Distributed node parameter prediction method based on physics-space-data multi-graph fusion

Abstract

The invention discloses a distributed node parameter prediction method based on physics-space-data multi-graph fusion, and the method comprises the steps: carrying out the dynamic prediction of distributed node parameters through a space-data-physics multi-graph fusion graph network model, and constructing a physical mechanism graph, a space geometric graph and a data driving graph of distributed nodes; various information such as physical relations, spatial distances and data statistical laws are included during composition, complex characteristics of the distributed nodes are described more comprehensively, a distributed node parameter prediction model has higher interpretability, and through multi-graph fusion, the distributed node parameter prediction model is more accurate. The distributed node parameter prediction model can better capture a complex nonlinear relationship in the distributed nodes, so that the prediction accuracy is improved; the method can adapt to diversified application scenes through multi-graph fusion, and the data driving graph can distribute dynamic changes of node production environment data and is suitable for dynamic prediction. The space geometric diagram and the physical mechanism diagram can capture static characteristics of distributed nodes, and the method is suitable for long-term prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and specifically relates to a distributed node parameter prediction method based on multi-graph fusion of physics-space-data. Background Art

[0002] During the injection and production process of a gas storage reservoir, the operating parameters of the gas storage reservoir have an accumulative effect. That is, the reservoir capacity and formation pressure will change with the injection and production activities of the gas. The change of the formation pressure of the gas storage reservoir directly affects the overall safety of the gas storage reservoir system. Overpressure or insufficient pressure will lead to accidents such as rupture and leakage of the gas storage reservoir. Moreover, the adjustment of the injection and production rates of the gas storage reservoir also needs to ensure that the formation pressure is within a safe range. Therefore, establishing a suitable prediction model to comprehensively predict the formation pressure of the gas storage reservoir is of great significance for the safety and reliability in the actual production operation of the gas storage reservoir.

[0003] Large gas storage reservoirs usually contain multiple injection-production well fields. During the production operation process, as the reservoir capacity of each well field increases, the formation pressure will also increase accordingly, and vice versa. To ensure the safety during the production operation of the gas storage reservoir, the relationship between the reservoir capacity and the formation pressure can be used to dynamically predict the formation pressure of each well field of the gas storage reservoir.

[0004] At present, domestic and foreign scholars usually use numerical simulation methods to predict the formation pressure of depleted oil and gas reservoir gas storage reservoirs, and conduct three-dimensional simulation calculations on the formation pressure of the gas storage reservoir. This method can predict the formation pressure to a certain extent and has a fast prediction speed. However, the accuracy of this method's model is low, and the calculation results deviate greatly from the actual situation, making it difficult to achieve a good prediction effect. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a distributed node parameter prediction method based on multi-graph fusion of physics-space-data, which can improve the prediction accuracy of distributed node parameters by fusing the physical mechanism graph, spatial geometry graph, and data-driven graph of distributed nodes.

[0006] To achieve the above purpose, the present invention provides the following technical solutions:

[0007] A distributed node parameter prediction method based on multi-graph fusion of physics-space-data includes the following steps:

[0008] Step 1: Establish a data set

[0009] Collect characteristic parameters related to the prediction parameters of the distributed nodes. The characteristic parameters include the production data of each node and the production environment data of the distributed nodes; conduct theoretical calculations on the prediction parameters to construct a data set;

[0010] Step 2: Create a graph for the distributed nodes

[0011] Based on the physical relationships among the nodes in the distributed nodes, determine the connection relationships and edge weights among the nodes, and construct a physical mechanism diagram of the distributed nodes;

[0012] Based on the spatial distances among the nodes in the distributed nodes, determine the connection relationships and edge weights among the nodes, and construct a spatial geometry diagram of the distributed nodes;

[0013] Based on the data statistical laws of the distributed nodes, determine the connection relationships and edge weights among the nodes, and construct a data-driven diagram of the distributed nodes;

[0014] Step Three: Perform multi-graph fusion based on the attention mechanism

[0015] Calculate the attention of each edge in the physical mechanism diagram, spatial geometry diagram, and data-driven diagram respectively, and fuse the weights of the corresponding edges in the physical mechanism diagram, spatial geometry diagram, and data-driven diagram to obtain the fused edge weights, thus obtaining a fused graph;

[0016] Step Four: Train the graph network model

[0017] Input the fused graph into the graph neural network model, and use the data set to train the graph neural network model to obtain a distributed node parameter prediction model;

[0018] Step Five: Parameter prediction

[0019] Collect the characteristic parameters associated with the prediction parameters of the distributed nodes, input the collected characteristic parameters into the distributed node parameter prediction model, and predict the prediction parameters.

