Casing damage prediction method based on rock mechanics theory fused with multilayer perceptron and graph convolutional neural network

Abstract

The invention discloses a casing damage prediction method based on a rock mechanics theory and fusing a multi-layer perceptron and a graph convolutional neural network, and the method comprises the steps: firstly, constructing a well position data attribute graph to describe a complex connection relation between wells, and secondly, designing a ground stress data enhancement module based on the rock mechanics theory, the data layering module is used for mapping oil field single well production data based on MLP; and the graph convolution module is used for stacking a multi-layer network to obtain high-order neighbor information. According to the scheme, the ground stress information is used as the characteristic attribute, mapping from single well production data to horizon production data is achieved through MLP, correction of the well node adjacency relation is achieved through geographical weighting, and the graph convolutional neural network is used for extracting the inter-well incidence relation, so that the aim of improving the casing damage prediction accuracy and reliability is achieved.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of data processing based on a specific computing model, and in particular to a casing damage prediction method based on rock mechanics theory and integrating a multi-layer perceptron and a graph convolutional neural network. Background Art

[0002] Casing damage (abbreviated as casing damage) refers to the gradual change in the structure of oil and water wells due to the combined effects of multiple factors such as geology, engineering and development during the long-term exploitation of oil fields. As the exploitation of major oil fields around the world enters the middle and late stages, the number of casing damage wells increases year by year, causing serious economic losses. The detection and prevention of casing damage has become one of the key issues to be urgently addressed in today's oil and gas development. By analyzing the on-site production data of oil fields, scientific and accurate predictions of casing damage are made, which is of great significance for seeking economical and efficient production methods, reasonable maintenance cycles and maintenance strategies, as well as extending the service life of casing and improving the economic benefits of oil fields.

[0003] Traditional casing damage evaluation and prediction methods start from a qualitative and mechanical perspective, study the influence of casing damage factors, conduct quantitative research on casing damage phenomena to establish casing damage mechanical models, and the main methods include analytical methods, numerical analysis methods, and measurement methods. Because the various factors that affect casing damage have characteristics such as nonlinearity, instability, and time-varying, it is difficult for the model to truly reflect the comprehensive impact mechanism of casing damage factors, and the cost is high and the model lacks general applicability. With the development of information technology, the oil industry has attached increasing importance to the development and utilization of data. As the shortcomings and limitations of traditional technologies in solving oil problems are becoming increasingly significant, machine learning technology can expand new methods and ideas for oilfield exploration and development.

[0004] At present, there have been many studies on machine learning methods in casing damage prediction. The use of BP neural network, SVM, XGBoost and other algorithms can obtain relatively accurate casing damage prediction results, but the generalization ability of the model is slightly insufficient. In addition, the existing research on casing damage prediction methods based on machine learning mostly takes single wells as the research object, lacks relevant information and methods for the analysis of casing damage influencing factors in the layer dimension, and does not refine the impact of layer factors in casing damage prediction research, which reduces the prediction accuracy and the actual application value. Therefore, there are certain limitations in the use of traditional machine learning technology for casing damage prediction research. Summary of the invention

[0005] In order to overcome the above technical defects, a casing damage prediction method based on rock mechanics theory and integrating multi-layer perceptron and graph convolutional neural network is provided in the embodiment of the present application. The specific scheme is as follows:

[0006] First, data enhancement is performed based on the formation stress information well layer static data in oil and water wells to increase the attribute scale and feature dimension diversity of well nodes;

[0007] Formation stress information includes rock elasticity-related parameters and vertical principal stress of the formation, and its related parameters include: shear wave time difference, longitudinal wave time difference, rock Poisson's ratio, Young's modulus, shear modulus, pore pressure, uniaxial tensile strength and vertical principal stress;

[0008] Secondly, a multi-layer perceptron (MLP) is used to form a data mapping layer. The single well production data, well layer static data and formation stress information are input into the corresponding data mapping layer. The single well production data is mapped to the production data of each layer, forming a well layer data attribute map of all layers in the block, thereby retaining the independent characteristics of each layer and better capturing the different influencing factors of each layer, thereby improving the prediction accuracy.

[0009] Finally, multiple graph convolutional neural networks (GCNs) layers are used to aggregate the information of the well layer data attribute graph and capture the multi-level features between well nodes. The extracted features are input into the fully connected layer and classified by the Softmax function to obtain the prediction results, making the final prediction results more accurate.

