the reservoir where the injection well is located, and the second static sample data includes the permeability, reservoir thickness, and location coordinates of the reservoir where the production well is located.

[0204] Among them, the permeability of the reservoir is a physical quantity that describes the ability of the rock to allow fluids to pass through its pores. The unit is usually darcy (D) or millidarcy (mD). The larger the value, the smaller the resistance of the fluid flowing in the rock.

[0205] The permeabilities of the reservoirs where the injection wells and production wells are located at different positions are different. For example, when the injection well is at position A, the permeability of the reservoir where it is located is 45 mD, and when the injection well is at position B, the permeability of the reservoir where it is located is 23 mD. Similarly, when the production well is at position C, the permeability of the reservoir where it is located is 14 mD, and when the production well is at position D, the permeability of the reservoir where it is located is 25 mD.

[0206] The reservoir thickness refers to the thickness of the reservoir rock in the vertical direction, which reflects the scale of the reservoir rock and its ability to store fluids, and is usually expressed in meters (m) or feet (ft).

[0207] The reservoir thicknesses of the reservoirs where the injection wells and production wells are located at different positions are different. For example, when the injection well is at position A, the reservoir thickness where it is located is 45 m, and when the injection well is at position B, the reservoir thickness where it is located is 40 m. Similarly, when the production well is at position C, the reservoir thickness where it is located is 38 m, and when the production well is at position D, the reservoir thickness where it is located is 42 m.

[0208] The location coordinates refer to the specific positions of the injection well or production well in the formation.

[0209] By obtaining the first static sample data of the water injection well and the second static sample data of the production well, it indicates that the impact of the geological conditions at the locations of the water injection well and the production well on the production work can be evaluated more accurately.

[0210] It can be understood that, first of all, the permeability of the reservoir not only directly affects the water injection effect of the water injection well, but also has a significant impact on the production efficiency of the production well. For example, a reservoir with high permeability allows fluids (including the injected water and the crude oil in the formation) to flow more easily, thereby improving the water injection efficiency of the water injection well and the oil production efficiency of the production well.

[0211] Secondly, the reservoir thickness not only determines the area of the reservoir into which the water injected by the water injection well can penetrate, but also determines the area of the reservoir from which the production well can effectively extract crude oil.

[0212] In addition, the position coordinates not only reflect the actual position of the water injection well in the formation, but also reflect the actual position of the production well in the formation.

[0213] Therefore, by obtaining the first static sample data of the water injection well and the second static sample data of the production well, the production work of the water injection well and the production well can be accurately evaluated according to the geological characteristics of their locations.

[0214] S504. Determine the initial adjacency matrix according to the first static sample data and the second static sample data.

[0215] Among them, the purpose of this step is to obtain a matrix that can represent the spatial influence relationship between the water injection well and the production well.

[0216] It can be understood that the first static sample data of the water injection well is used to characterize the physical properties of the reservoir at the location of the water injection well. These data include the permeability of the reservoir, which reflects the ease of fluid passing through the reservoir rock; the reservoir thickness, which indicates the scale of the reservoir rock in the vertical direction; and the position coordinates, which accurately locate the water injection well in the formation. These parameters together constitute a comprehensive description of the geological characteristics of the reservoir where the water injection well is located.

[0217] The second static sample data of the production well is used to characterize the production potential of the reservoir at the location of the production well. Similarly, these data also include key parameters such as the permeability of the reservoir, the reservoir thickness, and the position coordinates, which also together constitute a comprehensive description of the geological characteristics of the reservoir where the production well is located.

[0218] Therefore, by comprehensively considering the first static sample data of the water injection well and the second static data sample of the production well, it means that the internal relationship between these data can be further analyzed, so as to construct an initial adjacency matrix that can represent the spatial influence relationship between the water injection well and the production well.

[0219] Optionally, the present application provides a possible implementation method, including: using the following seepage resistance coefficient formula, combining the first static sample data and the second static sample data to determine the initial adjacency matrix.

[0220]

[0221] Among them, represents the inter-well seepage resistance coefficient between the water injection well and the oil production well, which is a dimensionless quantity; is the viscosity of crude oil; represents the distance between the water injection well and the oil production well; is the permeability at the location of the water injection well i, is the permeability at the location of the oil production well j; is the oil layer thickness at the location of the water injection well i, is the oil layer thickness at the location of the oil production well j.