[0020] Furthermore, the distributed nodes are well sites, and each well in the well site serves as a node.

[0021] Furthermore, the well site is a gas storage reservoir, and the prediction parameter is formation pressure.

[0022] Furthermore, in the first step, the production data of the well includes the daily gas injection volume per well and the wellhead pressure; the production environment data of the gas storage reservoir is well test data, including tubing resistance coefficient, mid-depth of gas layer, annual average wellhead temperature, geothermal gradient, critical temperature of natural gas, critical pressure of natural gas, inner diameter of tubing, relative density of natural gas, flow area of gas layer, permeability, viscosity of natural gas, and thickness of gas layer;

[0023] For each well, the bottom-hole pressure is:

[0024]

[0025] where: pwf is the bottom-hole pressure; pwh is the wellhead pressure; qin is the daily gas injection volume of a single well; s is the skin factor; λ is the tubing resistance coefficient; φ 2 is the inner diameter of the tubing; e is the natural logarithm; and:

[0026] w = 0.03415γ g H / (TZ)

[0027] where: γ g is the relative density of natural gas; H is the depth of the middle part of the gas reservoir;

[0028] The formation pressure is:

[0029]

[0030] where: Pr is the formation pressure; A is the flow area of the gas reservoir; k is the permeability; μ is the viscosity of natural gas; B g is the gas volume coefficient; L is the thickness of the gas reservoir; and:

[0031]

[0032] where: T is the average temperature of the moving gas column in the wellbore; Z is the average deviation coefficient of the moving gas column in the wellbore.

[0033] Furthermore, in the second step, when constructing the physical mechanism diagram, the physical relationships include reservoir connectivity and pressure propagation;

[0034] The method for determining the connection relationship and edge weights between nodes based on reservoir connectivity is as follows: Use logging data to construct the structural model and property model of the reservoir, and construct the three-dimensional geological model of the reservoir; According to the geological model, divide the reservoir into different units or regions and conduct connectivity analysis: If two well sites are located in the same reservoir unit, it is considered that the two well sites are geologically connected, and an edge is connected between the two well sites. Based on the permeability and porosity of the reservoir, calculate the weight of the edge. The higher the permeability, the stronger the connectivity, and the greater the weight; If there is a fluid flow path between two well sites, it is considered that the two well sites are fluidly connected, and an edge is connected between the two well sites. Based on the flow rate, calculate the weight of the edge. The higher the flow rate, the stronger the connectivity, and the greater the weight;

[0035] The method for determining the connection relationship and edge weights between nodes based on pressure propagation is as follows: Analyze the time series data of pressure changes based on the pressure test data between well sites; Conduct correlation analysis and time delay analysis on the time series data, and calculate the correlation coefficient of pressure changes between well sites; If the correlation coefficient between two well sites exceeds the set threshold, it is considered that there is a pressure propagation relationship between the two well sites, and an edge is connected between the two well sites; Calculate the weight based on the correlation and time delay of pressure propagation. The higher the correlation and the shorter the time delay, the greater the weight;

[0036] Use weighted average to comprehensively use reservoir connectivity and pressure propagation to obtain the edge weights in the physical mechanism diagram:

[0037]

[0038] Where: is the edge weight between well site i and well site j in the physical mechanism diagram; is the edge weight obtained based on reservoir connectivity; α is the weight coefficient based on reservoir connectivity; is the edge weight obtained based on pressure propagation; β is the weight coefficient based on pressure propagation.

