A Smart Method for Stratified Water Volume Allocation in Injection Wells Considering Time-Variating Factors

Abstract

This invention discloses an intelligent water volume splitting method for injection wells that considers time-varying factors, belonging to the field of water well volume splitting technology. The method includes the following steps: acquiring production data to construct a connectivity discrimination model dataset for injection and production wells; constructing and training the connectivity discrimination model to obtain connectivity coefficients; acquiring relative permeability, viscosity, capillary pressure, well coordinates, and geological data; calculating the time-varying parameters of the production wells; and weighting the time-varying parameters of the injection wells according to the connectivity coefficients; acquiring geological data, perforation data, and water absorption profile data of the injection wells; constructing and training a Transformer-based intelligent water volume splitting model; and using the trained model to predict the relative water absorption of each sub-layer. This invention can consider the influence of time-varying parameters between injection and production wells on water volume splitting in injection wells, enabling rapid and accurate water volume splitting.

Description

Technical Field

[0001] This invention belongs to the field of water well volume splitting technology, specifically relating to an intelligent splitting method for stratified water volume in injection wells that takes into account time-varying factors. Background Technology

[0002] After years of water injection development, water-driven reservoirs are now generally in a high water-cut stage, with an average water cut of 90%. However, the overall recovery rate is only around 30%, indicating significant development potential and making them a crucial ballast for stable oilfield production and profitability. Water injection development often faces three major contradictions: inter-layer contradictions, intra-layer contradictions, and planar contradictions. Vertically, due to differences in reservoir parameters such as permeability, porosity, and thickness among different layers, high-permeability layers have strong water absorption capacity, while low-permeability layers have weak water absorption capacity, easily leading to single-layer surges and water channeling, thus highlighting inter-layer contradictions. Planarly, due to the influence of permeability differences, high-permeability zones have low seepage resistance and fast water propagation speed, while low-permeability zones have high resistance and slow water propagation speed, resulting in a "tongue-like" phenomenon. Intra-layer, injected water surges along high-permeability areas, while water drive is weaker in areas with poor seepage, resulting in a "finger-like" phenomenon. Due to the aforementioned contradictions, it is difficult to accurately grasp the water absorption status of each sub-layer in the injection well. Clarifying the water absorption status of each sub-layer in the injection well is an important foundation for reservoir dynamic analysis, numerical simulation, and remaining oil analysis, and is crucial for formulating efficient water injection schemes and remaining oil tapping measures.

[0003] Currently, the main methods for stratified water volume analysis in injection wells include the KH splitting method, numerical simulation, water absorption profiling, and machine learning. The KH splitting method only considers the influence of static geological parameters such as permeability and thickness on the water absorption of each sub-layer, neglecting dynamic factors such as injection-production pressure gradient, viscosity, and relative permeability. Therefore, its calculation results cannot accurately reflect the water absorption status of each sub-layer in the well. Numerical simulation requires precise geological models and high-quality historical data fitting to achieve accurate stratified water volume analysis, making it difficult and time-consuming. The water absorption profiling method uses water absorption profile logging data to analyze the water absorption of each sub-layer in the injection well, making it the most accurate method for reflecting the actual underground water injection situation. However, the water absorption profiling method has high testing costs, limited and discontinuous water absorption profile data, making it difficult to grasp the water absorption status of each sub-layer throughout the entire injection period. In recent years, machine learning algorithms, due to their powerful nonlinear mapping, self-learning, and adaptive capabilities, have become a powerful tool for establishing nonlinear correlations, providing a new approach and method for stratified water volume analysis in injection wells. Constructing an intelligent water volume partitioning model for injection wells using machine learning methods can solve the problems of large computational load and long time consumption associated with traditional numerical simulation methods. Furthermore, it can partition the water volume over the entire historical injection period, making it an effective method for dynamic partitioning of water volume in injection wells. However, current water volume partitioning methods based on machine learning models only consider the influence of static and dynamic data, such as static parameters like permeability, porosity, and thickness, and dynamic parameters like injection volume and injection pressure. They do not consider the impact of time-varying factors between injection and production wells on partitioning water volume, such as relative permeability, injection-production pressure gradient, capillary pressure, and viscosity. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes an intelligent water volume splitting method for injection wells that considers time-varying factors. This method comprehensively considers the impact of time-varying factors on the splitting of water volume in injection wells based on the connectivity between injection and production wells. Furthermore, it takes into account the influence of each sub-layer on the splitting of water volume through an attention mechanism. This method can split the water volume of each sub-layer of the injection well with high accuracy and efficiency.

[0005] The technical solution of the present invention is as follows:

[0006] A method for intelligently dividing the stratified water volume of injection wells considering time-varying factors includes the following steps:

[0007] Step 1: Obtain production data to construct a dataset for the injection-production well connectivity discrimination model, and construct and train the injection-production well connectivity discrimination model to obtain connectivity coefficients;

[0008] Step 2: Obtain relative permeability, viscosity, capillary pressure, well coordinates and geological data; calculate the time-varying parameters of the production well; and obtain the time-varying parameters of the injection well by weighting according to the connectivity coefficient.

[0009] Step 3: Obtain geological data, perforation data, and water intake profile data of the injection wells to construct the dataset for the layered water volume intelligent splitting model;

[0010] Step 4: Build and train a Transformer-based intelligent water volume splitting model, and use the trained model to predict the relative water absorption of each sub-layer.

[0011] Furthermore, the specific process of step 1 is as follows:

[0012] Step 1.1: Obtain historical production volume data and bottom hole flowing pressure data of production wells, and historical water injection volume data and bottom hole flowing pressure data of injection wells to construct a dataset for the injection-production well connectivity discrimination model; specifically:

[0013] Obtain historical daily liquid production data of production wells and historical daily water injection data of injection wells, and then stitch together the daily water injection data and daily liquid production data to obtain an injection-production liquid volume data table;

[0014] Obtain historical bottom-hole flowing pressure data for production wells and injection wells, and then stitch the bottom-hole flowing pressure data of injection wells and production wells together to obtain a bottom-hole flowing pressure data table;

[0015] The injection / production fluid volume data table and the bottom hole flowing pressure data table are partitioned using a sliding window method to construct a dataset for the injection / production well connectivity discrimination model; specifically: using past... Data on the injected and produced fluid volume at each time step, and past data. Bottomhole flowing pressure data at each time step are used as input to the injection-production connectivity discrimination model to predict the production volume of the production well at the next time step; the injection-production volume data, bottomhole flowing pressure data, and production volume of the production well at all time steps constitute the injection-production connectivity discrimination model dataset.

[0016] Step 1.2: Construct a connectivity discrimination model between injection and production wells, and train the model using the dataset constructed in Step 1.1 to obtain the connectivity coefficients between injection and production wells.

[0017] Furthermore, in step 1.2, the injection-production well connectivity discrimination model includes a spectrum filtering and conversion module, a data embedding module, a differential attention module, and an effective water injection module; the specific working process of the injection-production well connectivity discrimination model is as follows:

[0018] Step 1.2.1: The spectrum filtering and conversion module performs denoising and smoothing on the input injection and production fluid volume data and bottom hole flowing pressure data. The entire denoising and smoothing process is performed individually for each well. The denoising and smoothing process for the injection volume data of each injection well is as follows:

[0019] Step 1.2.1.1: Convert the water injection volume data to the frequency domain using Fast Fourier Transform, retaining the highest frequency data. After inverse transforming each frequency component back to the time domain, the details are as follows:

[0020] First, the water injection data is processed using Fast Fourier Transform. Convert to the frequency domain and retain The frequency component with the largest amplitude is then subjected to an inverse fast Fourier transform to retain the preceding frequency components. The frequency component corresponding to the maximum amplitude Converted to the original time-domain water injection data ;

[0021] Step 1.2.1.2, for Smoothing is achieved using Hamming window technology; details are as follows:

[0022] First, define a size of The Hanming Window:

[0023] ;

[0024] in, Index of sample points within the window;

[0025] right Fill in:

[0026] ;

[0027] in, This refers to the water injection volume data after filling; for Sequence index; for The sequence length;

[0028] Perform convolution smoothing operation:

[0029] ;

[0030] in, The water injection volume data is smoothed. Indexed by time step number;

[0031] Step 1.2.1.3: Following the same principle as Step 1.2.1.1 to Step 1.2.1.2, the injection and production fluid volume data and the bottom hole flowing pressure data of each injection and production well are denoised and smoothed respectively. Then, the processed data are spliced ​​together to obtain the denoised and smoothed injection and production fluid volume data and bottom hole flowing pressure data.