[0222] The initial adjacency matrix is represented by . Each element in the initial adjacency matrix represents the reciprocal of the seepage resistance coefficient between the water injection well and the oil production well. The larger the value of this element, the stronger the inter-well connection relationship based on geological data and the greater the influence of the water injection well on the oil well. The initial adjacency matrix is represented by and its expression is as follows:

[0223]

[0224] Among them, represents the inter-well seepage resistance coefficient between the first oil production well and the first water injection well; represents the inter-well seepage resistance coefficient between the first oil production well and the nth water injection well; represents the inter-well seepage resistance coefficient between the mth oil production well and the first water injection well; represents the inter-well seepage resistance coefficient between the mth oil production well and the nth water injection well.

[0225] For example, if there are 4 water injection wells and 5 oil production wells, then based on the above information, first, it is necessary to obtain the first static sample data corresponding to each of the 4 water injection wells, and obtain the second static sample data corresponding to each of the 5 oil production wells; secondly, the initial adjacency matrix can be jointly determined according to the seepage resistance coefficient formula and in combination with these first static sample data and second static sample data. The initial adjacency matrix is represented as:

[0226]

[0227] Among them, the data in the first row represents the inter-well seepage resistance coefficient between the first water injection well and 5 production wells; the data in the second row represents the inter-well seepage resistance coefficient between the second water injection well and 5 production wells; the data in the third row represents the inter-well seepage resistance coefficient between the third water injection well and 5 production wells; the data in the fourth row represents the inter-well seepage resistance coefficient between the second water injection well and 5 production wells.

[0228] S505. Adjust the parameters and the initial adjacency matrix of the oil production prediction model based on the predicted oil production and the oil production label.

[0229] Among them, the explanation of step S505 is similar to the explanation of the above step S403, and will not be elaborated here.

[0230] The training method of the oil production prediction model provided by the embodiments of the present application first obtains sample data, which includes the first dynamic sample data of water injection wells, the second dynamic sample data of production wells, and the oil production label. At the same time, the first static sample data of water injection wells and the second static sample data of production wells are obtained. Subsequently, according to the first static sample data and the second static sample data, the initial adjacency matrix is determined, and the initial adjacency matrix is embedded into the feature extraction network of the oil production prediction model. Finally, these dynamic sample data are input into the oil production prediction model, and the prediction model can combine the initial adjacency matrix and the dynamic sample data to predict the oil production. Finally, the predicted oil production output by the oil production prediction model is compared with the actual oil production label, and the parameters and the initial adjacency matrix of the oil production prediction model are adjusted according to the comparison result to improve the prediction accuracy of the model.

[0231] By comprehensively considering the dynamic and static attributes of water injection wells and production wells, as well as the spatial relationship between them, the method effectively solves the technical problem that in the existing reservoir production prediction, only the attributes of production wells themselves are considered, while the spatial influence relationship between water injection wells and production wells is ignored, and the static attributes of water injection wells and production wells are not fully utilized, resulting in a reduction in the accuracy of production prediction, thereby significantly improving the accuracy of oil production prediction.

[0232] Figure 6 It is a schematic diagram of the prediction structure of the reservoir production prediction method based on the adaptive graph convolutional network provided by the present application, as Figure 6 shown, including:

[0233] The first step is to input the first dynamic data of water injection wells and the second dynamic data of production wells into the encoding network;

[0234] The second step is that the encoding network performs a convolution operation on the first dynamic data of water injection wells and the second dynamic data of production wells to obtain the initial features;

[0235] In the third step, input the initial features into the first convolutional layer of the feature extraction network. The first convolutional layer uses the first temporal convolutional network to process the initial features, obtaining the first sub-features output by the first temporal convolutional network, and then continues to process the features using the sigmoid activation function to obtain the processed first sub-features;

[0236] Meanwhile, the first convolutional layer uses the second temporal convolutional network to process the initial features, obtaining the first sub-features output by the second temporal convolutional network, and then continues to process the features using the tanh activation function to obtain the processed first sub-features;

[0237] In the fourth step, multiply the processed first sub-features and the processed second sub-features to obtain the first feature;

[0238] In the fifth step, input the first feature of the first convolutional layer into the adaptive graph convolutional network built in the feature extraction network to obtain the second feature output by the adaptive graph convolutional network;

[0239] In the sixth step, input the second feature output by the adaptive graph convolutional network into the second convolutional layer as the input vector of the second convolutional layer until the first features generated by each of the L convolutional layers are obtained;

[0240] In the seventh step, after the convolution operations of the first features generated by each of the L convolutional layers are completed, sum up the L first features to obtain the fused feature;

[0241] In the eighth step, input the fused feature into the decoding network, and decode the fused feature through the 2 decoding layers in the decoding network to obtain the predicted oil production.