[0010] Optionally, in a possible implementation of the embodiment of the present application, in S1, data enhancement is performed according to the formation stress information well layer static data in the oil and water wells, including:

[0011] S1.1, Construction of spatial topology graph: Based on the well layer static data, production data and layer stress attribute data of oil and water wells, multiple weighted attribute graphs containing oil production well nodes and water injection well nodes of each layer are constructed, and the relationship between wells in the block is modeled to obtain the weighted adjacency matrix of the attribute graph;

[0012] S1.2. Calculation of geographical weighting coefficient: First, the Gaussian kernel function is used to perform a weighted operation on the weighted adjacency matrix based on the Euclidean distance between wells, and then the small layer sedimentary phase is introduced to perform geographical weighting correction on the correlation between wells. The small layer sedimentary phases include four types of sedimentary phases: river channel, intra-surface, extra-surface and pinch-out.

[0013] Optionally, in a possible implementation of the embodiment of the present application, in S1.1, the weighted attribute graph G is H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] ,in H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] represents a finite set of all types of well nodes, and H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] = H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] , H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] , H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] are the number of oil production wells and water injection wells, H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] Representatives use the information of stratum sedimentary facies and well locations to determine the connection between production and injection wells. H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] represents a weighted adjacency matrix;

[0014] If the node H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] and H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] If there is an edge between H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] represents the edge weight, otherwise H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] , weighted adjacency matrix H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] for:

[0015] H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] .

[0016] Optionally, in a possible implementation of the embodiment of the present application, in S1.2, the distance weighting coefficient in the distance weighting coefficient matrix is ​​obtained based on the weighting operation of the Euclidean distance between wells. for: , Well and well The Euclidean distance of is the bandwidth parameter, which is the maximum distance between the central well and the block boundary. The central well is the well at the center of the block in the horizontal and vertical directions.

[0017] Optionally, in a possible implementation of the embodiment of the present application, the weight coefficient after the geographical weighting correction, that is, the geographical weighting weight coefficient for:

[0018] , where Well and well The geographically weighted weight coefficient between is the permeability corresponding to the channel, internal surface, external surface and pinch-out sedimentary facies, The channel, internal surface, external surface and pinch-out sedimentary facies are connected to the well. and If the sedimentary facies between the connected wells has only one type of pinch-out, then .

[0019] Optionally, in a possible implementation of the embodiment of the present application, in S2, the mapping relationship between the single well production data and the production data of each layer is:

[0020] , where Indicates the layer obtained after processing by the data mapping module Output feature data of is a nonlinear activation function; Represents the production data characteristics of a single well; Indicates the index of well production data, which refers to a certain type of data in the well production data; Represents production data from a single well Middle layer Weights for feature mapping.

[0021] Optionally, in a possible implementation of the embodiment of the present application, the expression of the layer-by-layer information transmission mechanism for realizing information aggregation of the well layer data attribute graph is:

[0022] H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] , where H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] yes H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] The degree matrix of H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] That is H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] Perform symmetric normalization operation, H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] Represents the adjacency matrix of a node after self-connection to obtain its own information, H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] for H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] Layer trainable weight matrices, H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] express H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] The embedding value after layer activation, H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] is the feature dimension, and H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] , H = [ h 1 , 1 h 1 , 2 ⋯ h 1 , n h 2 , 1 h 2 , 2 ⋯ h 2 , n ⋮ ⋮ ⋯ ⋮ h n , 1 h n 1 ⋯ h n , n ] is the activation function.

[0023] The above technical solution can achieve the following technical effects in the embodiments of the present application:

[0024] Four technical means were used in combination: using geostress information as a characteristic attribute, using MLP to map single well production data to layer production data, correcting the well node adjacency relationship through geographic weighting, and using graph convolutional neural network to extract the relationship between wells to improve the accuracy and reliability of casing damage prediction. Specifically:

[0025] 1) Application of geostress information as characteristic attributes: Acoustic logging data (such as transverse and longitudinal wave velocities, density, etc.) are used to calculate geostress-related parameters, including vertical principal geostress, rock Poisson’s ratio and shear modulus, etc. These parameters increase the diversity of node attributes and can effectively increase the sensitivity of the model to inter-well geological information, thereby improving prediction accuracy.

[0026] 2) Use MLP to map single-well production data to layer production data: Use multi-layer perceptron (MLP) to map single-well production data to production data for each layer, retaining the independent characteristics of each layer, so that the model can better capture the different influencing factors of each layer, thereby improving the prediction accuracy.