[0032] Step 1.2.2: Input the denoised and smoothed injection and production fluid volume data and bottom hole flowing pressure data into the data embedding module. The data is mapped to a high dimension through multiple fully connected layers and added to the spatial embedding vector to obtain the information representation of each injection and production well at the current time.

[0033] Step 1.2.3: Use the output of the data embedding module as the input of the differential attention module, and calculate the attention score between each well through differential attention;

[0034] Step 1.2.4: Convert the historical water injection volume into the effective water injection volume for the next time step through the effective water injection module.

[0035] Furthermore, the specific process of step 1.2.2 is as follows:

[0036] First, the noise-reduced and smoothed injection-production fluid volume data Denoising and smoothing bottom hole flowing pressure data After the second and third dimensions are swapped, they are input into the data embedding module. For the total time step, This refers to the number of injection wells. The number of production wells; the data embedding module consists of three parts. The first is the injection and production fluid volume data embedding part, which maps the injection and production fluid volume data to the dimension through multiple fully connected layers. :

[0037] ;

[0038] in, A high-dimensional embedding representation of the injected and produced fluid volume data; and These are the weight parameters for the fully connected layer; and These are the bias parameters for the fully connected layer; For activation functions;

[0039] The second part is the embedding of bottom hole flowing pressure data, which will incorporate past... Bottomhole flowing pressure data from injection and production wells at each time step are mapped to a high dimension through multiple fully connected layers:

[0040] ;

[0041] in, Data embedded for bottom hole flowing pressure; and These are the weight parameters for the fully connected layer; and These are the bias parameters for the fully connected layer;

[0042] The third part is creating a learnable parameter. As a representation of spatial information; then , and Addition as the output of the data embedding module :

[0043] .

[0044] Furthermore, the specific process of step 1.2.3 is as follows: First, obtain the query matrix of the differential attention mechanism. Bond matrix ; Query matrix Bond matrix Each was split along the second dimension to obtain , , , ; , The query matrix is ​​split into two sets of matrices; , These are the two sets of matrices resulting from splitting the key matrix;

[0045] Next, calculate the attention scores between each well:

[0046] ;

[0047] ;

[0048] in, , These are the first and second attention scores, respectively. For activation functions; This is a transpose operation;

[0049] Calculate the final attention score:

[0050] ;

[0051] in, This is the final attention score; These are learnable parameters;

[0052] Connectivity coefficient between injection and production wells Corresponding to the lower left part of the attention score:

[0053] ;

[0054] in, For the first Injection well and the first The connectivity coefficient between production wells;

[0055] Calculate the fluid production rate of each production well:

[0056] ;

[0057] in, For the first The fluid production rate of a production well; For the first The water injection volume of the injection well; For the first Injection well and the first The connectivity coefficient between production wells;

[0058] The specific process of step 1.2.4 is as follows: rewrite the liquid production formula as follows:

[0059] ;

[0060] in, For the first The effective water injection volume of a water injection well;

[0061] The effective water injection module consists of two fully connected layers. First, it processes the input past... Water injection volume at each time step Normalize; then output the effective water injection volume for the next time step, and finally reverse normalize:

[0062] ;

[0063] in, The effective injection volume before inverse normalization; , These are the weight parameters for the fully connected layer; , These are the bias parameters for the fully connected layer; It is the ReLU activation function;

[0064] The mean square error is used as the objective function, and the Adam algorithm is used to iteratively update the parameters of the injection-production well connectivity discrimination model.

[0065] Furthermore, the specific process of step 2 is as follows:

[0066] Step 2.1: Obtain relative permeability curve data, viscosity data, capillary pressure data, coordinate data of each well, and porosity and permeability data of the production well. Calculate the time-varying parameters of the production well, including the relative permeability, comprehensive viscosity, capillary pressure of the production well at each time point, and the pressure gradient between each injection and production well.

[0067] Step 2.2: Based on the connectivity coefficients between injection and production wells obtained in Step 1, the time-varying parameter set of each injection well at each time point is obtained by weighting.

[0068] Furthermore, the specific process of step 2.1 is as follows:

[0069] Step 2.1.1: Obtain relative permeability curve data and calculate the relative permeability and overall viscosity of the production well at various times. The specific process is as follows: First, normalize the relative permeability curve data and establish the relationship between relative permeability and water saturation using a power-law model. Based on the obtained relative permeability data, fit the relative permeability equation of the oil-water two-phase system. Then, obtain the water cut and oil-water two-phase viscosity data of the production well, calculate the water saturation according to the Buckley–Leverett water drive front equation, and then obtain the corresponding relative permeability and overall viscosity of the water phase based on the calculated water saturation.

[0070] Step 2.1.2: Obtain capillary pressure curve data, porosity and permeability data of production well, and calculate oil-water capillary pressure of production well at each time point; convert the capillary pressure curve into a relationship curve between capillary pressure and water saturation: calculate the capillary pressure of production well based on the water saturation obtained in step 2.1.1.

[0071] Step 2.1.3: Based on the bottomhole flowing pressure of the injection and production wells obtained in Step 1, calculate the injection-production pressure gradient between each injection and production well. The calculation process for the injection-production pressure gradient in one time step is as follows: Based on the bottomhole flowing pressure data obtained in Step 1, calculate the pressure difference between each injection well and each production well; obtain the coordinates of each well and calculate the well distance between each injection well and each production well; divide the pressure difference by the well distance to obtain the injection-production pressure gradient between each injection and production well.

[0072] Step 2.1.4: Repeat step 2.1.3 to calculate the injection-production pressure gradient between injection and production wells at all time steps.

[0073] Furthermore, the specific process of step 2.2 is as follows: after calculating the relative permeability, overall viscosity, and capillary pressure of the water phase at each moment in the production well, these three parameters are horizontally concatenated, and the data from all production wells are vertically concatenated to obtain a dimension of [missing information]. Time-varying parameter matrix The connectivity coefficient between injection and production wells calculated in step 1 is used. Multiplying these parameters by the time-varying parameter matrix yields the relative permeability, overall viscosity, and capillary pressure of the water phase in the injection well.

[0074] ;

[0075] in, The parameters include the relative permeability, overall viscosity, and capillary pressure of the water phase at various times in the injection well.

[0076] By concatenating the injection-production pressure gradient data from all time steps, a dimension of [dimensional value] is obtained. Injection-production pressure gradient data The connectivity coefficient between injection and production wells The pressure gradient of the injection well is obtained by weighted summation of the injection and production pressure gradient data:

[0077] ;

[0078] in, This represents the pressure gradient of the injection well at various times. For element-wise multiplication; This is matrix multiplication; It is a vector whose elements are all 1s;

[0079] The time-varying parameter set of the injection well is obtained by stitching together the data of relative permeability, overall viscosity, capillary pressure, and pressure gradient of the water phase.

[0080] ;

[0081] in, This is a set of time-varying parameters for each moment in the injection well.

[0082] Furthermore, the specific process of step 3 is as follows:

[0083] Step 3.1: Obtain perforation data, porosity, and permeability data for each injection well; based on the perforation data, determine the injection layer and corresponding perforation thickness of the injection layer at each time point; horizontally stitch together the perforation thickness, porosity, permeability, relative permeability of the water phase, overall viscosity, capillary pressure, injection-production pressure gradient, and injection volume data; vertically stitch together all sub-layers of the same injection well according to time steps to obtain the... The data dimensions of the injection well are ;in For the first The number of injection layers in a water injection well;

[0084] Step 3.2: Determine the maximum number of injection layers, set up virtual filling layers, and maintain the same dimensions for each injection well;

[0085] The data filling method is used to fill in the number of injection layers in the combined injection wells. First, the maximum number of injection layers in the target block is determined. The input data dimension of the stratified water volume intelligent splitting model is: For injection wells with a smaller number of injection layers than the maximum number of injection layers, the missing layers are filled in the last layer, denoted as a virtual filling layer, where all parameters of the virtual filling layer are set to 0; the data from all injection wells are concatenated to obtain a dimension of The input dataset;

[0086] Step 3.3: Obtain the water absorption profile data of the target block, specifically including the water absorption profile test time and the relative water absorption of each sub-layer. Based on the water absorption profile test time of each injection well, select the data of the corresponding time step in the input dataset of Step 3.2 as input, and use the relative water absorption as output to construct the dataset of the layered water volume intelligent splitting model.