[0039] Furthermore, in the second step, when constructing the spatial geometry diagram, obtain the coordinates of each well site, and use the coordinates of the well sites to calculate the Euclidean distance between the well sites:

[0040]

[0041] In the formula: (x i , y i ) and (x j , y j ) are the coordinates of well site i and well site j respectively; D ij is the distance between well site i and well site j;

[0042] Set the distance threshold θ. When D ij < θ, there is an edge between well site i and well site j; different weights are assigned according to the distance, and the closer the distance, the greater the weight. Then the edge weight in the spatial geometry diagram is:

[0043]

[0044] Where: is the edge weight between well site i and well site j in the spatial geometry diagram; σ is a parameter that controls the weight decay rate.

[0045] Furthermore, in the second step, when constructing the data-driven diagram, based on the historical data of the well sites, establish edges according to the cosine similarity:

[0046]

[0047] Where: S ij is the edge between well site i and well site j in the data-driven diagram; x i and x j are the reservoir capacity time series data of well site i and well site j respectively;

[0048] S ijThe larger it is, the more similar wellsite i and wellsite j are. Using the K-nearest neighbor strategy, each wellsite is only connected to the K most similar wellsites, and the normalized cosine similarity is used as the edge weight:

[0049]

[0050] Where: is the edge weight between wellsite i and wellsite j in the data-driven graph; S min is the minimum value among all S ij ; S max is the maximum value among all S ij .

[0051] Furthermore, in the third step, the weights of each edge in the physical mechanism graph, the spatial geometry graph, and the data-driven graph are different. A learnable MLP is used to calculate the attention of each edge in the physical mechanism graph, the spatial geometry graph, and the data-driven graph:

[0052]

[0053] Where: α k (i,j) is the edge weight between wellsite i and wellsite j in the k-th graph, k = 1, 2, 3, representing the physical mechanism graph, the spatial geometry graph, and the data-driven graph respectively; A k is the weighted adjacency matrix of the k-th graph; f(A k (i,j),x i ,x j ) is a neural network, namely a learnable MLP; x i and x j are node features;

[0054] The fused edge weight is:

[0055]

[0056] Where: A fusion (i,j) is the edge weight between wellsite i and wellsite j in the fused graph.

[0057] Furthermore, in the fourth step, the graph neural network model includes a graph convolutional network, a non-linear layer, and a fully connected layer; during the process of training the graph neural network model using the dataset, the Adam optimizer and the MSE loss function are adopted, and the initial learning rate is set to 0.001.

[0058] The beneficial effects of the present invention are as follows:

[0059] The distributed node parameter prediction method based on physical-space-data multi-graph fusion of the present invention uses a graph network model of space-data-physical multi-graph fusion to dynamically predict distributed node parameters, constructs a physical mechanism graph, a spatial geometry graph, and a data-driven graph of the distributed nodes. When constructing the graphs, various information such as physical relationships, spatial distances, and data statistical laws is included, more comprehensively describing the complex characteristics of the distributed nodes, making the distributed node parameter prediction model have higher interpretability. Through multi-graph fusion, the distributed node parameter prediction model can better capture the complex non-linear relationships in the distributed nodes, thereby improving the prediction accuracy; the distributed node parameter prediction model using multi-graph fusion can adapt to diverse application scenarios. The data-driven graph can adapt to the dynamic changes of the production environment data of the distributed nodes and is suitable for dynamic prediction; the spatial geometry graph and the physical mechanism graph can capture the static characteristics of the distributed nodes and are suitable for long-term prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to make the objectives, technical solutions, and beneficial effects of the present invention clearer, the present invention provides the following drawings for description:

[0061] Figure 1 It is a flowchart of the distributed node parameter prediction method based on physical-space-data multi-graph fusion of the present invention;

[0062] Figure 2 It is a schematic diagram of the distributed node parameter prediction method based on physical-space-data multi-graph fusion of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0063] The following further describes the present invention in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments cited are not intended to limit the present invention.

[0064] 1. Method Flow

[0065] The distributed node parameter prediction method based on physical-space-data multi-graph fusion in this embodiment includes the following steps:

[0066] Step 1: Establish a data set

[0067] Collect the characteristic parameters associated with the prediction parameters of the distributed nodes. The characteristic parameters include the production data of each node and the production environment data of the distributed nodes; perform theoretical calculations on the prediction parameters to construct a data set.