[0242] Figure 7 The structural schematic diagram of the reservoir production prediction device based on the adaptive graph convolutional network provided by the present application is as Figure 7 shown. The reservoir production prediction device 700 based on the adaptive graph convolutional network provided in this embodiment includes:

[0243] An acquisition module 701, configured to acquire the first dynamic data of the injection well and the second dynamic data of the production well, where the first dynamic data includes the injection rate and / or the bottom-hole flowing pressure of the injection well, and the second dynamic data includes at least one of the oil production rate, water production rate, liquid production rate, and bottom-hole flowing pressure of the production well;

[0244] An input module 702, configured to input the first dynamic data and the second dynamic data into a pre-trained prediction model to obtain the predicted oil production output by the prediction model, wherein a feature extraction network in the prediction model includes an adaptive graph convolutional network, and an adjacency matrix in the adaptive graph convolutional network is used to characterize the spatial influence relationship between the water injection wells and the oil production wells.

[0245] In a possible implementation manner, the input module 702 is further configured to input the first dynamic data and the second dynamic data into the encoding network to obtain initial features, use the initial features as input vectors of the first convolutional layer in the multiple cascaded convolutional layers, and for any convolutional layer, generate first features by passing the input vectors of the convolutional layer through a temporal processing module of the convolutional layer, and then generate second features by passing the first features through the adaptive graph convolutional network of the convolutional layer, and use the second features as input vectors of the next convolutional layer until the first features respectively generated by the multiple cascaded convolutional layers are obtained;

[0246] The input module 702 is specifically configured to input the first features respectively generated by the multiple cascaded convolutional layers into the decoding network to obtain the predicted oil production.

[0247] In a possible implementation manner, the input module 702 is further configured to input the input vectors of the convolutional layer into a first temporal convolutional network to obtain first sub-features;

[0248] The input module 702 is further configured to input the input vectors of the convolutional layer into a second temporal convolutional network to obtain second sub-features;

[0249] The apparatus further includes: a product module 703;

[0250] The product module 703 is configured to multiply the first sub-features and the second sub-features to obtain the first features;

[0251] The apparatus further includes: a generation module 704;

[0252] The generation module 704 is configured to generate the second features based on the first features and the adjacency matrix in the adaptive graph convolutional network.

[0253] In a possible implementation manner, the apparatus further includes: an accumulation module 705;

[0254] The accumulation module 705 is configured to accumulate the first features respectively generated by the multiple cascaded convolutional layers to obtain fused features;

[0255] The apparatus further includes: a decoding module 706;

[0256] The decoding module 706 is configured to decode the fused features through one or more decoding layers of the decoding network to obtain the predicted oil production.

[0257] The prediction device 700 for injection-production reservoir production provided in this embodiment can execute the method provided in the above method embodiment, and its implementation principle and technical effects are similar, which will not be elaborated here.

[0258] Figure 8 It is a schematic structural diagram of a training device for an oil production prediction model provided by this application. As Figure 8 shown, the training device 800 for the oil production prediction model provided in this embodiment includes:

[0259] An acquisition module 801, configured to acquire sample data, where the sample data includes first dynamic sample data of an injection well, second dynamic sample data of a production well, and an oil production label. Among them, the first dynamic sample data includes the injection rate and / or bottom hole flowing pressure of the injection well, and the second dynamic sample data includes at least one of the oil production rate, water production rate, liquid production rate, and bottom hole flowing pressure of the production well;

[0260] An input module 802, configured to input the first dynamic sample data and the second dynamic sample data into the oil production prediction model to obtain the predicted oil production output by the oil production prediction model. Among them, the feature extraction network in the oil production prediction model includes an adaptive graph convolutional network, and the initial adjacency matrix in the adaptive graph convolutional network is used to characterize the spatial influence relationship between the injection well and the production well;

[0261] An adjustment module 803, configured to adjust the parameters of the oil production prediction model and the initial adjacency matrix based on the predicted oil production and the oil production label.

[0262] In a possible implementation manner, the acquisition module 801 is further configured to acquire the first static sample data of the injection well and the second static sample data of the production well. The first static sample data includes the permeability, reservoir thickness, and position coordinates of the reservoir where the injection well is located, and the second static sample data includes the permeability, reservoir thickness, and position coordinates of the reservoir where the production well is located;

[0263] The device further includes: a determination module 804;

[0264] The determination module 804 is configured to determine the initial adjacency matrix according to the first static sample data and the second static sample data.

[0265] The training device 800 for the oil production prediction model provided in this embodiment can execute the method provided in the above method embodiment. The implementation principle and technical effect are similar, and will not be elaborated here in this embodiment.