[0087] Furthermore, the specific process of step 4 is as follows:

[0088] Step 4.1: Construct a hierarchical intelligent water volume splitting model based on Transformer; the hierarchical intelligent water volume splitting model includes a data embedding module, an encoder module, and a decoder module;

[0089] The data embedding module includes an embedding layer and positional encoding. The embedding layer consists of a fully connected layer, and the positional encoding uses absolute positional encoding. First, the input data is normalized. Then, the normalized data is passed through a fully connected layer and absolute positional encoding to obtain the output of the data embedding module. ;

[0090] The encoder module consists of a multi-head self-attention layer, a layer normalized activation function, a feedforward neural network, and residual connections. The output of the data embedding module is input into the multi-head self-attention layer, and then passed through three fully connected layers to obtain the query matrix of the smaller layers. Key matrix Sum matrix Through self-attention mechanism and mask matrix Calculate the attention score for each injection layer :

[0091] ;

[0092] The output of self-attention is obtained by weighted summation of the value matrix of each sub-layer using attention scores. :

[0093] ;

[0094] The output of the multi-head attention layer is obtained by using a multi-head attention mechanism. :

[0095] ;

[0096] ;

[0097] in, This is a multi-head attention mechanism; For the number of attention heads; , and The first A person's attention The query matrix, key matrix, and value matrix; These are the weights of the fully connected layer; For splicing operations;

[0098] Add residual connections and layer normalization to the encoder module:

[0099] ;

[0100] in, This is the output after layer normalization; For layer normalization activation function;

[0101] The feedforward neural network consists of two fully connected layers. Residual connections are used to add the input and output of the feedforward neural network. Finally, the output is normalized to obtain the output of the encoder module. :

[0102] ;

[0103] in, and These are the weight parameters of the fully connected layer; and These are the bias parameters for the fully connected layer;

[0104] The decoder module consists of a single fully connected layer, which takes the output of the encoder module as input to obtain the predicted value of relative water absorption. ;

[0105] Step 4.2: Divide the dataset from Step 3.3 into a training set and a validation set according to the proportion, set the basic parameters of the model, train the hierarchical water volume intelligent splitting model with the training set data, obtain the predicted value of relative water absorption of each layer through the decoder module, take the relative water absorption of the smaller layer as the label, use the mean squared error as the target loss function, calculate the error between the predicted value and the actual value of relative water absorption of each layer, calculate the weight gradient of each layer of the neural network based on the obtained error, and update the weights with the Adam algorithm until the error reaches the preset threshold and then stop training.

[0106] After the model training is completed, the input dataset of all time steps of each injection well obtained in step 3.3 is input into the layered water volume intelligent splitting model to obtain the relative water absorption of each sub-layer under all time steps of each injection well.

[0107] The beneficial effects of this invention are as follows: First, the connection relationship between injection and production wells is effectively determined through the injection-production discrimination module. Based on this connection relationship, the influence of time-varying factors between injection and production wells on the stratified water volume splitting is fully considered, avoiding the influence of traditional methods that only consider the dynamic and static data of the well itself. At the same time, Transformer is used as the intelligent stratified water volume splitting model. While mining the nonlinear relationship between dynamic and static factors and the water absorption of the sub-layer through the feedforward neural network, a multi-head attention mechanism is also used to mine the mutual influence relationship between each sub-layer. This not only considers the attributes of the sub-layer itself, but also the degree of influence of other sub-layers on the sub-layer and the influence of the sub-layer on other sub-layers, thus improving the accuracy of the intelligent stratified water volume splitting model for injection wells. Attached Figure Description

[0108] Figure 1 This is a flowchart of the method for dividing the water volume of injection wells in consideration of time-varying factors, which is part of the present invention.

[0109] Figure 2 This is a thermal diagram showing the connectivity of injection and production wells according to an embodiment of the present invention.

[0110] Figure 3 This is a comparison chart of the stratified water volume splitting results of a water injection well according to an embodiment of the present invention.

[0111] Figure 4 This is a historical water injection diagram of each sub-layer of a water injection well according to an embodiment of the present invention. Detailed Implementation

[0112] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0113] like Figure 1 As shown, a method for intelligently dividing the stratified water volume of injection wells considering time-varying factors specifically includes the following steps:

[0114] Step 1: Obtain production data to construct a dataset for the injection-production well connectivity discrimination model, and construct and train the injection-production well connectivity discrimination model to obtain connectivity coefficients; the specific process is as follows:

[0115] Step 1.1: Obtain historical production volume data and bottom hole flowing pressure data of production wells in the target block, and historical water injection volume data and bottom hole flowing pressure data of injection wells to construct a dataset for the injection-production well connectivity discrimination model.

[0116] This involves acquiring historical daily liquid production data from each production well and historical daily water injection data from each injection well in the target block, and then concatenating the daily water injection data and daily liquid production data to form a data dimension of [missing information]. Injection and production volume data table .in For the total time step, This refers to the number of injection wells. For the number of production wells, For in total In the first time step, the first The water injection volume of the injection well and the liquid production volume of the production well at each time step.

[0117] Historical bottom-hole flowing pressure data of each production well and each injection well in the target block are obtained. The bottom-hole flowing pressure data of the injection wells and the production wells are then concatenated to form a data structure with the following dimensions: Bottom hole flowing pressure data table .in For in total In the first time step, the first Bottom hole flowing pressure data for all wells at each time step.

[0118] The injection / production fluid volume data table and the bottom hole flowing pressure data table are divided using a sliding window method to construct a dataset for the injection / production well connectivity discrimination model. Specifically, the method involves using past... Injection and production volume data at each time step and the past Bottom hole flowing pressure data at each time step As input to the injection-production connectivity discrimination model, it predicts the production volume of the production well in the next time step. That is, the sliding window moves at a time step of one time step within a total time step of [missing information]. Slide sequentially along the timeline, each time capturing a continuous segment. The injection and production fluid volume data and bottom hole flowing pressure data of each time step are used as input parameters for a sample, and the production fluid production data of the next production well in the next time step is used as the label for that sample. Through this process, the injection and production fluid volume data and bottom hole flowing pressure data of each time step can be used as input parameters for that sample. The injection and production volume data and bottom hole flowing pressure data at each time step are segmented into... , , .in and These are the injection / production fluid volume data and bottom hole flowing pressure data input to the injection-production connectivity discrimination model, respectively. The production volume data output by the injection-production connectivity discrimination model. This represents the total number of samples.

[0119] Step 1.2: Construct a connectivity discrimination model between injection and production wells, and train the model using the dataset constructed in Step 1.1 to obtain the connectivity coefficients between injection and production wells. The connectivity discrimination model includes a spectral filtering and transformation module, a data embedding module, a differential attention module, and an effective water injection module. The specific working process of the connectivity discrimination model is as follows:

[0120] Step 1.2.1: The spectrum filtering and conversion module processes the input injection / production fluid volume data. and bottom hole flowing pressure data Denoising and smoothing processes are performed, with each well processed individually. The denoising and smoothing process for the injection volume data of each injection well is as follows:

[0121] Step 1.2.1.1: Transform the input time series data to the frequency domain using a Fast Fourier Transform, retaining the highest frequencies. After inverse transforming the main frequency components back to the time domain, the details are as follows:

[0122] First, use the Fast Fourier Transform to convert the past Water injection data at each time step Transforming to the frequency domain, the expression is as follows:

[0123] ;

[0124] in, For frequency components; It is an imaginary number; For indexing.

[0125] Then, keep The expression for filtering out unimportant frequencies from the frequency component with the largest amplitude is as follows:

[0126] ;

[0127] in, For the former The frequency components corresponding to the maximum amplitude; For operation functions; Used from Selected from The frequency component with the largest amplitude This is a hyperparameter.

[0128] Finally, the inverse fast Fourier transform is used to... Converted to the original time-domain water injection data Its expression is as follows:

[0129] ;

[0130] Step 1.2.1.2: Process the time-domain water injection data after step 1.2.1.1. Smoothing is performed using Hamming window technology to reduce spectral leakage; the details are as follows:

[0131] First, define a size of The Hamming window has the following window function expression:

[0132] ;

[0133] in, Index of sample points within the window; window size This is a hyperparameter.