[0068] Step 2: Construct graphs for the distributed nodes

[0069] Based on the physical relationships between the nodes in the distributed nodes, determine the connection relationships and edge weights between the nodes, and construct the physical mechanism graph of the distributed nodes.

[0070] Determine the connection relationships and edge weights among the nodes based on the spatial distances between the nodes in the distributed nodes, and construct a spatial geometric graph of the distributed nodes.

[0071] Determine the connection relationships and edge weights among the nodes based on the data statistical laws of the distributed nodes, and construct a data-driven graph of the distributed nodes.

[0072] Step 3: Perform multi-graph fusion based on the attention mechanism

[0073] Calculate the attention of each edge in the physical mechanism graph, spatial geometric graph, and data-driven graph respectively, and fuse the weights of the corresponding edges in the physical mechanism graph, spatial geometric graph, and data-driven graph to obtain the fused edge weights, thus obtaining a fused graph.

[0074] Step 4: Train the graph network model

[0075] Input the fused graph into the graph neural network model, and use the data set to train the graph neural network model to obtain a distributed node parameter prediction model.

[0076] Step 5: Parameter prediction

[0077] Collect the characteristic parameters associated with the prediction parameters of the distributed nodes, input the collected characteristic parameters into the distributed node parameter prediction model, and predict the prediction parameters.

[0078] In this embodiment, the distributed nodes are well sites, and each well site in the well site is used as a node. Specifically, taking a gas storage reservoir as an example, the specific implementation manner of the distributed node parameter prediction method based on physical-space-data multi-graph fusion in this embodiment is described. That is, in this embodiment, the well site is a gas storage reservoir, and the prediction parameter is the formation pressure.

[0079] As Figure 1 shown, for the distributed node parameter prediction method based on physical-space-data multi-graph fusion in this embodiment, a graph network model of each well site in the gas storage reservoir is constructed. A physical mechanism graph is constructed based on the reservoir connectivity and pressure propagation of the injection-production well sites in the gas storage reservoir, a spatial geometric graph is constructed based on the spatial distances of the injection-production well sites in the gas storage reservoir, and a data-driven graph is constructed based on the characteristic cosine similarity between the injection-production well sites in the gas storage reservoir. Then, the attention mechanism is used for multi-graph fusion. The fused graph contains information from three aspects: the physical mechanism, space, and data of the gas storage reservoir. The graph network model is trained using GNN, making the dynamic prediction of the formation pressure more accurate, and solving the problem that the change of the formation pressure in the production operation process of the gas storage reservoir affects the safety of the gas storage reservoir.

[0080] 2. Establish a data set

[0081] In this embodiment, the reservoir capacity of each well field of the gas storage reservoir is used as the input feature of the node, and the formation pressure of the corresponding well field is used as the target value of the node. A relational expression between the formation pressure and the cumulative reservoir capacity of the gas storage reservoir is established. According to the fitted relational expression and the actual injection and production gas volumes, with days as the smallest time step, the reservoir capacity of the gas storage reservoir is changed, and the change of the formation pressure within the corresponding evaluation period can be obtained.

[0082] Specifically, for an actually operating gas storage reservoir, production data and well test data are collected and sorted out. Among them, the production data of the well field includes the daily gas injection volume per well, the wellhead pressure, etc.; the production environment data of the gas storage reservoir is the well test data, and the well test data includes the tubing resistance coefficient, the mid-depth of the gas layer, the annual average temperature at the wellhead, the geothermal gradient, the critical temperature of natural gas, the critical pressure of natural gas, the inner diameter of the tubing, the relative density of natural gas, the flow area of the gas layer, the permeability, the viscosity of natural gas, and the thickness of the gas layer. Through the collected data, the reservoir capacity and formation pressure of each well field are theoretically calculated. The reservoir capacity can be calculated by the daily gas injection volume per well and the gas injection days. For the calculation of the formation pressure, the bottom-hole pressure is first obtained by using the vertical tubing flow equation in the wellbore. For each well field, the bottom-hole pressure is:

[0083]

[0084] where: pwf is the bottom-hole pressure; pwh is the wellhead pressure; qin is the daily gas injection volume per well; s is the skin factor; λ is the tubing resistance coefficient; T is the average temperature of the moving gas column in the wellbore; Z is the average deviation coefficient of the moving gas column in the wellbore, which can be obtained by referring to the general gas deviation coefficient chart according to the reduced temperature and reduced pressure; φ 2 is the inner diameter of the tubing; e is the natural logarithm; and:

[0085] s = 0.03415γ g H / (TZ)

[0086] where: γ g is the relative density of natural gas; H is the mid-depth of the gas layer;

[0087] The formation pressure is:

[0088]

[0089] where: Pr is the formation pressure; A is the flow area of the gas layer; k is the permeability; μ is the viscosity of natural gas; B g is the gas volume factor; L is the thickness of the gas layer; and:

[0090]

[0091] where: T is the average temperature of the moving gas column in the wellbore; Z is the average deviation coefficient of the moving gas column in the wellbore.

[0092] 3. Construct the physical mechanism diagram

[0093] The nodes of the physical mechanism diagram of the gas storage injection-production well site are set as each injection-production well site. The connection relationship between the nodes and the weights of the edges depend on two aspects: reservoir connectivity and pressure propagation. That is, when constructing the physical mechanism diagram, the physical relationships include reservoir connectivity and pressure propagation.

[0094] (1) Based on reservoir connectivity

[0095] The method for determining the connection relationship and edge weights between each node based on reservoir connectivity is as follows: Collect well logging data in the wellbore (such as acoustic logging, resistivity logging, etc.), use the well logging data to construct the structural model and property model of the reservoir, conduct the construction of the 3D geological model of the reservoir, divide the reservoir into 3D grids, and finally determine parameters such as the distribution, thickness, porosity, and permeability of the reservoir. According to the geological model, divide the reservoir into different units or regions and conduct connectivity analysis. If two well sites are located in the same reservoir unit, they are considered to be geologically connected, and an edge is connected between them. Based on the permeability and porosity of the reservoir, calculate the weight of the edge. The higher the permeability, the stronger the connectivity, and the greater the weight. If there is a fluid flow path between two well sites, they are considered to be fluidly connected, and an edge is connected between them. Based on the flow rate, calculate the weight of the edge. The higher the flow rate, the stronger the connectivity, and the greater the weight.

[0096] (2) Based on pressure propagation

[0097] The method for determining the connection relationship and edge weights between each node based on pressure propagation is as follows:

[0098] Collect the pressure test data between well sites (such as interference well testing, pressure build-up testing). Analyze the time series data of the pressure change. Conduct correlation analysis and time delay analysis on this time series data. Calculate the correlation coefficient of the pressure change between well sites. If the correlation coefficient exceeds a certain threshold, it is considered that there is a pressure propagation relationship between the well sites, and an edge is connected between them. Based on the correlation and time delay of the pressure propagation, calculate the weight. The higher the correlation and the shorter the time delay, the greater the weight.

[0099] If both reservoir connectivity and pressure propagation support the existence of an edge between two well site nodes, use weighted average to comprehensively use reservoir connectivity and pressure propagation to obtain the edge weight in the physical mechanism diagram:

[0100]

[0101] Where: is the edge weight between well site i and well site j in the physical mechanism diagram; is the edge weight obtained based on reservoir connectivity; α is the weight coefficient based on reservoir connectivity; is the edge weight obtained based on pressure propagation; β is the weight coefficient based on pressure propagation.

