An intelligent offshore oil and gas well network production prediction method and system

Abstract

The embodiment of the specification discloses an offshore oil and gas well network yield intelligent prediction method and system, wherein the intelligent prediction method comprises the following steps: acquiring time sequence data of gas production of each single well in the well network and time sequence data of production key parameters corresponding to each single well, and a static well-to-well relationship representing the relative distance between each single well in the well network; performing feature extraction based on the time sequence data of gas production of each single well and the time sequence data of production key parameters corresponding to each single well, to obtain a feature matrix corresponding to each single well; inputting the feature matrix corresponding to each single well into a trained prediction model to obtain a predicted gas production of the well network, wherein the trained prediction model comprises an adjacency matrix representing a well-to-well interference relationship. The problem of prediction inaccuracy of an existing data model for a complex well network system with well-to-well dynamic interference is solved, and accurate prediction of the productivity of the well network system is realized.

Description

TECHNICAL FIELD

[0001] Embodiments of the present specification relate to the technical field of offshore oil and gas field productivity prediction, and in particular to precision optimization of oil and gas well pattern productivity prediction. BACKGROUND

[0002] Due to the inherent complexity and dynamic characteristics of offshore oil and gas reservoirs, efficient development faces multiple challenges. First, offshore oil and gas reservoirs generally exhibit spatial heterogeneity and interwell dynamic interference phenomena, resulting in complex spatiotemporal dynamic characteristics of oil and gas well productivity evolution. Second, the variability and uncertainty factors of the marine environment, such as tides, currents, and temperature-salinity gradient changes, further increase the complexity of oil and gas reservoir dynamic behavior.

[0003] In recent years, with the rapid development of computer technology and artificial intelligence, intelligent prediction methods based on data-driven have attracted widespread attention. This kind of method uses historical production data of oil and gas fields to construct the mapping relationship between input and output through machine learning algorithms, and directly predicts future productivity dynamics. Among them, artificial intelligence technology represented by deep learning, with its powerful feature extraction and nonlinear mapping ability, has shown significant potential in the field of oil and gas productivity prediction.

[0004] However, the existing methods still have obvious limitations. The existing models are mainly limited to single-well scale modeling, lack effective expression of interwell spatial dependence, and are difficult to apply to complex well pattern systems with significant interwell interference effects, thus making it difficult to accurately predict the productivity of well pattern systems. SUMMARY

[0005] The embodiments of the present specification provide an offshore oil and gas well pattern production intelligent prediction method and system, which solves the prediction error problem of existing data models for complex well pattern systems with interwell dynamic interference phenomena, and realizes accurate prediction of well pattern system productivity.

[0006] The technical scheme is as follows:

[0007] In a first aspect, the embodiments of the present specification provide an offshore oil and gas well pattern production intelligent prediction method, comprising the following steps:

[0008] Obtaining time series data of gas production and time series data of production key parameters corresponding to each single well in the well pattern, and static interwell relationship representing the relative distance between each single well in the well pattern;

[0009] Performing feature extraction based on the time series data of gas production and the time series data of production key parameters corresponding to each single well, to obtain a feature matrix corresponding to each single well;

[0010] inputting the feature matrix corresponding to each single well into the trained prediction model to obtain the predicted gas production of the well pattern, the trained prediction model including an adjacency matrix representing the interwell interference relationship;

[0011] The trained prediction model includes a parameter matrix with initial values set based on the static interwell relationship, and the adjacency matrix is a learned parameter matrix corresponding to the trained prediction model.

[0012] As a preferred solution, the time series data of the gas production of each single well in the well pattern and the time series data of the production key parameter are obtained by:

[0013] The time series data of the gas production of each single well in the well pattern and the time series data of multiple production parameters are obtained.

[0014] Based on the similarity of the time series data of the gas production of each single well in the well pattern and the time series data of multiple production parameters, a production key parameter is selected from the multiple production parameters.

[0015] As a preferred solution, the production key parameter is selected from the multiple production parameters based on the similarity of the time series data of the gas production of each single well in the well pattern and the time series data of multiple production parameters, including:

[0016] Based on the time series data of the gas production of each single well in the well pattern and the time series data of multiple production parameters, similarity score ranking information of the multiple production parameters corresponding to each single well is obtained by dynamic time warping method.

[0017] Based on the similarity score ranking information of the multiple production parameters corresponding to each single well, a production key parameter is selected from the multiple production parameters.

[0018] As a preferred solution, the parameter matrix includes a first parameter matrix and a second parameter matrix representing the asymmetric interference between wells, and a third parameter matrix representing the characteristics of each single well in the well pattern.

[0019] As a preferred solution, the training method of the prediction model includes the following steps:

[0020] Setting the initial value of the parameter matrix based on the static interwell relationship;

[0021] Inputting the current parameter matrix and the feature matrix corresponding to each single well into the trained prediction model, updating the current parameter matrix based on the loss function of the trained prediction model, and repeating the step with the updated parameter matrix as the new current parameter matrix until the trained prediction model meets the convergence condition to obtain the trained prediction model.

