Oil reservoir development rule dynamic prediction method based on data driving

Abstract

The invention discloses an oil reservoir development law dynamic prediction method based on data driving, and the method comprises the steps: obtaining oil reservoir data, and carrying out the layer division and data processing; constructing a layered development rule dynamic prediction model, dividing data into a training set and a test set, and performing model training; longitudinal yield splitting is conducted, and the splitting yield coefficient of each small layer is calculated; determining a hierarchical development rule dynamic prediction model objective function, an optimization method and an optimization model; and according to the model prediction data index, calculating and analyzing the rule, and providing real-time and reliable dynamic index prediction for the oil reservoir. Static data and dynamic data of the oil reservoir are combined, and the characteristics and the operation state of the oil reservoir are comprehensively understood; each oil reservoir layer is predicted and analyzed more finely by constructing a layered development law dynamic prediction model; dynamic index prediction of oil reservoir development is generated in real time, oil field managers are helped to know oil reservoir development conditions in time, development strategies are adjusted in time, and the recovery efficiency and economic benefits are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of oil reservoir development law, and in particular relates to a data-driven dynamic prediction method for oil reservoir development law. Background Art

[0002] With the deepening of reservoir development, how to more effectively predict reservoir development laws has become an important topic. Traditional reservoir development law prediction methods often rely on empirical formulas and mathematical models, which have problems such as complex models and low prediction accuracy. Therefore, data-driven dynamic prediction methods for reservoir development laws have emerged. As a powerful data analysis tool, machine learning has broad application prospects in the field of reservoir development. Through machine learning algorithms, we can learn the patterns and laws hidden in massive reservoir data and realize dynamic prediction of reservoir development. Summary of the invention

[0003] The present invention is proposed to solve the problems of long calculation time and reliance on empirical formulas in the process of dynamic prediction of reservoir development by traditional reservoir engineering methods, and its purpose is to provide a data-driven dynamic prediction method for reservoir development laws.

[0004] The present invention is achieved through the following technical solutions:

[0005] A data-driven dynamic prediction method for reservoir development law includes the following steps:

[0006] S1. Acquire reservoir data, divide the reservoir data into layers and process the data;

[0007] S2. Build a hierarchical development law dynamic prediction model, divide the data into a training set and a test set in an 8:2 ratio, and perform model training;

[0008] S3, using the particle swarm optimization algorithm to split the production of the directional well vertically, and using the multi-layer optimization algorithm to calculate the splitting coefficient of each small layer;

[0009] S4. Determine the objective function and optimization method of the hierarchical development law dynamic prediction model, and use the optimization algorithm to optimize the model;

[0010] S5. Calculate and analyze the data indicators based on the model prediction to provide real-time and reliable dynamic indicator prediction for the reservoir.

[0011] In the above technical scheme, the reservoir data includes historical data, static data, dynamic data and small layer information data table; the static data includes logging interpretation data and perforation data; the logging interpretation data includes mud content, permeability, effective thickness, oil saturation, well name layer information and interpretation conclusions; the perforation data includes boundary oblique depth, top boundary depth, perforation oblique thickness, top boundary vertical depth, lower boundary vertical depth, effective vertical thickness and well name layer information; the dynamic data is production data; the production data includes water injection well data and production well data; the water injection well data includes daily injection volume and production time information; the production well data includes daily liquid production, daily oil production, daily water production, gas-oil ratio, bottom hole flow pressure, pump frequency and production time; the small layer information data table includes layer number and layer name.

[0012] In the above technical scheme, the layer segment division method of the reservoir data is as follows: first, according to the small layer information data table, the static data and dynamic data are preliminarily divided into layers; then, according to the well name layer segment information in the static data, the data is further divided into multiple segments according to the naming rules, and the dynamic data will be consistent with the static data according to the division results.

[0013] In the above technical solution, the data processing includes data cleaning and normalization processing; the data cleaning specifically processes missing values, abnormal values ​​and duplicate values ​​in the data.

[0014] In the above technical solution, the data cleaning specifically includes processing missing values, abnormal values ​​and duplicate values ​​in the data; and processing missing values ​​by combining deletion and filling.

[0015] In the above technical solution, the data processing method of the reservoir data is linear normalization processing; the conversion function formula of the linear normalization processing is as follows (1):

[0016]

[0017] Where: x' is the normalized sample data, the data range of x' is [0, 1]; x is the sample data; max(x) is the maximum value of the sample data; min(x) is the minimum value of the sample data.