[0027] 3) Correction of well node adjacency relationship through geographic weighting: First, the physical distance between wells is weighted using the Gaussian kernel function, and then the weight is corrected based on the information of the small layer sedimentary phase, which fully reflects the heterogeneity of geological conditions and enables the model to more accurately capture the spatial correlation between wells, thereby improving the reliability of the prediction.

[0028] 4) Apply graph convolutional neural networks to extract the correlation between wells: Through the information aggregation of multiple graph convolutional neural networks (GCNs) layers, the multi-level correlation information between well nodes is captured. At the same time, a residual network is introduced to alleviate the over-smoothing problem that may be caused by deepening the network, making the final prediction results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings exemplarily illustrate the embodiments and constitute a part of the specification, and together with the text description of the specification, are used to explain the exemplary implementation of the embodiments. The embodiments shown are for illustrative purposes only and do not limit the scope of the claims. In all drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0030] Figure 1 A model structure diagram of the GSMGN model in this application;

[0031] Figure 2 A schematic diagram of a process of a casing damage prediction method based on rock mechanics theory integrating a multi-layer perceptron and a graph convolutional neural network in this application;

[0032] Figure 3 A schematic diagram for constructing the weighted adjacency matrix in this application;

[0033] Figure 4 A schematic diagram of the structure of the MLP layer data stratification module in this application;

[0034] Figure 5 A structural diagram of the graph convolution module in this application;

[0035] Figure 6 This is a comparison chart of the K-fold cross-validation accuracy in this application;

[0036] Figure 7 This is a comparison chart of the K-fold cross-validation F1 value in this application. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0038] It should be noted that the descriptions involving "first", "second", etc. in the embodiments of the present application are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0039] In the description of the present application, it should be understood that the numerical labels before the steps do not indicate the order in which the steps are executed, but are only used to facilitate the description of the present application and to distinguish each step, and therefore should not be understood as a limitation on the present application.

[0040] In the design of oilfield exploration and production plans, the application research of geostress has become an indispensable part of basic research. Usually, the characteristics of the rock itself (Poisson's ratio, shear modulus and fracture pressure, etc.) are the internal causes of casing damage, and the change of geostress field is the external cause. Analyzing the relationship between casing damage and geostress helps to identify potential risk areas for casing damage. In addition, the use of logging data can efficiently obtain the formation stress profile, which can save the high cost of geostress measurement when conducting theoretical research on rock mechanics. Therefore, analyzing the rock strain characteristics of the casing damage layer through logging data is of great significance to the research on layer casing damage prediction.

[0041] With the continuous development and application of deep learning technology, its application in the oil field has gradually shown a widespread trend. With its powerful data processing and feature extraction capabilities, deep learning technology has achieved remarkable results in logging images, reservoir engineering, casing damage well prevention and control, etc. Therefore, how to fuse multi-source data and extract feature parameters, efficiently combine deep learning methods with oil and water well casing damage prediction, and improve the accuracy of casing damage prediction has important theoretical significance and practical value. Graph Convolutional Networks (GCNs) take nodes and edges in the graph structure as input, reflect the characteristics of the graph with Laplace matrix, use graph Fourier transform and graph convolution and other methods to capture the relationship and local features between nodes, and integrate the global feature information of nodes in the graph through a multi-layer GCN model, which can provide more comprehensive feature representation for tasks such as node classification, graph classification and link prediction. Therefore, graph convolutional neural networks can effectively process the complex geostress and spatial characteristics in oilfield production data, and provide new solutions and ideas for casing damage prediction research.

[0042] In summary, the dynamic and static production data of the production and injection wells of a certain oil production plant in Daqing Oilfield are analyzed. Based on the rock mechanics theory, this application proposes a casing damage prediction method based on the fusion of multi-layer perceptron and graph convolutional neural network (Geostatic-Stress Mlp-GCNNetwork, GSMGN): First, the rock mechanical properties are analyzed according to the logging data to extract the formation stress information; secondly, the multi-layer perceptron is combined with the formation static data to realize the conversion of single well production data to formation production data, and then the well formation data attribute map is generated according to the well connection relationship, well location coordinates, sedimentary phase and formation stress information; thirdly, based on the graph convolutional neural network, high-order neighborhood information is captured, spatial features are extracted and graphical modeling is completed, and the formation casing damage prediction results are output. Finally, the casing damage prediction problem is converted into a node classification problem, and the measurement indicators of similar models are compared and analyzed to verify the effectiveness of the method proposed in this application.