[0022] As a preferred solution, the loss function based on the to-be-trained prediction model updates the current parameter matrix, including:

[0023] The loss function based on the to-be-trained prediction model preliminarily updates the current parameter matrix;

[0024] The weight matrix of the two parameter matrices respectively corresponding to before and after the update is obtained through the multi-head attention mechanism;

[0025] Based on the two parameter matrices respectively corresponding to before and after the update and the weight matrix thereof, secondary update is performed to obtain the updated parameter matrix.

[0026] As a preferred solution, the time series data of the gas production of each single well and the time series data of the production key parameters are extracted to obtain the feature matrix corresponding to each single well, including:

[0027] For any single well, local fine features and global trend features are extracted based on the time series data of the gas production and the time series data of the production key parameters corresponding thereto;

[0028] The feature matrix corresponding to the single well is obtained by feature fusion based on the local fine features and the global trend features;

[0029] The above steps are repeated until the feature matrix corresponding to each single well is obtained.

[0030] As a preferred solution, the local fine features and the global trend features are extracted based on the time series data of the gas production and the time series data of the production key parameters corresponding to any single well, including:

[0031] For any single well, local fine features are extracted by using a sliding window, and global trend features are extracted by using global average pooling based on the time series data of the gas production and the time series data of the production key parameters corresponding thereto.

[0032] As a preferred solution, the prediction method further includes:

[0033] The parameter matrix is output to obtain a visual inter-well relationship diagram.

[0034] In a second aspect, the embodiments of the present specification provide an offshore oil and gas well pattern production intelligent prediction system, including a data acquisition module, a feature extraction module and a productivity prediction module;

[0035] The data acquisition module acquires time series data of gas production and time series data of production key parameters corresponding to each single well in the well pattern, and a static inter-well relationship representing the relative distance between the single wells in the well pattern;

[0036] The feature extraction module extracts features based on the time series data of the gas production of each single well and the time series data of the production key parameters, to obtain a feature matrix corresponding to each single well;

[0037] The productivity prediction module inputs the feature matrix corresponding to each single well into a trained prediction model to obtain the predicted gas production of the well pattern, the trained prediction model including an adjacency matrix representing the interwell interference relationship; wherein the prediction model to be trained includes a parameter matrix with an initial value set based on the static interwell relationship, and the adjacency matrix is a learned parameter matrix corresponding to the trained prediction model.

[0038] In a third aspect, an electronic device is provided, including a processor and a memory; the processor is connected with the memory; the memory is used to store executable program codes; the processor runs a program corresponding to the executable program codes by reading the executable program codes stored in the memory, to execute the steps of the first aspect of the above-mentioned embodiments.

[0039] In a fourth aspect, a computer storage medium is provided, which stores a plurality of instructions, the instructions being adapted to be loaded by a processor and executed to perform the steps of the first aspect of the above-mentioned embodiments.

[0040] The technical solutions provided by some embodiments of the present specification have at least the following beneficial effects:

[0041] 1. Existing models are mainly limited to single-well scale modeling, lack effective expression of interwell spatial dependence, and are difficult to apply to complex well pattern systems with significant interwell interference effects, so it is difficult to accurately predict the productivity of well pattern systems. The present method proposes an innovative framework of adaptive spatio-temporal graph neural network. On the one hand, the graph network structure is introduced to explicitly model the well pattern topology, and a graph attention mechanism is designed to adaptively learn the dynamic correlation strength between wells. On the other hand, based on the graph network, spatio-temporal convolution and attention modules are integrated to extract trend and periodic patterns in time series data from a multi-scale perspective. The spatio-temporal dynamic characteristics of the well pattern system and the interwell interference effect in the development of marine gas reservoirs are fully considered, solving the prediction error problem of existing data models for complex well pattern systems with interwell dynamic interference, and realizing accurate prediction of the productivity of well pattern systems.

[0042] 2. Modeling and training based on the relative distance between each single well in the well pattern and the time series data of each single well can effectively guide the model to take the correct direction at the beginning of training, improving the training speed.

[0043] 3. The interpretable feature analysis based on dynamic time warping improves the identification ability of key features.

[0044] 4、 Introduce high and low frequency attention fusion mechanism, capture multi-scale spatio-temporal dynamic pattern in time series data, and significantly enhance the description ability of complex production dynamics.

[0045] 5、 The present application visualizes the interference between each single well in the well pattern system, and the visualized analysis results corresponding to different production stages can directly reveal the evolution law of interwell interference in different development stages. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0047] Figure 1 is a flowchart of an offshore oil and gas well pattern production intelligent prediction method provided by an embodiment of the present application;

[0048] Figure 2 is a comparison chart of prediction effects of the adaptive spatio-temporal graph neural network model and the existing data model in the test set in the embodiment of the present application;

[0049] Figure 3 is an interwell dynamic relationship evolution visualization result chart;

[0050] Figure 4 is a structural schematic diagram of an offshore oil and gas well pattern production intelligent prediction system provided by an embodiment of the present application;

[0051] Figure 5 is a structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application.

[0053] In the specification and claims of the present application and the above drawings, the terms "first", "second", "third" and the like are used to distinguish different objects, and are not used to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed or can optionally include other steps or units inherent to the process, method, product or device.