[0134] After noise reduction Fill it to ensure its length and window size are correct. The matching and filling methods are as follows:

[0135] ;

[0136] in, This refers to the water injection volume data after filling; for Sequence index; for The sequence length.

[0137] forward each element and the last Each element is mirrored and padded to the beginning and end of the sequence. This padding method can be used to reduce the original sequence length. Extend to a new length This allows for smoothing of the convolution operation, and its expression is:

[0138] ;

[0139] in, This is the smoothed water injection volume data.

[0140] Step 1.2.1.3: Following the same principle as steps 1.2.1.1 to 1.2.1.2, the injection volume data of each injection well, the production volume data of each production well, and the bottom hole flowing pressure data of each injection well and production well can be denoised and smoothed respectively. Then, the processed data are spliced ​​together to obtain a sequence with the same dimensions as before denoising, which are the denoised and smoothed injection and production volume data respectively. Denoising and smoothing bottom hole flowing pressure data .

[0141] Step 1.2.2: Input the denoised and smoothed injection-production fluid volume data and bottom hole flowing pressure data into the data embedding module. These data are then mapped to a high dimension through multiple fully connected layers and added to the spatial embedding vector to obtain the current information representation of each injection-production well. The specific process is as follows:

[0142] First, and By exchanging the second and third dimensions, we obtain and Then, the injection-production data and bottom hole flowing pressure data are input into the data embedding module. The data embedding module consists of three parts. The first part is the injection-production fluid volume data embedding section, which embeds the past... The injected fluid volume data at each time step is mapped to a dimension through multiple fully connected layers. Its expression is as follows:

[0143] ;

[0144] ;

[0145] ;

[0146] ;

[0147] in, A high-dimensional embedding representation of the injected and produced fluid volume data; For the dimensions of the fully connected layer; , , , and These are the weight parameters for the fully connected layer; , , , and These are the bias parameters for the fully connected layer; and For activation functions; This is the embedded representation after passing through the first fully connected layer.

[0148] The second part is the embedding of bottom hole flowing pressure data, which will incorporate past... The bottomhole flowing pressure data of injection and production wells at each time step are mapped to a high dimension through multiple fully connected layers, and the expression is as follows:

[0149] ;

[0150] in, Data embedded for bottom hole flowing pressure; and These are the weight parameters for the fully connected layer; and These are the bias parameters for the fully connected layer.

[0151] The third part is creating a learnable parameter. As a spatial information representation, it is used to characterize the comprehensive static attributes of each injection and production well.

[0152] Then , and Addition as the output of the data embedding module Its expression is as follows:

[0153] ;

[0154] Step 1.2.3: Use the output of the data embedding module as the input of the differential attention module, and calculate the attention score between wells using differential attention. First, obtain the query matrix of the differential attention mechanism. Bond matrix Its expression is as follows:

[0155] ;

[0156] ;

[0157] in, and These are the weight parameters for the fully connected layer.

[0158] query matrix Bond matrix Divide into two groups along the second dimension on average: , , , The expression is as follows:

[0159] ;

[0160] ;

[0161] in, For splicing operations; , The query matrix is ​​split into two sets of matrices; , These are the two sets of matrices resulting from splitting the key matrix.

[0162] Then calculate the attention scores between each well. and , of which Line 1 The value of the column represents the first Koujing and the first The correlation between wells is expressed as follows:

[0163] ;

[0164] ;

[0165] in, , These are the first and second attention scores, respectively. for Activation function; This is a transpose operation.

[0166] The final attention score is obtained by subtracting the two attention scores, and is used as the connectivity coefficient between the wells. Its expression is as follows:

[0167] ;

[0168] ;

[0169] in, This is the final attention score; As weight; , , , It is a learnable vector.

[0170] Connectivity coefficient between injection and production wells The expression corresponding to the lower left or upper right part of the attention score is as follows:

[0171] ;

[0172] in, For the first Injection well and the first The connectivity coefficient between production wells.

[0173] The inter-well connectivity coefficient, also known as the allocation factor, is defined as follows: Injection well and the first Connectivity coefficient between production wells For the first The injected fluid from the injection well is distributed to the first The volume percentage of a production well is determined by the fact that a single production well is often influenced by multiple injection wells. Therefore, the production volume of a production well is the sum of the flow rates from each injection well to that production well, expressed as follows:

[0174] ;

[0175] in, For the first The fluid production rate of a production well; For the first The water injection volume of the injection well;

[0176] Step 1.2.4: However, the propagation of injected water in the formation exhibits time lag and attenuation. To account for these time lags and attenuation during the injection process, the historical injection volume is converted into the effective injection volume for the next time step using an effective injection module. Therefore, the formula for the production well's fluid production can be rewritten as follows:

[0177] ;

[0178] in, For the first The effective water injection volume of a water injection well.

[0179] Specifically, the effective water injection module consists of two fully connected layers. First, the input is... Water injection volume at each time step Then, the input data is normalized using a minimum-maximum method, the expression of which is as follows:

[0180] ;

[0181] in, This is the normalized water injection volume; It is the minimum value in the sample; It represents the maximum value in the sample.

[0182] Then, two fully connected layers are set up, and the activation function is adopted. The activation function, the output of the effective water injection module is the effective water injection volume for the next time step, and finally, inverse normalization is performed. Its expression is:

[0183] ;

[0184] ;

[0185] in, The effective injection volume before inverse normalization; The effective injection volume after denormalization; , These are the weight parameters for the fully connected layer; , These are the bias parameters for the fully connected layer; for Activation function;

[0186] The mean squared error is used as the objective function to calculate the error between the predicted and actual production of the production well. The Adam algorithm is then used to iteratively update the parameters of the injection-production well connectivity discrimination model. The loss function expression is as follows:

[0187] ;

[0188] ;

[0189] in, Mean squared error; , The first production well number The predicted and actual liquid production at each time step.

[0190] Step 2: Obtain relative permeability, viscosity, capillary pressure, well coordinates, and geological data; calculate the time-varying parameters of the production well; and obtain the time-varying parameters of the injection well by weighting based on the connectivity coefficient. The specific process is as follows:

[0191] Step 2.1: Obtain the relative permeability curve data, viscosity data, capillary pressure data, well coordinate data, and porosity and permeability data of the production wells in the target block. Calculate the time-varying parameters of the production wells, including the relative permeability, overall viscosity, capillary pressure, and pressure gradient between injection and production wells at various times. The specific process is as follows:

[0192] Step 2.1.1: Obtain the relative permeability curve data of the target block and calculate the relative permeability and overall viscosity of the production well at various times. The specific process is as follows: First, normalize the relative permeability curve data and establish the relationship between relative permeability and water saturation using a power-law model. The expression is as follows:

[0193] ;

[0194] ;

[0195] ;

[0196] ;

[0197] in, This represents the normalized water saturation. This represents the normalized oil saturation. The water saturation level before normalization; To bind water saturation; Residual oil saturation; The relative permeability of the aqueous phase; The relative permeability of the oil phase; This is the proportionality coefficient of the relative permeability of the aqueous phase; It is an index of the relative permeability of the aqueous phase; This is the proportionality coefficient of the relative permeability of the oil phase; It is an index representing the relative permeability of the oil phase; The normalized exponent is Water saturation; The normalized exponent is The oil saturation.

[0198] Then, based on the obtained relative permeability data, the equation for the relative permeability of the oil-water two phases is fitted to obtain... , , and Then, the relative permeability of the water phase at each time point can be calculated based on the water cut at each time step. Specifically: First, obtain the water cut and oil-water two-phase viscosity data for each time step of the production well. Calculate the water saturation using the Buckley-Leverett water drive front equation. That is, the relationship between water cut and water saturation can be obtained from the relationship between water cut and relative permeability. Therefore, given the water cut, the corresponding water saturation can be calculated by interpolation. Then, the corresponding relative permeability of the water phase can be obtained based on the calculated water saturation. Wherein, the water cut... The expression is as follows:

[0199] ;

[0200] in, Water production; Oil production; The viscosity of the aqueous phase; This represents the viscosity of the oil phase.