[0043] Secondly, in order to facilitate those skilled in the art to understand the technical solutions provided in the embodiments of the present application, the relevant technologies are described below:

[0044] In order to make full use of the spatial characteristics of the geographical distribution of wells in the block, the interconnection relationship between wells and the well's own production data in the problem of casing damage prediction of oil and water wells; the geological influencing factors involved in various main ground stresses in the formation; this paper introduces a graph structure to describe non-Euclidean spatial data such as the interconnection relationship between wells in the block, the geographical location information of the wells and the ground stress field (the sum of the stress state of each well in the spatial distribution). An attribute map of the well layer data in the block is established, and the GSMGN model is proposed. Figure 1As shown in the figure, the GSMGN model consists of a geostress data enhancement module, an MLP layer data mapping module, a graph convolution module, and a fully connected layer. The geostress data enhancement module calculates the rock mechanics and formation principal stress related parameters in the single well layer dimension through logging data; the MLP layer data mapping module is composed of multiple sub-modules with the same structure, and the number of sub-modules is adjusted according to the number of single well production data types; the graph convolution module is composed of multiple GCNs layers, and the inter-layer connection method uses a residual network for connection. The following will describe each part of the model in detail.

[0045] See also Figure 2 In the embodiment of the present application, the casing damage prediction method based on rock mechanics theory integrating multi-layer perceptron and graph convolutional neural network includes: S1. Data enhancement is performed based on the formation stress information and well layer static data in oil and water wells to increase the attribute scale and feature dimension diversity of the well nodes.

[0046] Specifically, the formation stress information includes rock elastic mechanics related parameters and formation vertical principal stress, and its related parameters include: Young's modulus, shear modulus, pore pressure, formation skeleton density, uniaxial tensile strength and mud content.

[0047] Specifically, the calculation of the above-mentioned related parameters is shown in formula (5) to formula (12): (5), (6), (7), (8), (9), (10), (11), (12); in, and is the time difference between transverse and longitudinal waves, is the formation density, is the rock Poisson's ratio, is Young's modulus, is the shear modulus, is the pore pressure, is the stratum skeleton density, , is the uniaxial tensile strength, is the mud content, is the relative value of natural gamma, is the empirical coefficient related to the stratigraphic age, and the new stratigraphic age is , old strata , is the vertical principal stress, For the depth of the well, is the depth of the stratum, It is a function of the density of the formation as it changes with its depth.

[0048] The S1 technical solution can bring the following benefits: Use acoustic logging data (such as transverse and longitudinal wave velocity, density, etc.) to calculate geostress related parameters, including vertical principal geostress, rock Poisson's ratio and shear modulus and other geostress information as characteristic attributes. These parameters increase the diversity of node attributes, which can effectively increase the sensitivity of the model to inter-well geological information, thereby improving prediction accuracy.

[0049] Data enhancement refers to the use of a standardized data enhancement process, combining rock mechanics theory to build a geostress data enhancement module to expand the existing common logging data such as transverse and longitudinal wave velocity and density, calculate rock elastic mechanics related parameters and vertical principal stress of the formation as the main features in casing damage prediction research, and align them with the static data of each well position, thereby improving the performance of the GSMGN model.

[0050] Specifically, the specific steps of data enhancement in step S1 include: S1.1, Construction of spatial topology map: Based on the static data of well layers, production data and layer stress attribute data of oil and water wells, multiple weighted attribute graphs containing oil production well nodes and water injection well nodes in each layer are constructed, and the relationship between wells in the block is modeled to obtain the weighted adjacency matrix of the attribute graph.

[0051] The constructed spatial topology is as follows Figure 3 As shown, the inter-well relationship represents the various characteristic values ​​and connection relationships of multiple wells in the observed area.

[0052] The weighted attribute graph is denoted by G, that is, ,in, represents a finite set of all types of well nodes, and , , are the number of oil production wells and water injection wells, Represents the edge between well nodes, that is, the connection relationship between production and injection wells is determined by using the layer sedimentary facies and well location information. Represents a weighted adjacency matrix, if the node and If there is an edge between represents the edge weight, otherwise , weighted adjacency matrix As shown in formula (1): (1).

[0053] The edge weight represented needs to go through two steps of "distance weighting" and "geographic weighting", so we also call it "geographically weighted weight".

[0054] S1.2. Calculation of geographical weighting coefficient: In summary, the calculation of the geographic weighted weight coefficient is to first use the Gaussian kernel function to perform a weighted operation on the weighted adjacency matrix based on the Euclidean distance between wells, and then introduce small-layer sedimentary phases to correct the correlation between wells. The small-layer sedimentary phases include four types of sedimentary phases: river channel, intra-surface, extra-surface and pinch-out.