[0054] The following description provides examples, and is not intended to limit the scope, applicability or examples set forth in the claims. Alterations can be made in the elements' functions and arrangements described without departing from the scope of the present description. Various examples can omit, substitute, or add various procedures or components as appropriate. For instance, the methods described can be performed in an order different than described, and various steps can be added, omitted, or combined. Also, features described with respect to some examples can be combined in other examples.

[0055] The present application aims to solve the technical challenges faced by multi-well productivity prediction in the development of complex offshore oil and gas reservoirs. Traditional methods predict the predicted value of each single well by single-well timing prediction model and aggregate it as the total production prediction value of the well pattern, which is difficult to effectively model the dynamic evolution law of the well pattern system, so the prediction accuracy and adaptability are limited. Therefore, an intelligent method is needed that can fully tap the spatio-temporal dynamic characteristics contained in the well pattern production data and accurately predict the dynamic productivity of multi-wells, providing important support for scientific decision-making and efficient development of offshore gas fields.

[0056] Referring to Figure 1 As shown in the figure, Figure 1 A flowchart of an offshore oil and gas well pattern production intelligent prediction method provided by an embodiment of the present application can include at least the following steps:

[0057] Step 202, obtaining the time series data of the gas production of each single well in the well pattern and the time series data of the production key parameters, and the static interwell relationship representing the relative distance between each single well in the well pattern.

[0058] Illustratively, the gas production and production key parameters of each single well in the offshore oil and gas well pattern are continuously collected, and through data cleaning, normalization and other pretreatments, the time series data of the gas production and the time series data of the production key parameters corresponding to each single well are obtained.

[0059] Illustratively, the relative distance between each single well in the well pattern is determined according to the static interwell relationship of the well pattern, for example, the distance between any two wells is calculated from the coordinate values of each single well, and the distance matrix X of the well pattern is obtained after numbering each single well, the elements in the distance matrix X are the inverses of the distances between any two wells, the closer the distance, the stronger the physical correlation, and the higher the corresponding interference, which is used as the initial value of the parameter matrix. The initial value of other forms of parameter matrix set on the basis of the relative distance between each single well is within the protection scope of the present embodiment.

[0060] It should be noted that due to the complexity of reality, such as the existence of non-uniform geological structures such as cracks, faults, etc. in underground reservoirs, two wells that are physically close in distance may be isolated by faults, and far away wells may be connected through cracks, so the well interference relationship in the well pattern needs to be learned rather than being directly determined by the physical distance. The distance matrix X is used as the initial state of the well interference relationship, and then adjusted and corrected during the model training process (such as unmarked faults). Using the distance matrix X as the initial state of the well interference relationship provides a scientific "learning starting point" for the model, so that the model does not start from scratch and learn in the air. According to the distance relationship, a more accurate initial relationship sketch is drawn for it. The main task of this initial sketch is to guide the model to the correct direction in the early stage of training. Once the model training starts, the model will continuously learn and correct based on this sketch, forming a more accurate dynamic relationship network, i.e. the adjacency matrix, for prediction.

[0061] In an embodiment of the present specification, the time series data of the gas production of each single well in the well pattern corresponding to each single well and the time series data of the production key parameters are obtained, including:

[0062] Step 2022, obtaining the time series data of the gas production of each single well in the well pattern corresponding to each single well and the time series data of multiple production parameters;

[0063] Step 2024, selecting the production key parameters from the multiple production parameters based on the similarity of the time series data of the gas production of each single well in the well pattern corresponding to each single well and the time series data of multiple production parameters.

[0064] Explanatorily, by evaluating the correlation between each production parameter and the time series data of the gas production, the production key parameters with high similarity are selected, so that the parameters irrelevant to the gas production or with low correlation are excluded, and interference on the well pattern gas production prediction is avoided.

[0065] Illustratively, the production key parameters can include multiple production parameters, and a preset number of production parameters are selected as the production key parameters according to the correlation degree from large to small.

[0066] Exemplarily, the selected production key parameters are wellhead temperature, wellhead pressure and choke opening.

[0067] In an embodiment of the present specification, the production key parameters are selected from the multiple production parameters based on the similarity of the time series data of the gas production of each single well in the well pattern corresponding to each single well and the time series data of multiple production parameters, including:

[0068] Based on the time series data of the gas production of each single well in the well pattern corresponding to each single well and the time series data of multiple production parameters, the similarity score ranking information of the multiple production parameters corresponding to each single well is obtained by the dynamic time warping method;

[0069] The similarity score ranking information based on all the respective multiple production parameters of each single well is selected from the production key parameters among the multiple production parameters.

[0070] Illustratively, the dynamic time warping (DTW) method is a method for measuring the similarity of two time series, finding the most natural alignment between sequences, even if they are different in length, frequency, phase offset, and local region deformation. The dynamic time warping method is used to calculate the correlation degree between each respective production parameter and the time series data of gas production of each single well, and the calculation formula is:

[0071]

[0072] wherein, represents the time series data of the i-th production parameter, and y represents the time series data of the gas production. The correlation degree score of the corresponding production parameter is higher, indicating that the dynamic similarity between the production parameter and the gas production is stronger. The similarity score ranking information is arranged from high to low according to the correlation degree score of the production parameter, and the first few production parameters in the respective similarity score ranking information of each single well are selected, and the preset number of production parameters with the highest occurrence frequency are counted as the production key parameters.