[0201] After obtaining the water saturation of the production well at various times, the overall viscosity of the production well at each time can be calculated. The overall viscosity is the average effective viscosity of the oil-water two-phase fluid in the reservoir, and its expression is:

[0202] ;

[0203] Step 2.1.2: Obtain capillary pressure curve data for the target block, and porosity and permeability data for each sublayer of the production well. Calculate the oil-water capillary pressure at each time point in the production well. The obtained capillary pressure curve is usually a curve showing the relationship between capillary pressure and mercury saturation. Therefore, it needs to be converted into a curve showing the relationship between capillary pressure and water saturation, as shown in the following expression:

[0204] ;

[0205] ;

[0206] in, The radius of the capillary tube; This refers to the capillary pressure of mercury. The surface tension of mercury; The wetting angle of mercury; This refers to the capillary pressure of the oil and water system. The interfacial tension between oil and water; The wetting angle of oil and water; For the J function; For the permeability of the small layer; Porosity of the small layer;

[0207] To reflect the change of capillary pressure with water saturation, it is necessary to establish the relationship between capillary pressure and water saturation. A power-law model is used to establish the relationship between the J function and water saturation. The rewritten expression of the J function is defined as follows:

[0208] ;

[0209] in, , For parameters; The parameters before normalization are Water saturation;

[0210] Based on the obtained capillary pressure data, the relationship between the J function and water saturation was fitted, and the parameters were obtained. and Combining the rewritten J function and the relationship between the J function and water saturation, we obtain the expression for capillary pressure and water saturation:

[0211] ;

[0212] Therefore, based on the water saturation of the production well at any given time obtained above, and after obtaining the permeability and porosity of each sub-layer of the production well, the capillary pressure of the oil-water two-phase at any given time in each sub-layer of the production well can be calculated.

[0213] Step 2.1.3: Based on the bottom hole flowing pressure data of the production well obtained in Step 1. and bottom hole flowing pressure data of water injection wells The pressure difference between each injection well and each production well is calculated, and its expression is as follows:

[0214] ;

[0215] ;

[0216] in, This is the pressure difference matrix between injection wells and production wells; For the first Injection well and the first Pressure difference between production wells; For the first Injection well and the first Pressure difference between production wells; For the first Bottom-hole flowing pressure of a water injection well For the first Bottom-hole flowing pressure of a production well.

[0217] Obtain the coordinates of the production wells and the injection wells, and calculate the well distance between each injection well and each production well. The expression is as follows:

[0218] ;

[0219] ;

[0220] in, This is the distance matrix between injection wells and production wells; For the first Injection well and the first The distance between production wells; For the first Injection well and the first The distance between production wells; , The first The x-axis and y-axis of the injection well; , The first The horizontal and vertical coordinates of the production well.

[0221] Dividing the injection-production pressure difference by the well spacing yields the injection-production pressure gradient between each injection-production well. Its expression is as follows:

[0222] ;

[0223] Step 2.1.4 It is worth noting that step 2.1.3 above is a time step injection-production pressure gradient calculation process. Repeating step 2.1.3 can calculate the injection-production pressure gradient between injection and production wells in all time steps.

[0224] Step 2.2: Based on the connectivity coefficient between injection and production wells obtained in Step 1, the relative permeability, comprehensive viscosity, capillary pressure, and pressure gradient of the water phase at each time point of each injection well are obtained by weighting.

[0225] After calculating the relative permeability, overall viscosity, and capillary pressure of the water phase at various times in the production wells, these three parameters were horizontally concatenated, and the data from all production wells were vertically concatenated to obtain a dimension of [missing information]. Time-varying parameter matrix The connectivity coefficient between injection and production wells calculated in step 1 is used. Multiplying these parameters by the time-varying parameter matrix yields the relative permeability, overall viscosity, and capillary pressure of the water phase in the injection well, expressed as follows:

[0226] ;

[0227] in, The parameters are the relative permeability, overall viscosity, and capillary pressure of the water phase at various times in the injection well.

[0228] By concatenating the injection-production pressure gradient data from all time steps, a dimension of [dimensional value] is obtained. Injection-production pressure gradient data The connectivity coefficient between injection and production wells The pressure gradient of the injection well is obtained by weighted summation of the injection and production pressure gradient data, and its expression is as follows:

[0229] ;

[0230] in, This represents the pressure gradient of the injection well at various times. For element-wise multiplication; This is matrix multiplication; It is a vector whose elements are all 1s.

[0231] The time-varying parameter set of the injection well is obtained by stitching together the data of relative permeability, overall viscosity, capillary pressure, and pressure gradient of the water phase. The expression is as follows:

[0232] ;

[0233] in, This is a set of time-varying parameters for each moment in the injection well.

[0234] Step 3: Obtain geological data, perforation data, and water intake profile data from the injection wells to construct the dataset for the layered water volume intelligent splitting model. The specific process is as follows:

[0235] Step 3.1: Obtain perforation data for each injection well and porosity and permeability data for each sub-layer of each injection well; the perforation data includes perforation layer, perforation thickness, and perforation time; the perforation thickness, porosity, permeability, relative permeability, overall viscosity, capillary pressure, injection-production pressure gradient, and injection volume are horizontally stitched together to obtain a dynamic and static parameter dataset;

[0236] Based on perforation data, the injection layer and corresponding perforation thickness of each injection well at each time point are determined. The perforation thickness, porosity, permeability, relative permeability of the aqueous phase, overall viscosity, capillary pressure, injection-production pressure gradient, and injection volume data are then horizontally stitched together; that is, the dimension of each sub-layer is... By longitudinally splicing all the sub-layers of the same injection well according to time steps, the first layer is obtained. The data dimensions of the injection well are .in For the first The number of water injection layers in a water injection well.

[0237] Step 3.2: Determine the maximum number of injection layers, set up virtual filling layers, and maintain the same dimensions for each injection well;

[0238] Since the number of combined injection layers varies among the injection wells, a filling method is used to ensure that the input data has the same dimensionality. First, the maximum number of combined injection layers in the target block's injection wells is determined. The input data dimension of the stratified water volume intelligent splitting model is: For combined betting layers smaller than the maximum combined betting layer For each injection well, the missing layers are uniformly filled in the last layer, denoted as the virtual filling layer. All eight feature parameters of the virtual filling layer are set to 0, thus making the data dimension of all injection wells zero. By concatenating the data from all injection wells, a dimension of [dimensional value] is obtained. The input dataset, i.e., Each time step There are 10 water injection wells, where the input dimension for each water injection well is 1. .

[0239] Step 3.3: Obtain water absorption profile data of the target block, and use the relative water absorption as the training label to establish a dataset for the layered water volume intelligent splitting model.

[0240] Obtain the water absorption profile data of the target block, specifically including the water absorption profile test time and the relative water absorption of each sub-layer. Based on the water absorption profile test time of each injection well, select the data from the corresponding time step in the input dataset of step 3.2 as input, and use the relative water absorption as the output, i.e., the label of the stratified water volume intelligent partitioning model, to construct the dataset of the stratified water volume intelligent partitioning model, where the input dimension is... The output dimension is , The number of samples for the water absorption profile test is shown. It is worth noting that the water absorption profile data does not include the water absorption of the virtual layer. The virtual layer is only used to ensure dimensional consistency, so the relative water absorption of the virtual layer is also set to 0.

[0241] Step 4: Construct and train a Transformer-based intelligent water volume partitioning model, and use the trained model to predict the relative water absorption of each sub-layer. This includes the following steps:

[0242] Step 4.1: Construct a Transformer-based intelligent water volume partitioning model. The intelligent water volume partitioning model mainly includes a data embedding module, an encoder module, and a decoder module. The data embedding module consists of an embedding layer and a position encoding part. The embedding layer is composed of a fully connected layer, and the position encoding part uses absolute position encoding technology. Since parameters such as permeability, porosity, and perforation thickness have different dimensions, it is necessary to normalize each parameter in the input data to eliminate the influence of different parameter dimensions. The expression is as follows:

[0243] ;

[0244] in, This is the normalized input parameter matrix; This is the input parameter matrix before normalization; It is the set of minimum values ​​of each parameter before normalization; This is the set of maximum values ​​for each parameter before normalization.

[0245] Then, the normalized data is passed through a fully connected layer, mapping the 8-dimensional input features to a 1-dimensional feature map. The activation function uses The activation function is expressed as follows:

[0246] ;

[0247] in, An embedded representation of the input parameters; These are the weight parameters for the fully connected layer; These are the bias parameters for the fully connected layer.