[0102] 4. Construct a spatial geometric graph

[0103] The nodes of the spatial geometric graph of the gas storage injection-production well field are set as each injection-production well field. The connection relationship between the nodes and the weights of the edges depend on the spatial distance between the well fields. When constructing the spatial geometric graph, obtain the coordinates (longitude and latitude) of each well field, and use the coordinates of the well fields to calculate the Euclidean distance between the well fields:

[0104]

[0105] In the formula: (x i , y i ) and (x j , y j ) are the coordinates of well field i and well field j respectively; D ij is the distance between well field i and well field j;

[0106] Set a distance threshold θ. When D ij < θ, there is an edge between well field i and well field j. In this embodiment, a sensitivity analysis is performed on the setting of the threshold θ to analyze the influence of different thresholds on the graph structure and determine the final threshold. Different weights are assigned according to the distance, and the closer the distance, the greater the weight. The Gaussian weight is selected. Then the edge weight in the spatial geometric graph is:

[0107]

[0108] Among them: is the edge weight between well field i and well field j in the spatial geometric graph; σ is a parameter that controls the weight decay rate.

[0109] 5. Construct a data-driven graph

[0110] The nodes of the data-driven graph of the gas storage injection-production well field are set as each injection-production well field, and the connection relationship between the wells and their weight coefficients are constructed using the statistical laws of the data itself. Based on the historical data of the wells (storage capacity and formation pressure), an edge is established according to the cosine similarity:

[0111]

[0112] Among them: S ij is the edge between well field i and well field j in the data-driven graph; x i and x j are the storage capacity time series data of well field i and well field j respectively;

[0113] S ijThe larger it is, the more similar well site i and well site j are. Using the K-nearest neighbor strategy, each well site is only connected to the K most similar well sites, and the normalized cosine similarity is used as the edge weight:

[0114]

[0115] Where: is the edge weight between well site i and well site j in the data-driven graph; S min is the minimum value of all S ij ; S max is the maximum value of all S ij .

[0116] 5. Perform physical-space-data multi-graph fusion based on the attention mechanism

[0117] Use the attention mechanism to achieve edge-level graph fusion. First, calculate the edge attention of each graph. Each edge has different weights in the physical mechanism graph, spatial geometry graph, and data-driven graph. Use a learnable MLP to calculate the attention of each edge in the physical mechanism graph, spatial geometry graph, and data-driven graph:

[0118]

[0119] Where: α k (i,j) is the edge weight between well site i and well site j in the k-th graph, k = 1, 2, 3, representing the physical mechanism graph, spatial geometry graph, and data-driven graph respectively; A k is the weighted adjacency matrix of the k-th graph; f(A k (i,j),x i ,x j ) is a neural network, that is, a learnable MLP; x i and x j are node features;

[0120] The fused edge weight is:

[0121]

[0122] Where: A fusion (i,j) is the edge weight between well site i and well site j in the fused graph.

[0123] 6. Train the graph network model

[0124] After constructing the physical mechanism graph, spatial geometry graph, and data-driven graph of the gas storage injection-production well site, use the attention mechanism for multi-graph fusion, and use the MLP to train the attention weights α ij, the weighted adjacency matrices of the three graphs and the node features of the three graphs are used as the inputs of the MLP. After training, the attention weights of the three graphs are obtained. Then, the fused edge weights are calculated based on the attention weights, and finally the fused graph is obtained. The fused graph is input into a graph neural network (GNN) model, which includes a graph convolutional network (GCN), a non-linear layer, and a fully connected layer. The model uses an Adam optimizer and an MSE loss function, and the initial learning rate is 0.001. The structure of the graph network model is as Figure 2 shown.

[0125] 7. Parameter Prediction

[0126] After the distributed node parameter prediction model is trained, the characteristic parameters associated with the prediction parameters of the distributed nodes are collected. The collected characteristic parameters are input into the distributed node parameter prediction model to change the node characteristic values, that is, the storage capacity of each well site. Then, the prediction parameters can be predicted through the graph network model, that is, the formation pressure of each well site can be obtained.

[0127] 8. Technical Effects

[0128] In this embodiment, a graph network model with spatial-data-physical multi-graph fusion is used to dynamically predict the formation pressure of the gas storage well sites. The physical mechanism graph, spatial geometry graph, and data-driven graph of each well site of the gas storage are constructed. When composing the graphs, various information such as reservoir connectivity, pressure propagation, spatial distance, and feature cosine similarity is included, which more comprehensively describes the complex characteristics of the gas storage. Moreover, the model has higher interpretability. Through multi-graph fusion, the model can better capture the complex non-linear relationships in the gas storage, thereby improving the prediction accuracy. And using the graph network model with multi-graph fusion to predict the formation pressure can adapt to diverse application scenarios. The data-driven graph can reflect the dynamic changes of the reservoir capacity data and is suitable for dynamic prediction. The spatial geometry graph and the physical mechanism graph can capture the static characteristics of the gas storage and are suitable for long-term prediction.