[0055] The similarity between nodes is usually used as the basis for constructing edges. Under conventional geological conditions, the closer the distance between two wells, the easier it is for the wells to affect each other. In order to capture the complex structure of the data more efficiently, the Gaussian kernel function is used to perform a weighted operation based on the Euclidean distance between wells to obtain a distance weighted coefficient matrix.

[0056] The well location coordinate vector data format is and ,but The calculation formula is shown in formula (2)-formula (3): (2), (3), where It's a well and well The distance weighting coefficient between nodes, Well and well The Euclidean distance of the block is selected as the central well in the horizontal and vertical directions. is the bandwidth parameter, which is the maximum distance between the central well and the block boundary.

[0057] Considering only the physical distance between wells cannot fully reflect the strength of the correlation between wells. Therefore, this application introduces small-layer sedimentary phases to further correct and supplement the correlation between wells, namely, geographical weighted correction. Small-layer sedimentary phases reflect the sedimentary environment and lithological characteristics of the strata. The analysis of small-layer sedimentary phases can provide important indications for the judgment of the correlation between wells.

[0058] Due to the large number of sedimentary facies, based on the actual situation on site, this paper only retains four types of sedimentary facies: river channel, surface interior, surface exterior, and pinch-out. (referred to as geographical weighted weight coefficient) The calculation formula is shown in formula (4): (4), where Well and well The geographically weighted weight coefficient between It is the permeability corresponding to the channel, intra-surface, extra-surface and pinch-out sedimentary phases. The channel, internal surface, external surface and pinch-out sedimentary facies are connected to the well. and If the sedimentary facies between the connected wells has only one type of pinch-out, then .

[0059] S2. Use the multi-layer perceptron MLP to form a data mapping layer, input the single well production data, well layer static data and formation stress information into the corresponding data mapping layer, map the single well production data into the production data of each layer, and form a well layer data attribute map for all layers in the block.

[0060] Traditional casing damage prediction schemes use data from the entire well to establish a machine learning-based casing damage prediction model to predict the casing status of a single well. The data used is relatively rough and cannot be located at a specific layer. If the layer casing damage analysis is performed without processing, it will have a significant impact on the prediction results.

[0061] Therefore, this paper constructs a data mapping module based on a multi-layer perceptron (MLP), which uses existing layer data to map the production data of a single well into the production data of each layer of the well. This module is composed of multiple identical sub-modules stacked together. The number of sub-modules is dynamically adjusted according to the number of well production data types, and contains multiple data mapping layers with the same structure. The detailed structure is as follows: Figure 4 shown.

[0062] The input of a single submodule includes: single-type production data of the whole well and static data such as porosity, permeability, effective thickness and rock mechanical parameters of each layer. The final mapping is to the production data of the corresponding layer through formula (13).

[0063] (13) In the formula, Indicates the layer obtained after processing by the data mapping module Output feature data of is a nonlinear activation function; Represents the production data characteristics of a single well; Indicates the index of well production data, which refers to a certain type of data in the well production data; Represents production data from a single well Middle layer Weights for feature mapping.

[0064] The data of the layer static data, rock mechanical parameters, etc. are numbered with the layer to which they belong. They are aligned one by one with the data mapping layers in the submodule according to the numbers, and the characteristic items in the single well production data are input into each mapping layer in sequence. The data stratification module and the geostress data enhancement module are sorted according to the layer number and aligned with the dimension of the weighted adjacency matrix corrected in S1.2 above to form the layer data attribute map of all layers in the study block.

[0065] The technical solution of S2 can bring the following benefits: using multi-layer perceptron (MLP) to map single well production data to production data of each layer, realizing the mapping of single well production data to layer production data, which can effectively retain the independent characteristics of each layer, so that the model can better capture the different influencing factors of each layer, thereby improving the prediction accuracy.

[0066] S3. Through multiple graph convolutional neural network (GCNs) layers, information aggregation of well layer data attribute graph is realized to capture multi-level features between well nodes. The extracted features are input into the fully connected layer and classified through the Softmax function to obtain the prediction results.

[0067] The calculation of graph convolutional neural network is an iterative process in which nodes continuously collect all neighbor and self-feature information, and then update themselves after weighted summation. The multi-layer graph convolutional network realizes the information aggregation of well layer data attribute graph through a layer-by-layer information transmission mechanism. Its transmission mechanism is shown in formula (14): (14) In the formula, yes The degree matrix of That is Perform symmetric normalization operation, Represents the adjacency matrix of a node after self-connection to obtain its own information, for Layer trainable weight matrices, express The embedding value after layer activation, is the feature dimension, and , is the activation function.