[0073] Step 204, based on the time series data of the gas production and the time series data of the production key parameters of each single well, feature extraction is performed to obtain the respective feature matrix of each single well.

[0074] Illustratively, the present embodiment proposes an innovative framework of adaptive spatio-temporal graph neural network. The adaptive spatio-temporal graph neural network mainly includes a dynamic interwell relationship graph learning module and a spatio-temporal convolution representation module. The dynamic interwell relationship graph learning module explicitly models the well pattern topological relationship by introducing a graph network structure, and designs a graph attention mechanism to adaptively learn the dynamic correlation strength between wells. The spatio-temporal convolution representation module integrates spatio-temporal convolution and attention modules on the basis of the graph network, to extract the trend and periodic patterns in the time series data from a multi-scale perspective.

[0075] Illustratively, the spatio-temporal convolution representation module extracts time series features for each single well to obtain the respective feature matrix as the feature input for subsequent prediction of well pattern production.

[0076] Specifically, in one embodiment of the present specification, step 204 performs feature extraction based on the time series data of the gas production and the time series data of the production key parameters of each single well to obtain the respective feature matrix of each single well, including:

[0077] For any single well, based on the time series data of its corresponding gas production and the time series data of the production key parameters, local fine features and global trend features are extracted respectively;

[0078] Based on the local fine features and the global trend features, a feature matrix corresponding to the single well is obtained through feature fusion;

[0079] Repeat the above steps until the feature matrix corresponding to each single well is obtained.

[0080] Explanatorily, the spatio-temporal convolution representation module designs multi-scale convolution kernels and pooling operations to realize high-low frequency fusion design. By adopting parallel high-low frequency self-attention structure, local fine features (short span) and global trend features (long span) in time series are modeled respectively to depict the feature evolution law under different time spans, and then the adaptive fusion of high-low frequency features is realized through the fusion mechanism based on the gating unit.

[0081] Specifically, the fusion process is as follows:

[0082]

[0083]

[0084] wherein, is a sigmoid activation function, represents element-wise multiplication, and are learnable parameters, G is the weight matrix of the local fine features, 1-G is the weight matrix of the global trend features, is the feature matrix after fusion.

[0085] Further, in an embodiment of the present specification, for any single well, based on the time series data of its corresponding gas production and the time series data of the production key parameters, local fine features and global trend features are extracted, including:

[0086] For any single well, based on the time series data of its corresponding gas production and the time series data of the production key parameters, local fine features are extracted by using a sliding window method, and global trend features are extracted by using a global average pooling method.

[0087] Illustratively, the extraction of local fine features uses a sliding time window to process high-frequency components in time series data, capturing hourly or daily fluctuations; the extraction of global trend features uses global pooling to process low-frequency components in time series data, capturing monthly or annual trends.

[0088] Specifically:

[0089] (1) For high-frequency components, a sliding time window is used to apply self-attention mechanism to capture short-term fluctuations in time series, and the high-frequency attention is calculated by the following formula:

[0090]

[0091]

[0092] where w is the window length, t is the starting position of the window, is a time window in the input time series, are all learnable weight matrices, respectively used to generate Query, Key and Value, Linear(·) is a linear transformation function, which maps the input to a new space through a weight matrix, representing the linear layer (full connection layer) of the neural network, are the Query, Key and Value matrices of the high-frequency attention, which are passed to the subsequent attention scoring step. is the high-frequency attention weight, is the vector dimension, which is converted into a probability distribution by the Softmax function.

[0093] (2) For low-frequency components, a global average pooling method is used to generate a low-frequency representation of the time series, and the low-frequency attention is calculated by the following formula:

[0094]

[0095]

[0096] where, are all learnable weight matrices, respectively used to generate Query, Key and Value, (Average Pooling) compresses the information in the time dimension of the input X, retaining the overall distribution information, Linear(·) is a linear transformation function, which maps the input to a new space through a weight matrix, are the Query, Key and Value matrices of the low-frequency attention, which are passed to the subsequent attention scoring step, is the low-frequency attention weight.

[0097] Step 206, input each single well's respective feature matrix into the trained prediction model to obtain the predicted gas production of the well pattern, the trained prediction model including an adjacency matrix representing the interference relationship between wells.

[0098] wherein the prediction model to be trained includes a parameter matrix with initial values set based on static well-to-well relationships, and the adjacency matrix is a learned parameter matrix corresponding to the trained prediction model.

[0099] Illustratively, the parameter matrix in the prediction model takes the distance matrix X set based on the static interwell relationship in the above example as the initial value of the parameter matrix to represent the estimated interwell interference relationship at the current development stage, and combines the respective feature matrix of each single well to predict the predicted gas production of the well pattern. The parameter matrix corresponding to the trained prediction model is the adjacency matrix representing the interwell interference relationship closest to the real interwell interference relationship at the current development stage.