[0266] Figure 9 It is a schematic structural diagram of the reservoir production prediction device based on the adaptive graph convolutional network provided in this application. As Figure 9 shown, the electronic device 900 provided in this embodiment includes: at least one processor 901 and a memory 902. Optionally, the device 900 further includes a communication component 903. Among them, the processor 901, the memory 902, and the communication component 903 are connected through a bus 904.

[0267] In the specific implementation process, at least one processor 901 executes the computer-executable instructions stored in the memory 902, so that at least one processor 901 executes the above method.

[0268] For the specific implementation process of the processor 901, reference can be made to the above method embodiment. The implementation principle and technical effect are similar, and will not be elaborated here in this embodiment.

[0269] In the above embodiment, it should be understood that the processor may be a central processing unit (English: Central Processing Unit, abbreviated as: CPU), or other general-purpose processors, digital signal processors (English: Digital Signal Processor, abbreviated as: DSP), application-specific integrated circuits (English: Application Specific Integrated Circuit, abbreviated as: ASIC), etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the invention can be directly embodied as being executed and completed by a hardware processor, or executed and completed by a combination of hardware and software modules in the processor.

[0270] The memory may include a high-speed memory (Random Access Memory, RAM), and may also include a non-volatile memory (Non-volatile Memory, NVM), such as at least one disk memory.

[0271] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, an Extended Industry Standard Architecture (EISA) bus, or the like. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the buses in the drawings of this application are not limited to only one bus or one type of bus.

[0272] This application also provides a computer program product, including a computer program which, when executed by a processor, implements the above method.

[0273] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above method.

[0274] The above-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk, or an optical disk. The readable storage medium can be any available medium accessible by a general-purpose or special-purpose computer.

[0275] An exemplary readable storage medium is coupled to the processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an Application Specific Integrated Circuit (ASIC). Of course, the processor and the readable storage medium can also exist as discrete components in a device.

[0276] The division of units is only a logical function division. In actual implementation, there can be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection between each other can be through some interfaces, and the indirect coupling or communication connection of devices or units can be in an electrical, mechanical, or other forms.

[0277] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0278] In addition, in each embodiment of the present invention, the various functional units may be integrated in a processing unit, may exist separately as individual physical units, or two or more units may be integrated in one unit.

[0279] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in each embodiment of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0280] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When this program is executed, it executes the steps including the above method embodiments; and the aforementioned storage medium includes: various media such as ROM, RAM, magnetic disks, or optical discs that can store program codes.

[0281] Finally, it should be noted that: after considering the specification and practicing the invention disclosed herein, those skilled in the art will readily think of other implementation schemes of the present invention. The present invention aims to cover any variations, uses, or adaptive changes of the present invention, and these variations, uses, or adaptive changes follow the general principles of the present invention and include common general knowledge or conventional technical means in the technical field not disclosed in the present invention. It is not limited to the precise structures already described and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present invention is only limited by the appended claims.

Claims

1. A method for predicting oil reservoir production based on an adaptive graph convolutional network, characterized in that: include: Acquire first dynamic data of the water injection well and second dynamic data of the oil production well, wherein the first dynamic data includes the water injection rate and / or bottom hole flow pressure of the water injection well, and the second dynamic data includes at least one of the oil production rate, water production rate, liquid production rate, and bottom hole flow pressure of the oil production well; Inputting the first dynamic data and the second dynamic data into a pre-trained prediction model to obtain a predicted oil production output by the prediction model, wherein the feature extraction network in the prediction model includes an adaptive graph convolution network, and the adjacency matrix in the adaptive graph convolution network is used to characterize the spatial influence relationship between the water injection well and the oil production well; The prediction model includes an encoding network, a feature extraction network and a decoding network, wherein the feature extraction network includes a plurality of cascaded convolutional layers, and each of the convolutional layers includes a timing processing module and the adaptive graph convolutional network; The step of inputting the first dynamic data and the second dynamic data into a pre-trained prediction model to obtain a predicted oil production output by the prediction model comprises: Inputting the first dynamic data and the second dynamic data into the encoding network to obtain initial features, using the initial features as input vectors of the first convolutional layer among the multiple cascaded convolutional layers, and for any convolutional layer, passing the input vector of the convolutional layer through the timing processing module of the convolutional layer to generate a first feature, then passing the first feature through the adaptive graph convolutional network of the convolutional layer to generate a second feature, and using the second feature as the input vector of the next convolutional layer, until the first features generated by each of the multiple cascaded convolutional layers are obtained; Inputting the first features generated by each of the plurality of cascaded convolutional layers into the decoding network to obtain the predicted oil production; The timing processing module includes a first timing convolutional network and a second timing convolutional network, wherein the input vector of the convolutional layer is passed through the timing processing module of the convolutional layer to generate a first feature, and then the first feature is passed through the adaptive graph convolutional network of the convolutional layer to generate a second feature, including: Inputting the input vector of the convolutional layer into the first temporal convolutional network to obtain a first sub-feature; Inputting the input vector of the convolutional layer into the second temporal convolutional network to obtain a second sub-feature; multiplying the first sub-feature and the second sub-feature to obtain the first feature; The second feature is generated based on the first feature and an adjacency matrix in the adaptive graph convolutional network.