[0248] After the embedding layer, absolute positional encoding is used to obtain the positional relationships between each sub-layer and between features. The positional encoding expression is as follows:

[0249] ;

[0250] ;

[0251] in, For position encoding functions; This refers to the position index of the smaller layer in the input data; Indexes for each feature; This is the output of the data embedding module.

[0252] The encoder module consists of a multi-head self-attention layer, layer normalized activation functions, a feedforward neural network, and residual connections. The output of the data embedding module is input into the multi-head self-attention layer. In this layer, the output passes through three fully connected layers to obtain the query matrix, key matrix, and value matrix of each sub-layer, as expressed below:

[0253] ;

[0254] ;

[0255] ;

[0256] in, , and These are the query matrix, value matrix, and key matrix for each sub-layer; , and These are the weight parameters of the fully connected layer; , and These are the bias parameters for the fully connected layer.

[0257] Through a self-attention mechanism, each injection layer can pay attention to the features of other injection layers, thereby learning the influence of other sub-layers on the current sub-layer and the influence of the current sub-layer on other sub-layers, which is the attention score in the attention mechanism. Its expression is as follows:

[0258] ;

[0259] To maintain dimensionality consistency, a virtual padding layer is added to the input data. Since this virtual padding layer does not affect other smaller layers, the mask matrix is ​​used to ignore its influence when calculating the attention score. Therefore, the above formula can be rewritten as:

[0260] ;

[0261] ;

[0262] in, It is a mask matrix; For the first The smaller layers and the first The masking relationship between the sub-layers This indicates that the two sub-layers influence each other, meaning that both sub-layers are actual water injection layers. This means that the two sub-layers do not affect each other, that is, at least one of the two sub-layers is a virtual layer.

[0263] Finally, the attention scores are used to evaluate the value matrix of each sub-layer. We perform a weighted summation to obtain the output of self-attention. Its expression is as follows:

[0264] ;

[0265] To improve the learning efficiency of the self-attention mechanism, a multi-head attention mechanism is used instead of the self-attention mechanism. The multi-head attention mechanism uses a matrix... , and The multi-head attention is divided into multiple matrices of the same size, computed in parallel in different multi-head spaces, and then the outputs of the multiple attention heads are concatenated. Finally, a linear transformation is performed to obtain the final multi-head attention output, the expression of which is:

[0266] ;

[0267] ;

[0268] ;

[0269] ;

[0270] ;

[0271] in, This is the output of the multi-head attention mechanism module; This is a multi-head attention mechanism; For splicing operations; For the number of attention heads; , and The first A person's attention The query matrix, key matrix, and value matrix; These are the weights of the fully connected layer.

[0272] Meanwhile, to alleviate the problems of gradient explosion and gradient vanishing during training and accelerate model convergence, residual connections and layer normalization are added to the encoder module, with the following expressions:

[0273] ;

[0274] in, This is the output after layer normalization; This is the layer normalization activation function.

[0275] The feedforward neural network consists of two fully connected layers, where the ReLU activation function is used. A residual connection is used to sum the input and output of the feedforward neural network, and finally, the output is normalized and denoted as the output of the encoder module. Its expression is as follows:

[0276] ;

[0277] in, It is the ReLU activation function; and These are the weight parameters of the fully connected layer; and These are the bias parameters for the fully connected layer.

[0278] The decoder module consists of a fully connected layer that maps the high-dimensional data output by the encoder to the target space, i.e., the predicted relative water absorption. Its expression is as follows:

[0279] ;

[0280] in, These are the weight parameters for the fully connected layer; These are the bias parameters for the fully connected layer.

[0281] Step 4.2: Use the dataset obtained in Step 3.3 to train and validate the stratified water volume intelligent splitting model, and use the trained stratified water volume intelligent splitting model to split the relative water absorption of each small layer of each injection well.

[0282] The dataset from step 3.3 is randomly divided into a training set and a validation set in an 8:2 ratio. The basic parameters of the model are set, and then the Transformer hierarchical water volume intelligent splitting model is trained using the training set data. The decoder module obtains the predicted relative water absorption value for each layer. Using the relative water absorption value of each layer as the label, the mean squared error is used as the objective loss function to calculate the error between the predicted and actual relative water absorption values ​​for each layer. Based on the obtained error, the weight gradient of each layer of the neural network is calculated, and the weights are updated using the Adam algorithm. Training stops when the error meets the requirements (i.e., reaches a pre-set threshold). The objective loss function expression is as follows:

[0283] ;

[0284] in, Mean squared error; For the first Predicted relative water absorption for each sample; For the first The actual relative water absorption of each sample; The number of samples;

[0285] After the model training is completed, the input datasets for all time steps of each injection well obtained in step 3.3 are input into the stratified water volume intelligent splitting model to obtain the relative water absorption of each sub-layer at all time steps of each injection well. Then, based on the principle that the water injection volume of the injection well is equal to the sum of the water absorption of each sub-layer, the water injection volume of each sub-layer is normalized to obtain the actual water injection volume of each sub-layer, the expression of which is as follows:

[0286] ;

[0287] in, For small layers The predicted water injection volume; This refers to the water injection volume of the injection well; For small layers Predicted relative water absorption; For a small number of layers.

[0288] To demonstrate the feasibility and superiority of the present invention, the following embodiments are provided.

[0289] A certain oil reservoir will undergo water injection development from 2012 to 2025, including 11 water injection wells (I1-I11) and 25 production wells (P1-P25). The inverse seven-point well pattern is used for production. The average reservoir depth is 2300 meters, the initial reservoir pressure is 22.8 MPa, the average reservoir porosity is 0.175, the average reservoir permeability is 108 md, the oil and water viscosities are 70 mPa*s and 1 mPa*s, respectively, and the current overall water cut of the reservoir is 58.6%, with a recovery rate of 15.9%. The reservoir is divided into 11 layers, with each well having a different production layer.

[0290] First, production dynamic data from 11 injection wells and 25 production wells were acquired. Using injection-production fluid volume data and bottom hole flowing pressure data from the past 128 time steps, the production volume of the production wells in the next time step was predicted. A dataset was created using a sliding window method, resulting in 4617 samples from 4745 time steps. A well connectivity discrimination model was constructed, dividing the dataset into training and validation sets in an 8:2 ratio. The model was trained using the training set, with a maximum of 2000 iterations and a learning rate of 0.003, decreasing by 0.02% after each iteration. The mean squared error between the predicted and actual production volumes was calculated as the model loss. The Adam optimizer was used, with backpropagation to update relevant model parameters. Training was stopped when the accuracy exceeded 0.8, and the connectivity coefficients between injection and production wells were derived. The heatmap of injection-production connectivity is shown below. Figure 2 As shown.

[0291] Then, based on the relative permeability curves, viscosity data, capillary pressure curves, well coordinates, and geological data of the production wells collected in step 2, the relative permeability, overall viscosity, capillary pressure, and injection-production pressure gradient between injection and production wells for all time steps of the 25 production wells were calculated; according to Figure 2 The relative permeability, overall viscosity, capillary pressure, and injection-production pressure gradient of all time steps of the 11 injection wells were obtained by weighting the connectivity coefficients.

[0292] The porosity, permeability, perforation thickness and relative permeability, overall viscosity, capillary pressure, injection-production pressure gradient, and injection volume of each sub-layer from the 11 injection wells collected in step 3 are horizontally stitched together. Then, the data of all sub-layers are vertically stitched together in chronological order to obtain the input for the liquid volume splitting model. The data dimensions of each well are {4745×5×8, 4745×2×8, 4745×5×8, 4745×7×8, 4745×4×8, 4745×5×8, 4745×2×8, 4745×4×8, 4745×3×8, 4745×5×8, 4745×5×8}, where the first dimension is the total time step, the second dimension is the number of injection layers in each injection well, and the third dimension is the number of input features. Since the number of combined injection layers varies among the 11 injection wells, a filling method is used. The maximum number of combined injection layers in this reservoir is 7. Therefore, the input data dimension of the intelligent water volume splitting model is 4745×7×8. For injection wells with fewer than 7 combined injection layers, the missing layers are uniformly filled in the last layer, denoted as a virtual filling layer. The 8 feature parameters of the virtual filling layer are all set to 0. The data from all injection wells are concatenated to obtain an input dataset with a dimension of (4745×11)×7×8. A total of 105 water absorption profile test data were collected for this reservoir, resulting in only 105 training samples. The input data corresponding to the water absorption profile test time is selected, and the relative water absorption is used as the label to construct the dataset for the intelligent water volume splitting model. The input dimension is 105×7×8, the output dimension is 105×7×1, and the remaining 52090 samples are the data to be split.