[0100] It should be noted that at different development stages, the interwell interaction of the well pattern will change, and the learning and training of the prediction model need to be re-performed according to the time series data at the corresponding development stage. Each trained prediction model corresponds to a fixed adjacency matrix. The difference between the adjacency matrices corresponding to different development stages can show the dynamic evolution law of the interwell interference.

[0101] In an embodiment of the present specification, the parameter matrix includes a first parameter matrix and a second parameter matrix representing the asymmetric interference between the wells, and a third parameter matrix representing the characteristics of each single well in the well pattern.

[0102] Illustratively, the first parameter matrix W1 and the second parameter matrix W2 are used to generate an anti-symmetric matrix The first parameter matrix W1 captures the dominant direction influence, and the second parameter matrix W2 describes the secondary reverse action, thereby capturing the asymmetry of the interwell influence. The third parameter matrix represents the characteristics of each single well, and adds a bias correction term to each single well to correct the characteristics of the single well.

[0103] Illustratively, the dynamic interwell relationship graph learning module continuously learns the parameter matrix adaptively through the anti-symmetric matrix and the design of correcting the characteristics of each single well, and finally obtains the adjacency matrix capturing the complex asymmetric influence relationship between the wells. The loss function is used to measure the deviation between the predicted value and the true value, and the loss function is used to update the gradient of the current parameter matrix (initially the distance matrix X), thereby updating the first parameter matrix, the second parameter matrix and the third parameter matrix in the parameter matrix.

[0104] Specifically, the expression of the parameter matrix is:

[0105]

[0106] where ReLU is an activation function, W1 and W2 are learnable parameter matrices, is the product of the transposes of W1 and W2, is the product of the transpose of W1 and W2, and β is a learnable diagonal element parameter vector, diag(β) is a diagonal matrix of the diagonal element parameter vector β.

[0107] In an embodiment of the present specification, the training method of the prediction model comprises the following steps:

[0108] Setting an initial value of the parameter matrix based on the static interwell relationship;

[0109] Inputting the current parameter matrix and the respective feature matrix of each single well into the prediction model to be trained, updating the current parameter matrix based on the loss function of the prediction model to be trained, and repeating the step with the updated parameter matrix as the new current parameter matrix until the loss function of the prediction model to be trained meets the convergence condition to obtain the trained prediction model.

[0110] Illustratively, the deviation between the predicted value and the true value is measured by calculating the loss function, and the loss function is used to update the current parameter matrix by back propagation and gradient descent, so as to train the model until the loss function of the prediction model to be trained meets the convergence condition, and the training of the prediction model is completed.

[0111] Specifically, the parameters of the prediction model are derived, and the gradients of the loss function with respect to the parameters in the expression of the parameter matrix are calculated, and the parameters of the prediction model and the parameters of the parameter matrix are updated simultaneously using the optimizer (such as Adam).

[0112] In an embodiment of the present specification, updating the current parameter matrix based on the loss function of the prediction model to be trained comprises:

[0113] Preliminary updating the current parameter matrix based on the loss function of the prediction model to be trained;

[0114] Obtaining the weight matrix of the two parameter matrices corresponding to before and after updating respectively through the multi-head attention mechanism;

[0115] Based on the two parameter matrices corresponding to before and after updating respectively and their weight matrix, performing secondary updating to obtain the updated parameter matrix.

[0116] Illustratively, the dynamic interwell relationship learning module learns the interwell dynamic correlation strength adaptively through the graph attention mechanism, and captures the mutual interference effect between the wells. In addition to the dynamic updating of the parameter matrix, the key technical point in this module also includes the aggregation strategy of multi-head attention. The gradient update of the parameter matrix through back propagation and gradient descent is the preliminary update, and the updated parameter matrix is , the parameter matrix before updating is , further, the multi-head attention mechanism is used to balance the new parameter matrix and the old parameter matrix learned, comprising the following calculation steps:

[0117] Calculating the query matrix Q, the key matrix K and the value matrix V:

[0118]

[0119] wherein, are learnable parameter matrices, is the transposed matrix of the concatenation of the two parameter matrices corresponding to before and after the update respectively, and then multiplied by three learnable parameter matrices to obtain the query matrix Q, the key matrix K and the value matrix V;

[0120] Single-head attention calculation:

[0121]

[0122] Single-head attention is a standard attention mechanism, as explained above. Multiple independent "attention heads" are executed in parallel, and each head performs independent attention calculation. Since each head uses random initialization parameters, the corresponding are different, so the multi-head attention mechanism can extract information from different angles;

[0123] Calculation of the weight matrix of the two parameter matrices corresponding to before and after the update:

[0124]

[0125] In this embodiment, the multi-head attention mechanism includes two independent "attention heads", and the attention calculation results correspond to , After concatenating the two by Concat(·), the updated parameter matrix is generated by MLP(·) multi-layer perceptron , and the weight matrix corresponding to the parameter matrix before the update is ;

[0126] Second update of the parameter matrix:

[0127]

[0128] wherein, is the updated parameter matrix, represents element-level multiplication, which integrates the results of the two "attention heads" to capture more rich feature information.