2. The method according to claim 1, characterized in that The step of inputting the first features generated by each of the plurality of cascaded convolutional layers into the decoding network to obtain the predicted oil production comprises: Accumulating the first features generated by each of the multiple cascaded convolutional layers to obtain a fused feature; The fused features are decoded by one or more decoding layers of the decoding network to obtain the predicted oil production.

3. The method according to claim 1 or 2, characterized in that: The adjacency matrix is ​​obtained by training the prediction model based on the initial adjacency matrix, and the initial adjacency matrix is ​​determined based on the first static data of the water injection well and the second static data of the oil production well. The first static data includes the permeability, reservoir thickness and location coordinates of the reservoir where the water injection well is located, and the second static data includes the permeability, reservoir thickness and location coordinates of the reservoir where the oil production well is located.

4. A training method for an oil production prediction model, characterized in that: include: Acquire sample data, wherein the sample data includes first dynamic sample data of a water injection well, second dynamic sample data of an oil production well, and an oil production label, wherein the first dynamic sample data includes the water injection rate and / or bottom hole flow pressure of the water injection well, and the second dynamic sample data includes at least one of the oil production rate, water production rate, liquid production rate, and bottom hole flow pressure of the oil production well; Inputting the first dynamic sample data and the second dynamic sample data into an oil production prediction model to obtain a predicted oil production output by the oil production prediction model, wherein the feature extraction network in the oil production prediction model includes an adaptive graph convolution network, and the initial adjacency matrix in the adaptive graph convolution network is used to characterize the spatial influence relationship between the water injection well and the oil production well; adjusting the parameters of the oil production prediction model and the initial adjacency matrix based on the predicted oil production and the oil production label; The oil production prediction model includes an encoding network, a feature extraction network and a decoding network, wherein the feature extraction network includes a plurality of cascaded convolutional layers, each of which includes a time series processing module and the adaptive graph convolutional network; The step of inputting the first dynamic sample data and the second dynamic sample data into an oil production prediction model to obtain a predicted oil production output by the oil production prediction model comprises: Inputting the first dynamic sample data and the second dynamic sample data into the encoding network to obtain initial sample features, using the initial sample features as input vectors of the first convolutional layer among the multiple cascaded convolutional layers, and for any convolutional layer, passing the input vector of the convolutional layer through the time series processing module of the convolutional layer to generate first sample features, then passing the first sample features through the adaptive graph convolutional network of the convolutional layer to generate second sample features, and using the second sample features as input vectors of the next convolutional layer, until first sample features generated by each of the multiple cascaded convolutional layers are obtained; Inputting the first sample features generated by each of the plurality of cascaded convolutional layers into the decoding network to obtain the predicted oil production; The timing processing module includes a first timing convolutional network and a second timing convolutional network, wherein the input vector of the convolutional layer is passed through the timing processing module of the convolutional layer to generate a first sample feature, and then the first sample feature is passed through the adaptive graph convolutional network of the convolutional layer to generate a second sample feature, including: Inputting the input vector of the convolution layer into the first temporal convolutional network to obtain a first sub-sample feature; Inputting the input vector of the convolution layer into the second temporal convolutional network to obtain a second sub-sample feature; Multiplying the first sub-sample feature and the second sub-sample feature to obtain the first sample feature; The second sample feature is generated based on the first sample feature and an initial adjacency matrix in the adaptive graph convolutional network.

5. The method according to claim 4, characterized in that Also includes: Acquire the first static sample data of the water injection well and the second static sample data of the oil production well, wherein the first static sample data includes the permeability, reservoir thickness and position coordinates of the reservoir at the location of the water injection well, and the second static sample data includes the permeability, reservoir thickness and position coordinates of the reservoir at the location of the oil production well; The initial adjacency matrix is ​​determined according to the first static sample data and the second static sample data.

6. An oil reservoir production prediction device based on an adaptive graph convolutional network, characterized in that: include: Memory, processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory, so that the processor performs the method according to any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 5 when executed by a processor.

8. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 5 when being executed by a processor.

Images