[0068] The technical solution of S3 can bring the following benefits: by aggregating information through multiple graph convolutional neural network (GCNs) layers, it can capture the multi-level correlation information between well nodes, and at the same time introduce the residual network to realize the application of graph convolutional neural network to extract the correlation relationship between wells, so as to alleviate the over-smoothing problem that may be caused by deepening the network, making the final prediction results more accurate.

[0069] In order to efficiently extract features from oil and water well production data, ground stress related data, well topology information within the block, and block global information, this paper constructs a multi-layer graph convolutional neural network, such as Figure 5 Select the image. As the model input, by stacking multiple graph convolution layers, the model can extract the high-order neighborhood information of the well node. At the same time, a residual module is added between the graph convolution layers to alleviate the over-smoothing problem caused by the deepening of the network. Finally, it is input into the fully connected layer and classified by the Softmax function, as shown in formula (15): (15), where is the predicted node category probability distribution, is the fully connected layer weight matrix, is the output of the last layer.

[0070] In the embodiments of the present application, we comprehensively applied four technical means: 1) applied geostress information as feature attributes, 2) used MLP to realize the mapping of single well production data to layer production data, 3) corrected the adjacency relationship of well nodes through geographic weighting, and 4) used graph convolutional neural network to extract the correlation relationship between wells to achieve the purpose of improving the accuracy and reliability of casing damage prediction.

[0071] Finally, in order to verify the effectiveness of the GSMGN model in the casing damage prediction task of production and injection wells, this paper selects a variety of GNN models as baseline models, and designs baseline model performance comparison experiments and ablation experiments to further verify the prediction performance of this model.

[0072] 1482 layers in 317 marked wells in a block of Daqing Oilfield were selected as samples, and the dynamic and static production data of oil and water wells were used as data sets to carry out casing damage mechanism analysis and method research. Among them, 63 wells were casing damaged wells and 254 were non-casing damaged wells. The number of casing damaged layers was 398, and the number of non-casing damaged layers was 1084.

[0073] Since the dynamic and static data of oil and water wells are complicated and highly professionally related, it is necessary to design a casing damage evaluation index system based on the characterization of casing damage influencing factors and dynamic and static data of oilfield development, and pre-screen the data. Finally, the dynamic data of water injection wells (such as intact years, cumulative production time, time rate, high-pressure water injection time, high-pressure water injection time rate, cumulative water injection volume, monthly water injection pressure, monthly water injection volume, water injection intensity, injection completion ratio, apparent water absorption index, etc.), dynamic data of oil production wells (such as cumulative liquid production, oil pressure, monthly oil production, monthly water production, liquid production intensity, water content, time rate, cumulative production time, intact years, shut-in frequency, etc.), static data of oilfield development (such as well-to-well connectivity, small layer sedimentary phase data, well location coordinate information, sandstone effective thickness, perforation effective thickness, etc.) and logging data (such as natural potential, natural gamma, acoustic wave time difference, resistivity and well diameter, etc.) in the dynamic data of oilfield development are selected to establish a layer casing damage prediction data set. In order to reduce the complexity of the model, the time series data in the oilfield production dynamic data, including the monthly data of oil wells (such as liquid production, water content, oil pressure, etc.) and the monthly data of water injection wells (such as monthly water injection volume, monthly water injection pressure, and apparent water absorption index), are subjected to dimensionality reduction processing.

[0074] In order to more comprehensively evaluate the model performance, the initial learning rate of the GSMGN model is set to 0.0005, the activation function is Relu, the Dropout is set to 0.5, the batch is set to 64, and the Adam training model is used for a maximum of 100 epochs. The other baseline models are set according to the optimal parameters of the experiment in the actual application scenario.

[0075] In order to accurately evaluate the prediction performance of the model in this paper, accuracy, precision, recall and F1 value are used as model performance evaluation indicators.

[0076] In the baseline model classification comparison experiment, SGC, MPNN, Graph CNN and GraphSAGE were selected as comparison models to verify the effect of the GSMGN proposed in this paper on the layer casing damage prediction data set. Since the baseline models generally do not have a data stratification module, in order to avoid affecting the verification of the experimental results, this paper unified the input of each baseline model into the layer casing damage prediction data set processed by the data stratification module. The accuracy values ​​of each experiment of the five-fold cross validation of all models were counted, and the results are shown in the figure. Figure 6 and Figure 7 As shown, the values ​​on the horizontal axis represent the experimental number of each five-fold cross validation.