[0129] The above-described embodiments are only the preferred embodiments given to fully illustrate the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are all within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A distributed node parameter prediction method based on physical-spatial-data multi-graph fusion, characterized by: The steps include: Step 1: Create a dataset Collect characteristic parameters associated with the distributed nodes and the prediction parameters, the characteristic parameters including the production data of each node and the production environment data of the distributed nodes; perform theoretical calculations on the prediction parameters to construct a data set; Step 2: Build a graph for distributed nodes Based on the physical relationship between the nodes in the distributed nodes, the connection relationship and edge weight between the nodes are determined, and the physical mechanism diagram of the distributed nodes is constructed; Based on the spatial distance between each node in the distributed nodes, the connection relationship and edge weight between each node are determined, and the spatial geometric graph of the distributed nodes is constructed; Determine the connection relationship and edge weight between nodes based on the data statistics of distributed nodes, and construct a data-driven graph of distributed nodes; Step 3: Multi-image fusion based on attention mechanism The attention of each edge in the physical mechanism graph, the spatial geometry graph, and the data-driven graph is calculated respectively, and the weights of the corresponding edges in the physical mechanism graph, the spatial geometry graph, and the data-driven graph are fused to obtain the fused edge weights and the fused graph; Step 4: Training the graph network model The fusion graph is input into the graph neural network model, and the graph neural network model is trained using the data set to obtain a distributed node parameter prediction model; Step 5: Parameter prediction The characteristic parameters associated with the distributed nodes and the prediction parameters are collected, and the collected characteristic parameters are input into the distributed node parameter prediction model to predict the prediction parameters.

2. The distributed node parameter prediction method based on physical-spatial-data multi-graph fusion according to claim 1 is characterized in that: The distributed nodes are well sites, and each well site in the well site serves as a node.

3. The distributed node parameter prediction method based on physical-spatial-data multi-graph fusion according to claim 2 is characterized by: The well site is a gas storage facility, and the prediction parameter is formation pressure.

4. The distributed node parameter prediction method based on physical-spatial-data multi-graph fusion according to claim 3 is characterized by: In the step 1, the production data of the well site includes the daily gas injection volume of a single well and the wellhead pressure; the production environment data of the gas storage reservoir is the well test data, including the oil pipe resistance coefficient, the well depth in the middle of the gas layer, the annual average temperature of the wellhead, the geothermal gradient, the critical temperature of natural gas, the critical pressure of natural gas, the inner diameter of the oil pipe, the relative density of natural gas, the flow area of ​​the gas layer, the permeability, the viscosity of natural gas and the thickness of the gas layer; For each well site, the bottom hole pressure is: Among them: pwf is the bottom hole pressure; pwh is the wellhead pressure; qin is the daily gas injection volume of a single well; s is the skin coefficient; λ is the tubing resistance coefficient; φ 2 is the inner diameter of the oil pipe; e is the natural logarithm; and: s=0.03415γ g H / (TZ) Where: γ g is the relative density of natural gas; H is the depth of the middle of the gas layer; The formation pressure is: Where: Pr is the formation pressure; A is the gas layer flow area; k is the permeability; μ is the natural gas viscosity; B g is the gas volume coefficient; L is the gas layer thickness; and: Where: T is the average temperature of the dynamic air column in the wellbore; Z is the average deviation coefficient of the dynamic air column in the wellbore.