[0129] By integrating the dynamic interwell relationship graph learning module and the high-low frequency fusion spatiotemporal convolution prediction module described in the above embodiments, an adaptive spatiotemporal graph neural network model is constructed. As shown in Figure 2As shown, the prediction performance of the model is evaluated using the root mean square error (RMSE) and the mean absolute error (MAE), and the spatio-temporal graph neural network model (AST-GNN) in the application is compared with multiple traditional time series prediction models (CNN, GUR, LSTM, and Transformer). The experimental results show that the method proposed in the application is significantly better than traditional machine learning and time series prediction models in terms of gas production prediction accuracy and multi-step prediction stability, and provides a new technical approach for efficient development of offshore gas fields.

[0130] It can be understood that the gas production is predicted by the adaptive spatio-temporal graph neural network model, and the predicted gas production in the previous step is re-input as the input for the next step prediction, so as to realize recursive calling and self-updating of the model at different time scales.

[0131] In an embodiment of the present specification, the intelligent production prediction further comprises:

[0132] The visual interwell relationship graph is output based on the parameter matrix and the distance matrix.

[0133] Explanatorily, the interwell interaction of the well pattern changes at different development stages, and the adjacency matrix corresponding to the current development stage needs to be re-learned. The adjacency matrices corresponding to different development stages respectively present the interwell interference relationship at different development stages, and the difference between the adjacency matrices corresponding to different development stages respectively presents the dynamic evolution law of the interwell interference.

[0134] Exemplarily, as Figure 3 As shown, the adjacency matrices corresponding to different development stages can be displayed by the visual analysis technology, so as to intuitively reveal the dynamic evolution law of the gas well interference at different development stages, and fully excavate the spatio-temporal dynamic characteristics contained in the well pattern production data. Compared with other methods which focus on numerical prediction, the visual analysis of the present application provides more intuitive and understandable auxiliary decision-making information for oil and gas field engineers.

[0135] Working principle:

[0136] Step 1: data preprocessing.

[0137] Step 2: feature selection. The dynamic similarity between each production parameter and the gas well production is calculated by the dynamic time warping (DTW) method, and the dynamic correlation between different features and the production is evaluated.

[0138] Step 3: build a dynamic interwell relationship graph learning module. Update the parameter matrix to obtain the adjacency matrix.

[0139] Step 4: Constructing a high-low frequency fusion spatio-temporal convolution prediction module. High-frequency and low-frequency self-attention structures are designed in parallel to capture local fine features and global trend features in the time series, respectively, and adaptive feature fusion is realized through a gating unit.

[0140] Step 5: Constructing an adaptive spatio-temporal graph neural network model. The output of the dynamic interwell relationship graph learning module is input into the high-low frequency fusion spatio-temporal convolution prediction module to realize synchronous modeling of the dynamic evolution of interwell relationships and multi-scale patterns of production time series.

[0141] Step 6: Production prediction. The trained model is used to predict the production of the test set data.

[0142] Step 7: Visualization analysis. The dynamic interwell relationships learned by the model are visualized to reveal the dynamic evolution law of the interaction between the wells in different development stages.

[0143] The above describes certain embodiments of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be executed in an order different than the order in the embodiments and still achieve the desired result. In addition, the processes depicted in the figures do not necessarily require the particular order shown or sequential order to achieve the desired results. In certain implementations, multitasking and parallel processing are also permissible or advantageous.

[0144] Next, please refer to Figure 4 , Figure 4 A structure schematic diagram of an offshore oil and gas well pattern production intelligent prediction system provided by an embodiment of the present specification is shown.

[0145] The intelligent prediction system 400 includes a data acquisition module 401, a feature extraction module 402, and a production prediction module 403.

[0146] The data acquisition module 401 acquires time series data of gas production and time series data of production key parameters corresponding to each single well in the well pattern, as well as static interwell relationships representing the relative distance between each single well in the well pattern.

[0147] The feature extraction module 402 extracts features based on the time series data of gas production and the time series data of production key parameters corresponding to each single well, to obtain a feature matrix corresponding to each single well.

[0148] The production prediction module 403 inputs the feature matrix corresponding to each single well into a trained prediction model to obtain the predicted gas production of the well pattern, the trained prediction model including an adjacency matrix representing the interwell interference relationship; wherein the prediction model to be trained includes a parameter matrix with initial values set based on the static interwell relationship, and the adjacency matrix is a learned parameter matrix corresponding to the trained prediction model.

[0149] Each of the embodiments in the specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each of the embodiments focuses on the difference from other embodiments. In particular, for the yield intelligent prediction system embodiment, since it is basically similar to the yield intelligent prediction method embodiment, the description is relatively simple, and the relevant parts can be referred to the part of the yield intelligent prediction method embodiment.

[0150] Please refer to Figure 5 The electronic device provided by the embodiment of the specification is shown in a structural schematic diagram.

[0151] As Figure 5 The electronic device 500 can include at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0152] The communication bus 502 can be used to realize the connection and communication of the above-mentioned components.

[0153] The user interface 503 can include a key, and the optional user interface can also include a standard wired interface, a wireless interface.

[0154] The network interface 504 can include but is not limited to a Bluetooth module, an NFC module, a Wi-Fi module, etc.