[0077] Figure 6 and Figure 7The results show that the GSMGN model performs more stably and well in each experiment. The accuracy and F1 values ​​of each model are similarly arranged. The detailed experimental results of the GSMGN model and each baseline model are shown in Table 1, where the Acc, Prec, Rec and F1 values ​​of each model are the average results obtained after five cross-validations. Compared with all baseline models, the GSMGN model achieved the best results in each evaluation indicator.

[0078] According to the chart data, the GraphSAGE model performs relatively well, with an accuracy rate of 74.56%~76.05%. The reason may be that GraphSAGE enhances the flexibility of the model by random subgraph sampling, but loses some node information in the graph. The performance of MPNN and Graph-CNN is slightly inferior to GraphSAGE, among which the accuracy of the Graph-CNN model is 73.16%~74.86%, and the accuracy of the MPNN model is 71.24%~73.67%. The performance of the SGC model is poor, with an accuracy rate of 65.41%~66.87%. This may be due to the nonlinearity, uncertainty and time-varying characteristics of the factors affecting casing damage. SGC improves the computational efficiency by simplifying the graph convolution operation, removing the weight matrix and activation function, but loses the nonlinear feature learning ability in GCN. The highest accuracy and F1 value achieved by the GSMGN model are 86.72% and 84.52% respectively.

[0079] Table 1

[0080] The GSMGN model has the ability to efficiently complete casing damage prediction tasks, because the construction of the stratigraphic map can help the model better capture the relationship between wells and the global information within the block. In addition, the graph convolutional neural network performs spectral decomposition on the Laplace matrix of the graph, takes the nodes and edges in the graph structure as input, reflects the characteristics of the graph with the Laplace matrix, uses the graph Fourier transform to capture the relationship and local features between nodes, and counts the new features of the node as the weighted average of itself and its high-order neighbors, so that the information of the production and injection well nodes can be propagated throughout the stratigraphic map.

[0081] In summary, compared with all baseline models, the GSMGN model has certain advantages in casing loss prediction tasks.

[0082] In order to explore the contribution of each module in the model, this paper designs four variants of the GSMGN model and conducts comparative experiments with itself as follows: (1) GSMGN-SL: The GSMGN model without self-connection in the adjacency matrix.

[0083] (2) GSMGN-A: The adjacency matrix is ​​not modified using geographically weighted weight coefficients.

[0084] (3) MGN: Removes the geostress data enhancement module based on rock mechanics theory and ignores the potential geological connections within blocks and between wells.

[0085] (4) GSGN: Remove the data stratification module, ignore the difference between single well data and layer data of production and injection wells, and set the single well production data except layer static data as each layer layer data.

[0086] (5) GCN: remove all modules and use only the basic graph convolutional neural network.

[0087] The experimental results are shown in Table 2. Analysis of the data shows that (1) stripping off node self-loops will lead to a decline in model performance. This shows that the node's own attribute information has an improved effect on the prediction of casing damage in production and injection wells. (2) The lack of a geographically weighted adjacency matrix makes it difficult to feedback the connection strength between well nodes, resulting in a decline in model performance. (3) The lack of a geostress data module has led to a significant decline in the accuracy of the model, indicating that the potential geological correlation information in the stratum plays an important role in casing damage prediction. (4) Single well data cannot be regarded as the stratum data of each well and each layer in the block. If the data is not "splitting" ("splitting" means allocating the production data of the entire well to each layer), the model performance will decline due to the coarseness of the data.

[0088] Table 2

[0089] This paper constructs a well layer data attribute graph for oil and water well casing damage prediction. The oil wells and water injection wells are taken as independent nodes in the graph, and the edges and weights between the nodes are established by considering the relationship between the wells in the block. The casing damage prediction problem of oil and water wells is transformed into a node classification problem, and the concept of ground stress is introduced. A casing damage prediction method based on a multi-layer graph convolutional neural network is proposed. The common and easily available data and logging data of oil field production are used, combined with rock mechanics theory, to effectively supplement and improve the node attribute information. Experiments show that the proposed GSMGN model has good inter-well information extraction and information fusion capabilities, which verifies the effectiveness and rationality of the model. At present, many researchers try to use multi-channel feature extraction and attention mechanism to improve the performance of the model. On the basis of improving the model proposed in this paper, the application of attention mechanism will also be taken into consideration in the future.