[0293] A stratified water volume intelligent partitioning model was constructed. 105 training samples were divided into training and validation sets in an 8:2 ratio. The model was trained using the training set data, with a learning rate of 0.005 and 2000 iterations. The learning rate was reduced by 0.02% after every 15 iterations. The mean squared error between the model's predicted relative water absorption and the actual relative water absorption was calculated as the model's loss. The Adam optimizer was used, with backpropagation to update relevant model parameters until training was complete, and the model was saved. In this case, the test set error was 0.11. The saved model was then used to partition water volume across all time steps, yielding the stratified water volume partitioning results for 11 wells from 2012 to 2025. Figure 3This is a comparison chart of the stratified water volume splitting results based on the test set using different models. KH splitting represents the splitting result using the KH method, which only considers permeability and perforation thickness. XGBoost, considering the influence of time-varying parameters, represents the splitting result obtained by inputting eight parameters into the XGBoost model. Transformer, also considering the influence of time-varying parameters, represents the splitting result obtained by inputting eight parameters into the splitting model of this invention. Transformer, without considering the influence of time-varying parameters, represents the splitting result obtained by inputting only four parameters: permeability, porosity, perforation thickness, and injection volume into the splitting model of this invention. Figure 3 It can be seen that the splitting model that considers time-varying parameters such as relative permeability and capillary pressure yields results closer to the field test values ​​than the splitting model that does not consider time-varying parameters; at the same time, when both consider time-varying parameters, the Transformer splitting model of this invention is more accurate than the XGBoost splitting model.

[0294] Figure 4 This is the water volume breakdown result for a certain well from 2012 to 2025, where layer 1 to layer 3 are different sub-layers. Figure 4 It can be seen that layer 1 is the main water-absorbing layer. From 2014 to 2022, the water absorption rate of layer 1 was greater than 60%, while the water absorption rates of layer 2 and layer 3 fluctuated below 40%.

[0295] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for intelligently dividing the stratified water volume of injection wells considering time-varying factors, characterized in that, Includes the following steps: Step 1: Obtain production data to construct a dataset for a connectivity discrimination model of injection and production wells, and construct and train the connectivity discrimination model of injection and production wells to obtain connectivity coefficients; the specific process is as follows: Step 1.1: Obtain historical production volume data and bottom hole flowing pressure data of production wells, and historical water injection volume data and bottom hole flowing pressure data of injection wells to construct a dataset for the injection-production well connectivity discrimination model; specifically: Obtain historical daily liquid production data of production wells and historical daily water injection data of injection wells, and then stitch together the daily water injection data and daily liquid production data to obtain an injection-production liquid volume data table; Obtain historical bottom-hole flowing pressure data for production wells and injection wells, and then stitch the bottom-hole flowing pressure data of injection wells and production wells together to obtain a bottom-hole flowing pressure data table; The injection / production fluid volume data table and the bottom hole flowing pressure data table are partitioned using a sliding window method to construct a dataset for the injection / production well connectivity discrimination model; specifically: using past... Data on the injected and produced fluid volume at each time step, and past data. Bottomhole flowing pressure data at each time step are used as input to the injection-production connectivity discrimination model to predict the production volume of the production well at the next time step; the injection-production volume data, bottomhole flowing pressure data, and production volume of the production well at all time steps constitute the injection-production connectivity discrimination model dataset. Step 1.2: Construct a well connectivity discrimination model and train it using the dataset constructed in Step 1.1 to obtain the connectivity coefficients between wells. The well connectivity discrimination model includes a spectral filtering and conversion module, a data embedding module, a differential attention module, and an effective water injection module. The specific working process of the well connectivity discrimination model is as follows: Step 1.2.1: The spectrum filtering and conversion module performs denoising and smoothing on the input injection and production fluid volume data and bottom hole flowing pressure data. The entire denoising and smoothing process is performed individually for each well. The denoising and smoothing process for the injection volume data of each injection well is as follows: Step 1.2.1.1: Convert the water injection volume data to the frequency domain using Fast Fourier Transform, retaining the highest frequency data. After inverse transforming each frequency component back to the time domain, the details are as follows: First, the water injection data is processed using Fast Fourier Transform. Convert to the frequency domain and retain The frequency component with the largest amplitude is then subjected to an inverse fast Fourier transform to retain the preceding frequency components. The frequency component corresponding to the maximum amplitude Converted to the original time-domain water injection data ; Step 1.2.1.2, for Smoothing is achieved using Hamming window technology; details are as follows: First, define a size of Hanming Window: ； in, Index of sample points within the window; right Fill in: ； in, This refers to the water injection volume data after filling; for Sequence index; for The sequence length; Perform convolution smoothing operation: ； in, The water injection volume data is smoothed. Indexed by time step number; Step 1.2.1.3: Following the same principle as Step 1.2.1.1 to Step 1.2.1.2, the injection and production fluid volume data and the bottom hole flowing pressure data of each injection and production well are denoised and smoothed respectively. Then, the processed data are spliced ​​together to obtain the denoised and smoothed injection and production fluid volume data and bottom hole flowing pressure data. Step 1.2.2: Input the denoised and smoothed injection and production fluid volume data and bottom hole flowing pressure data into the data embedding module. The data is mapped to a high dimension through multiple fully connected layers and added to the spatial embedding vector to obtain the information representation of each injection and production well at the current time. Step 1.2.3: Use the output of the data embedding module as the input of the differential attention module, and calculate the attention score between each well through differential attention; Step 1.2.4: Convert the historical water injection volume into the effective water injection volume for the next time step through the effective water injection module; Step 2: Obtain relative permeability, viscosity, capillary pressure, well coordinates and geological data; calculate the time-varying parameters of the production well; and obtain the time-varying parameters of the injection well by weighting according to the connectivity coefficient. Step 3: Obtain geological data, perforation data, and water intake profile data of the injection wells to construct the dataset for the layered water volume intelligent splitting model; Step 4: Build and train a Transformer-based intelligent water volume splitting model, and use the trained model to predict the relative water absorption of each sub-layer.

2. The intelligent water volume splitting method for injection wells considering time-varying factors according to claim 1, characterized in that, The specific process of step 1.2.2 is as follows: First, the noise-reduced and smoothed injection-production fluid volume data Denoising and smoothing bottom hole flowing pressure data After the second and third dimensions are swapped, they are input into the data embedding module. For the total time step, This refers to the number of injection wells. The number of production wells; the data embedding module consists of three parts. The first is the injection and production fluid volume data embedding part, which maps the injection and production fluid volume data to the dimension through multiple fully connected layers. : ； in, A high-dimensional embedding representation of the injected and produced fluid volume data; and These are the weight parameters for the fully connected layer; and These are the bias parameters for the fully connected layer; For activation functions; The second part is the embedding of bottom hole flowing pressure data, which will incorporate past... Bottomhole flowing pressure data from injection and production wells at each time step are mapped to a high dimension through multiple fully connected layers: ； in, Data embedded for bottom hole flowing pressure; and These are the weight parameters for the fully connected layer; and These are the bias parameters for the fully connected layer; The third part is creating a learnable parameter. As a representation of spatial information; then , and Addition as the output of the data embedding module : 。 3. The intelligent water volume splitting method for injection wells considering time-varying factors according to claim 2, characterized in that, The specific process of step 1.2.3 is as follows: First, obtain the query matrix of the differential attention mechanism. Bond matrix ; Query matrix Bond matrix Each was split along the second dimension to obtain , , , ; , The query matrix is ​​split into two sets of matrices; , These are the two sets of matrices resulting from splitting the key matrix; Next, calculate the attention scores between each well: ； ； in, , These are the first and second attention scores, respectively. For activation functions; This is a transpose operation; Calculate the final attention score: ； in, This is the final attention score; These are learnable parameters; Connectivity coefficient between injection and production wells Corresponding to the lower left part of the attention score: ； in, For the first Injection well and the first The connectivity coefficient between production wells; Calculate the fluid production rate of each production well: ； in, For the first The fluid production rate of a production well; For the first The water injection volume of the injection well; For the first Injection well and the first The connectivity coefficient between production wells; The specific process of step 1.2.4 is as follows: rewrite the liquid production formula as follows: ； in, For the first The effective water injection volume of a water injection well; The effective water injection module consists of two fully connected layers. First, it processes the input past... Water injection volume at each time step Normalize; then output the effective water injection volume for the next time step, and finally reverse normalize: ； in, The effective injection volume before inverse normalization; , These are the weight parameters for the fully connected layer; , These are the bias parameters for the fully connected layer; It is the ReLU activation function; The mean square error is used as the objective function, and the Adam algorithm is used to iteratively update the parameters of the injection-production well connectivity discrimination model.