[0155] The processor 501 can include one or more processing cores. The processor 501 connects various parts in the electronic device 500 through various interfaces and lines, executes various functions of the electronic device 500 and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 505, and calling data stored in the memory 505. Optionally, the processor 501 can be realized by at least one of the hardware forms of DSP, FPGA, and PLC. The processor 501 can integrate one or a combination of CPU, GPU, and modem. Among them, the CPU mainly processes the operating system, user interface, and application program; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; and the modem is used for processing wireless communication. It can be understood that the above-mentioned modem can also not be integrated into the processor 501, but be realized by a separate chip.

[0156] The memory 505 can include a RAM and can also include a ROM. Optionally, the memory 505 includes a non-transitory computer-readable medium. The memory 505 can be used to store instructions, programs, codes, code sets, or instruction sets. The memory 505 can include a program storage area and a data storage area, where the program storage area can store the instructions for implementing the operating system, the instructions for at least one function (such as a touch function, a sound playing function, an image playing function, etc.), the instructions for implementing the various method embodiments described above, etc.; and the data storage area can store the data involved in the various method embodiments above, etc. The memory 505 can also optionally be at least one storage device located away from the aforementioned processor 501. The memory 505, as a computer storage medium, can include an operating system, a network communication module, a user interface module, and a yield intelligent prediction application program. The processor 501 can be used to invoke the yield intelligent prediction application program stored in the memory 505, and execute the steps of the yield intelligent prediction method mentioned in the foregoing embodiments.

[0157] The embodiments of the present specification also provide a computer-readable storage medium, which stores instructions, and when the instructions run on a computer or a processor, the computer or the processor executes the steps of one or more of the above-mentioned yield intelligent prediction method embodiments. If each component module of the above-mentioned electronic device is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0158] In the above embodiments, all or part of the methods can be implemented by software, hardware, firmware, or any combination thereof. When implemented by software, all or part of the methods can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present specification are generated. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in or transmitted by a computer readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center through a wired (such as a coaxial cable, an optical fiber, a digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) manner. The computer readable storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center, etc. containing one or more available media sets. The available media can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a digital versatile disc (DVD)), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0159] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by a computer program instructing related hardware, which can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments. The storage medium includes ROM, RAM, magnetic or optical disc, and other program code storage media. In the case of no conflict, the technical features in the embodiments and the implementation solutions can be combined arbitrarily.

[0160] The above embodiments are only described as the preferred embodiment of the present specification, and do not limit the scope of the present specification. Without departing from the design spirit of the present specification, various modifications and improvements of the technical solutions of the present specification made by a person of ordinary skill in the art should fall within the protection scope determined by the claims of the present specification.

Claims

1. An intelligent offshore oil and gas field production prediction method, characterized in that, The method comprises the following steps: obtaining time series data of gas production and time series data of production key parameters of each single well in the well pattern, and static interwell relationship representing the relative distance between each single well in the well pattern; extracting features based on the time series data of gas production and the time series data of production key parameters of each single well to obtain a feature matrix corresponding to each single well; specifically, for any single well, local fine features and global trend features are extracted based on the time series data of gas production and the time series data of production key parameters corresponding to the single well; the feature matrix corresponding to the single well is obtained by feature fusion based on the local fine features and the global trend features; the above steps are repeated until the feature matrix corresponding to each single well is obtained; inputting the feature matrix corresponding to each single well into a trained prediction model to obtain the predicted gas production of the well pattern, wherein the trained prediction model comprises an adjacency matrix representing the interwell interference relationship; wherein the prediction model to be trained comprises a parameter matrix with initial values set based on the static interwell relationship, the parameter matrix comprising a first parameter matrix and a second parameter matrix representing the asymmetric interference between wells, and a third parameter matrix representing the characteristics of each single well in the well pattern, and the adjacency matrix is a learned parameter matrix corresponding to the trained prediction model; the training method of the prediction model comprises the following steps: setting initial values of the parameter matrix based on the static interwell relationship; inputting the current parameter matrix and the feature matrix corresponding to each single well into the prediction model to be trained, preliminarily updating the current parameter matrix based on the loss function of the prediction model to be trained, obtaining the weight matrix of the two parameter matrices corresponding to before and after the update through the multi-head attention mechanism, and performing secondary update based on the two parameter matrices and their weight matrices to obtain the updated parameter matrix, and repeating the step with the updated parameter matrix as the new current parameter matrix until the loss function of the prediction model to be trained meets the convergence condition to obtain the trained prediction model.