[0090] Obviously, those skilled in the art should understand that the modules or steps of the above-mentioned embodiments of the present application can be implemented by general-purpose computer devices, they can be concentrated on a single computer device, or distributed on a network composed of multiple computer devices, optionally, they can be implemented by executable program codes of computer devices, so that they can be stored in a storage device and executed by the computer device, and in some cases, the steps shown or described can be executed in a different order from that herein, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present application are not limited to any specific combination of hardware and software.

[0091] It should be noted that the above are only preferred embodiments of the present application, and the patent protection scope of the present application is not limited thereto. Any equivalent structure or equivalent process transformation made using the contents of the specification and drawings of the present application, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A casing damage prediction method based on rock mechanics theory integrating multi-layer perceptron and graph convolutional neural network, characterized in that: include: S1. Perform data enhancement on well layer static data based on formation stress information in oil and water wells to increase the attribute scale and feature dimension diversity of well nodes; Formation stress information includes rock elasticity-related parameters and vertical principal stress of the formation, and its related parameters include: shear wave time difference, longitudinal wave time difference, rock Poisson's ratio, Young's modulus, shear modulus, pore pressure, uniaxial tensile strength and vertical principal stress; S2. Use a multi-layer perceptron MLP to form a data mapping layer, input single well production data, well layer static data and formation stress information into the corresponding data mapping layer, map the single well production data to the production data of each layer, and form a well layer data attribute map for all layers in the block; S3. Through multiple graph convolutional neural network (GCNs) layers, information aggregation of well layer data attribute graph is realized to capture multi-level features between well nodes. The extracted features are input into the fully connected layer and classified through the Softmax function to obtain the prediction results.

2. The casing damage prediction method according to claim 1, characterized in that: In S1, data enhancement based on formation stress information and well layer static data in oil and water wells includes: S1.1, Construction of spatial topology graph: Based on the well layer static data, production data and layer stress attribute data of oil and water wells, multiple weighted attribute graphs containing oil production well nodes and water injection well nodes of each layer are constructed, and the relationship between wells in the block is modeled to obtain the weighted adjacency matrix of the attribute graph; S1.

2. Calculation of geographical weighting coefficient: First, the Gaussian kernel function is used to perform a weighted operation on the weighted adjacency matrix based on the Euclidean distance between wells, and then the small layer sedimentary phase is introduced to perform geographical weighting correction on the correlation between wells. The small layer sedimentary phases include four types of sedimentary phases: river channel, intra-surface, extra-surface and pinch-out.

3. The casing damage prediction method according to claim 2, characterized in that: In S1.1, the weighted attribute graph G is ,in represents a finite set of all types of well nodes, and = , , are the number of oil production wells and water injection wells, Representatives use the information of stratum sedimentary facies and well locations to determine the connection between production and injection wells. represents a weighted adjacency matrix; If the node and If there is an edge between represents the edge weight, otherwise , weighted adjacency matrix for: 。 4. The casing damage prediction method according to claim 2, characterized in that: In S1.2, the distance weighting coefficient in the distance weighting coefficient matrix is ​​obtained based on the weighting operation of the Euclidean distance between wells. for: , Well and well The Euclidean distance of is the bandwidth parameter, which is the maximum distance between the central well and the block boundary. The central well is the well at the center of the block in the horizontal and vertical directions.

5. The casing damage prediction method according to claim 4, characterized in that: The weight coefficient after geographical weighting correction, that is, the geographical weighting weight coefficient for: , where Well and well The geographically weighted weight coefficient between is the permeability corresponding to the channel, internal surface, external surface and pinch-out sedimentary facies, The channel, internal surface, external surface and pinch-out sedimentary facies are connected to the well. and If the sedimentary facies between the connected wells has only one type of pinch-out, then .

6. The casing damage prediction method according to claim 1, characterized in that: In S2, the mapping relationship between single well production data and production data of each layer is: , where Indicates the layer obtained after processing by the data mapping module Output feature data of is a nonlinear activation function; Represents the production data characteristics of a single well; Indicates the index of well production data, which refers to a certain type of data in the well production data; Represents production data from a single well Middle layer Weights for feature mapping.

7. The casing damage prediction method according to claim 1, characterized in that: The expression of the layer-by-layer information transmission mechanism for realizing information aggregation of well layer data attribute graph is: , where yes The degree matrix of That is Perform symmetric normalization operation, Represents the adjacency matrix of a node after self-connection to obtain its own information, for Layer trainable weight matrices, express The embedding value after layer activation, is the feature dimension, and , is the activation function.

Images