4. The intelligent water volume splitting method for injection wells considering time-varying factors according to claim 3, characterized in that, The specific process of step 2 is as follows: Step 2.1: Obtain relative permeability curve data, viscosity data, capillary pressure data, coordinate data of each well, and porosity and permeability data of the production well. Calculate the time-varying parameters of the production well, including the relative permeability, comprehensive viscosity, capillary pressure of the production well at each time point, and the pressure gradient between each injection and production well. Step 2.2: Based on the connectivity coefficients between injection and production wells obtained in Step 1, the time-varying parameter set of each injection well at each time point is obtained by weighting.

5. The intelligent water volume splitting method for injection wells considering time-varying factors according to claim 4, characterized in that, The specific process of step 2.1 is as follows: Step 2.1.1: Obtain relative permeability curve data and calculate the relative permeability and overall viscosity of the production well at various times. The specific process is as follows: First, normalize the relative permeability curve data and establish the relationship between relative permeability and water saturation using a power-law model. Based on the obtained relative permeability data, fit the relative permeability equation of the oil-water two-phase system. Then, obtain the water cut and oil-water two-phase viscosity data of the production well, calculate the water saturation according to the Buckley–Leverett water drive front equation, and then obtain the corresponding relative permeability and overall viscosity of the water phase based on the calculated water saturation. Step 2.1.2: Obtain capillary pressure curve data, porosity and permeability data of production well, and calculate oil-water capillary pressure of production well at each time point; convert the capillary pressure curve into a relationship curve between capillary pressure and water saturation: calculate the capillary pressure of production well based on the water saturation obtained in step 2.1.

1. Step 2.1.3: Based on the bottomhole flowing pressure of the injection and production wells obtained in Step 1, calculate the injection-production pressure gradient between each injection and production well. The calculation process for the injection-production pressure gradient in one time step is as follows: Based on the bottomhole flowing pressure data obtained in Step 1, calculate the pressure difference between each injection well and each production well; obtain the coordinates of each well and calculate the well distance between each injection well and each production well; divide the pressure difference by the well distance to obtain the injection-production pressure gradient between each injection and production well. Step 2.1.4: Repeat step 2.1.3 to calculate the injection-production pressure gradient between injection and production wells at all time steps.

6. The intelligent water volume splitting method for injection wells considering time-varying factors according to claim 5, characterized in that, The specific process of step 2.2 is as follows: After calculating the relative permeability, overall viscosity, and capillary pressure of the water phase at each moment in the production well, these three parameters are horizontally concatenated, and the data from all production wells are vertically concatenated to obtain a dimension of... Time-varying parameter matrix The connectivity coefficient between injection and production wells calculated in step 1 is used. Multiplying these parameters by the time-varying parameter matrix yields the relative permeability, overall viscosity, and capillary pressure of the water phase in the injection well. ； in, The parameters include the relative permeability, overall viscosity, and capillary pressure of the water phase at various times in the injection well. By concatenating the injection-production pressure gradient data from all time steps, a dimension of [dimensional value] is obtained. Injection-production pressure gradient data The connectivity coefficient between injection and production wells The pressure gradient of the injection well is obtained by weighted summation of the injection and production pressure gradient data: ； in, This represents the pressure gradient of the injection well at various times. For element-wise multiplication; This is matrix multiplication; It is a vector whose elements are all 1s; The time-varying parameter set of the injection well is obtained by stitching together the data of relative permeability, overall viscosity, capillary pressure, and pressure gradient of the water phase. ； in, This is a set of time-varying parameters for each moment in the injection well.

7. The intelligent water volume splitting method for injection wells considering time-varying factors according to claim 6, characterized in that, The specific process of step 3 is as follows: Step 3.1: Obtain perforation data, porosity, and permeability data for each injection well; based on the perforation data, determine the injection layer and corresponding perforation thickness of the injection layer at each time point; horizontally stitch together the perforation thickness, porosity, permeability, relative permeability of the water phase, overall viscosity, capillary pressure, injection-production pressure gradient, and injection volume data; vertically stitch together all sub-layers of the same injection well according to time steps to obtain the... The data dimensions of the injection well are ;in For the first The number of injection layers in the injection well; Step 3.2: Determine the maximum number of injection layers, set up virtual filling layers, and maintain the same dimensions for each injection well; The data filling method is used to fill in the number of injection layers in the combined injection wells. First, the maximum number of injection layers in the target block is determined. The input data dimension of the stratified water volume intelligent splitting model is: For injection wells with a smaller number of injection layers than the maximum number of injection layers, the missing layers are filled in the last layer, denoted as a virtual filling layer, where all parameters of the virtual filling layer are set to 0; the data from all injection wells are concatenated to obtain a dimension of The input dataset; Step 3.3: Obtain the water absorption profile data of the target block, specifically including the water absorption profile test time and the relative water absorption of each sub-layer. Based on the water absorption profile test time of each injection well, select the data of the corresponding time step in the input dataset of Step 3.2 as input, and use the relative water absorption as output to construct the dataset of the layered water volume intelligent splitting model.

8. The intelligent water volume splitting method for injection wells considering time-varying factors according to claim 7, characterized in that, The specific process of step 4 is as follows: Step 4.1: Construct a hierarchical intelligent water volume splitting model based on Transformer; the hierarchical intelligent water volume splitting model includes a data embedding module, an encoder module, and a decoder module; The data embedding module includes an embedding layer and positional encoding. The embedding layer consists of a fully connected layer, and the positional encoding uses absolute positional encoding. First, the input data is normalized. Then, the normalized data is passed through a fully connected layer and absolute positional encoding to obtain the output of the data embedding module. ; The encoder module consists of a multi-head self-attention layer, a layer normalized activation function, a feedforward neural network, and residual connections. The output of the data embedding module is input into the multi-head self-attention layer, and then passed through three fully connected layers to obtain the query matrix of the smaller layers. Key matrix Sum matrix Through self-attention mechanism and mask matrix Calculate the attention score for each injection layer : ； The output of self-attention is obtained by weighted summation of the value matrix of each sub-layer using attention scores. : ； The output of the multi-head attention layer is obtained by using a multi-head attention mechanism. : ； ； in, This is a multi-head attention mechanism; For the number of attention heads; , and The first A person's attention The query matrix, key matrix, and value matrix; These are the weights of the fully connected layer; For splicing operations; Add residual connections and layer normalization to the encoder module: ； in, This is the output after layer normalization; For layer normalization activation function; The feedforward neural network consists of two fully connected layers. Residual connections are used to add the input and output of the feedforward neural network. Finally, the output is normalized to obtain the output of the encoder module. : ； in, and These are the weight parameters of the fully connected layer; and These are the bias parameters for the fully connected layer; The decoder module consists of a single fully connected layer, which takes the output of the encoder module as input to obtain the predicted value of relative water absorption. ; Step 4.2: Divide the dataset from Step 3.3 into a training set and a validation set according to the proportion, set the basic parameters of the model, train the hierarchical water volume intelligent splitting model with the training set data, obtain the predicted value of relative water absorption of each layer through the decoder module, take the relative water absorption of the smaller layer as the label, use the mean squared error as the target loss function, calculate the error between the predicted value and the actual value of relative water absorption of each layer, calculate the weight gradient of each layer of the neural network based on the obtained error, and update the weights with the Adam algorithm until the error reaches the preset threshold and then stop training. After the model training is completed, the input dataset of all time steps of each injection well obtained in step 3.3 is input into the layered water volume intelligent splitting model to obtain the relative water absorption of each sub-layer under all time steps of each injection well.

Images