2. The method of claim 1, wherein, The method comprises the following steps: obtaining time series data of gas production and time series data of production key parameters of each single well in the well pattern, and static interwell relationship representing the relative distance between each single well in the well pattern; extracting features based on the time series data of gas production and the time series data of production key parameters of each single well to obtain a feature matrix corresponding to each single well; specifically, for any single well, local fine features and global trend features are extracted based on the time series data of gas production and the time series data of production key parameters corresponding to the single well; the feature matrix corresponding to the single well is obtained by feature fusion based on the local fine features and the global trend features; the above steps are repeated until the feature matrix corresponding to each single well is obtained; 3. The method of claim 2, wherein, inputting the feature matrix corresponding to each single well into a trained prediction model to obtain the predicted gas production of the well pattern, wherein the trained prediction model comprises an adjacency matrix representing the interwell interference relationship; wherein the prediction model to be trained comprises a parameter matrix with initial values set based on the static interwell relationship, the parameter matrix comprising a first parameter matrix and a second parameter matrix representing the asymmetric interference between wells, and a third parameter matrix representing the characteristics of each single well in the well pattern, and the adjacency matrix is a learned parameter matrix corresponding to the trained prediction model; the training method of the prediction model comprises the following steps: setting initial values of the parameter matrix based on the static interwell relationship; inputting the current parameter matrix and the feature matrix corresponding to each single well into the prediction model to be trained, preliminarily updating the current parameter matrix based on the loss function of the prediction model to be trained, obtaining the weight matrix of the two parameter matrices corresponding to before and after the update through the multi-head attention mechanism, and performing secondary update based on the two parameter matrices and their weight matrices to obtain the updated parameter matrix, and repeating the step with the updated parameter matrix as the new current parameter matrix until the loss function of the prediction model to be trained meets the convergence condition to obtain the trained prediction model. The method comprises the following steps: obtaining time series data of gas production and time series data of production key parameters of each single well in the well pattern, and static interwell relationship representing the relative distance between each single well in the well pattern; extracting features based on the time series data of gas production and the time series data of production key parameters of each single well to obtain a feature matrix corresponding to each single well; specifically, for any single well, local fine features and global trend features are extracted based on the time series data of gas production and the time series data of production key parameters corresponding to the single well; the feature matrix corresponding to the single well is obtained by feature fusion based on the local fine features and the global trend features; the above steps are repeated until the feature matrix corresponding to each single well is obtained; inputting the feature matrix corresponding to each single well into a trained prediction model to obtain the predicted gas production of the well pattern, wherein the trained prediction model comprises an adjacency matrix representing the interwell interference relationship; wherein the prediction model to be trained comprises a parameter matrix with initial values set based on the static interwell relationship, the parameter matrix comprising a first parameter matrix and a second parameter matrix representing the asymmetric interference between wells, and a third parameter matrix representing the characteristics of each single well in the well pattern, and the adjacency matrix is a learned parameter matrix corresponding to the trained prediction model; the training method of the prediction model comprises the following steps: setting initial values of the parameter matrix based on the static interwell relationship; inputting the current parameter matrix and the feature matrix corresponding to each single well into the prediction model to be trained, preliminarily updating the current parameter matrix based on the loss function of the prediction model to be trained, obtaining the weight matrix of the two parameter matrices corresponding to before and after the update through the multi-head attention mechanism, and performing secondary update based on the two parameter matrices and their weight matrices to obtain the updated parameter matrix, and repeating the step with the updated parameter matrix as the new current parameter matrix until the loss function of the prediction model to be trained meets the convergence condition to obtain the trained prediction model.

4. The method of claim 1, wherein, The local fine features and the global trend features are extracted based on the time series data of the corresponding gas production and the time series data of the production key parameters of any single well, including: The local fine features are extracted based on the time series data of the corresponding gas production and the time series data of the production key parameters of any single well in a sliding window manner, and the global trend features are extracted in a global average pooling manner.

5. The method of claim 1, wherein, It also includes: The parameter matrix is used to output a visual interwell relationship diagram.

6. An intelligent offshore oil and gas field production prediction system, characterized by, It includes a data acquisition module, a feature extraction module and a productivity prediction module. The data acquisition module acquires the time series data of the corresponding gas production and the time series data of the production key parameters of each single well in the well pattern, as well as the static interwell relationship representing the relative distance between each single well in the well pattern. The feature extraction module extracts features based on the time series data of the corresponding gas production and the time series data of the production key parameters of each single well, to obtain the feature matrix corresponding to each single well; specifically, the local fine features and the global trend features are extracted based on the time series data of the corresponding gas production and the time series data of the production key parameters of any single well; the feature matrix corresponding to the single well is obtained by feature fusion based on the local fine features and the global trend features; the above steps are repeated until the feature matrix corresponding to each single well is obtained. The productivity prediction module inputs the feature matrix corresponding to each single well into the trained prediction model to obtain the predicted gas production of the well pattern, and the trained prediction model includes an adjacency matrix representing the interwell interference relationship; wherein the prediction model to be trained includes a parameter matrix with initial values set based on the static interwell relationship, the parameter matrix includes a pair of first and second parameter matrices representing asymmetric interwell interference, and a third parameter matrix representing the characteristics of each single well in the well pattern, and the adjacency matrix is the learned parameter matrix corresponding to the trained prediction model. The training of the prediction model sets the initial values of the parameter matrix based on the static interwell relationship; the current parameter matrix and the feature matrix corresponding to each single well are input into the prediction model to be trained, the current parameter matrix is preliminarily updated based on the loss function of the prediction model to be trained, the weight matrix of the two parameter matrices corresponding to before and after the update is obtained through the multi-head attention mechanism, the two parameter matrices before and after the update and their weight matrix are updated again based on the two parameter matrices before and after the update and their weight matrix, to obtain the updated parameter matrix, and the updated parameter matrix is used as the new current parameter matrix to repeat the step until the trained prediction model meets the convergence condition.

Images