5. The distributed node parameter prediction method based on physical-spatial-data multi-graph fusion according to claim 3 is characterized by: In the step 2, when constructing the physical mechanism diagram, the physical relationship includes reservoir connectivity and pressure propagation; The method for determining the connection relationship and edge weight between nodes based on reservoir connectivity is as follows: using logging data to construct a structural model and an attribute model of the reservoir, and constructing a three-dimensional geological model of the reservoir; according to the geological model, the reservoir is divided into different units or regions and a connectivity analysis is performed: if two well sites are located in the same reservoir unit, it is considered that the two well sites are geologically connected, and an edge is connected between the two well sites, and the weight of the edge is calculated based on the permeability and porosity of the reservoir. The higher the permeability, the stronger the connectivity and the greater the weight; if there is a fluid flow path between the two well sites, it is considered that the two well sites are fluidically connected, and an edge is connected between the two well sites, and the weight of the edge is calculated based on the flow rate. The higher the flow rate, the stronger the connectivity and the greater the weight; The method for determining the connection relationship and edge weight between nodes based on pressure propagation is: analyzing the time series data of pressure changes based on the pressure test data between well sites; Perform correlation analysis and time delay analysis on the time series data to calculate the correlation coefficient of pressure changes between well sites. If the correlation coefficient between two well sites exceeds the set threshold, it is considered that there is a pressure propagation relationship between the two well sites, and an edge is connected between the two well sites. The weight is calculated based on the correlation and time delay of pressure propagation. The higher the correlation and the shorter the time delay, the greater the weight. The edge weights in the physical mechanism graph are obtained by using a weighted average to comprehensively use reservoir connectivity and pressure propagation: in: is the edge weight between well site i and well site j in the physical mechanism graph; is the edge weight obtained based on reservoir connectivity; α is the weight coefficient based on reservoir connectivity; is the edge weight obtained based on pressure propagation; β is the weight coefficient based on pressure propagation.

6. The distributed node parameter prediction method based on physical-spatial-data multi-graph fusion according to claim 3 is characterized by: In step 2, when constructing the spatial geometric graph, the coordinates of each well site are obtained, and the Euclidean distance between the well sites is calculated using the coordinates of the well sites: Where: (x i ,y i ) and (x j ,y j ) are the coordinates of well site i and well site j respectively; D ij is the distance between well site i and well site j; Set the distance threshold θ, when D ij When <θ, there is an edge between well site i and well site j. Different weights are assigned according to the distance. The closer the distance, the greater the weight. The edge weight in the spatial geometric graph is: in: The edge weight between well site i and well site j in the spatial geometric graph; σ is a parameter that controls the weight decay speed.

7. The distributed node parameter prediction method based on physical-spatial-data multi-graph fusion according to claim 3 is characterized by: In step 2, when constructing a data-driven graph, based on the historical data of the well site, edges are established according to cosine similarity: Where: S ij is the edge between well site i and well site j in the data driven graph; x i and x j are the storage capacity time series data of well site i and well site j respectively; S ij The larger it is, the more similar well site i is to well site j. Using the K nearest neighbor strategy, each well site is only connected to the most similar K well sites, and the normalized cosine similarity is used as the edge weight: in: is the edge weight between well site i and well site j in the data driven graph; S min For all S ij The minimum value in S max For all S ij The maximum value in .

8. The distributed node parameter prediction method based on physical-spatial-data multi-graph fusion according to claim 1 is characterized by: In step 3, the weight of each edge in the physical mechanism graph, the spatial geometry graph, and the data-driven graph is different, and a learnable MLP is used to calculate the attention of each edge in the physical mechanism graph, the spatial geometry graph, and the data-driven graph: Where: α k (i, j) is the edge weight between well site i and well site j in the kth graph, k = 1, 2, 3, representing the physical mechanism graph, spatial geometry graph, and data-driven graph, respectively; A k is the weighted adjacency matrix of the kth graph; f(A k (i,j),x i ,x j ) is a neural network, i.e. a learnable MLP; x i and x j is the node feature; The edge weight after fusion is: Among them: A fusion (i,j) is the edge weight between well site i and well site j in the fusion graph.

9. The distributed node parameter prediction method based on physical-spatial-data multi-graph fusion according to claim 1 is characterized by: In step 4, the graph neural network model includes a graph convolutional network, a nonlinear layer and a fully connected layer; in the process of training the graph neural network model using the data set, the Adam optimizer and MSE loss function are used, and the initial learning rate is set to 0.001.

Images