[0018] In the above technical solution, step S2 specifically includes the following steps:

[0019] S21. Determine the main controlling factors of development indicators at different levels during crude oil production

[0020] The Spearman coefficient and Pearson coefficient of each factor were calculated, with the Spearman coefficient as the main factor and the Pearson coefficient as the supplement. The model threshold was set to 0.4 to screen out the main controlling factors.

[0021] The calculation formula of the Spearman coefficient is as follows (2):

[0022]

[0023] Where: ρ is the Spearman coefficient of factor x and factor y, which is used to measure the correlation between the two. Its value range is [-1,1]. The closer it is to 1, the greater the positive correlation between factor x and factor y. The closer it is to -1, the greater the negative correlation between factor x and factor y. The closer it is to 0, the smaller the correlation. i is the data of factor x, is the mean of factor x; y i is the data of factor y, is the mean of factor y;

[0024] The calculation formula of the Pearson coefficient is as follows (3):

[0025]

[0026] Where: X,Y is the Pearson coefficient of factor x and factor y, which is used to measure the linear correlation between the two. Its value range is [-1,1]. The closer it is to 1, the greater the positive linear correlation between factor x and factor y. The closer it is to -1, the greater the negative linear correlation between factor x and factor y. The closer it is to 0, the smaller the correlation. cov(X,Y) is the covariance between factor x and factor y. X , σ Y are the standard deviations of factor x and factor y, μ X , μ Y are the means of factors x and y respectively;

[0027] S22. A hierarchical prediction method with embedded conduction factors in each direction is proposed to build a dynamic prediction model for hierarchical development rules. A neural network algorithm is used to calculate flow on the plane. The model input parameters are injection and production data with known observation values, and the output parameters are production data. The conduction factors are used to characterize the injection and production relationship. In the model, the relationship equation for conduction in each direction is:

[0028]

[0029] Where: C t is the total compressibility factor, in Pa -1 , V p is the drainage pore volume in m 3 ; Yes V p The average pressure in Pa; t is the time step in seconds, and X(t) is the injection rate in m 3 / s; Y(t) is the liquid production rate, in m 3 / s; when the total compression factor is less than 10 -6 Pa -1 It can be assumed that the compression effect of the fluid is minimal, resulting in almost no inflow or outflow of liquid;

[0030] The model infers the injection-production relationship between the injection well and the production well through the transmission factor. The input parameter data is defined as follows after the transmission factor:

[0031]

[0032] Where: ε ij is the injection-production relationship matrix σ ij , which is the independent variable of the conduction factor;

[0033] Due to the multiple solutions of injection-production relationship analysis, uniformly distributed random values ​​are used to initialize model parameters. By calculating the inverse of the matrix elements, we can get:

[0034]

[0035] Where: A is the number of injection wells, B is the number of production wells, ρ ij is the inverse of a random value uniformly distributed between the injection well and the production well;

[0036] In the layered development law dynamic prediction model, a series of fully connected layers are used to approximate the fluid change rate in the control volume; in actual reservoirs, linear models cannot accurately predict the nonlinear flow characteristics in reservoirs; therefore, the nonlinear mapping ability of neural networks combined with multiple nonlinear layers can better simulate the target output; therefore, nonlinear mapping is used to represent nonlinear relationships:

[0037] Q=N(mP)......(7)

[0038] Where: Q represents production, m is a constant, and P is pressure. In actual oil fields, m is usually not available, so the differential form of the nonlinear relationship is:

[0039]

[0040] At constant pressure, this can be simplified to:

[0041]

[0042] Multiply both sides of the above equation by C Vj We can get:

[0043]

[0044] For production well j, considering B production wells, the fully connected layer neural network is used for approximation:

[0045]

[0046] Where: Net represents the connection layer in the network. The flow characteristics in the reservoir are simulated through nonlinear mapping relationships and multiple connection layers. The nonlinear ability of the neural network is used to better reflect the complex dynamic behavior of the reservoir.

[0047] A production module is used to calculate the output of the production well. According to the mobility calculation of the neural network model and the hierarchical prediction method of the conduction factors in each direction, the output of the production well can be expressed as:

[0048]

[0049] Where: represents the predicted data of production well j, T j represents the real data of production well j, Net(Q j ) represents the data predicted by the connected layers in the network;

[0050] S23, dividing the data processed in step S1 into a training set (80%) and a test set (20%) according to the layer segments.

[0051] In the above technical solution, the speed update formula of the particles in the particle swarm optimization algorithm in step S3 is:

[0052]

[0053] Where: is the velocity of particle i at the next iteration t+1, is the velocity of particle i at the current iteration t, pbest i The individual best position found by particle i so far, gbest is the best position found by the entire particle group so far, is the position of particle i at the current iteration t;

[0054] The position update formula is:

[0055]

[0056] Where: The position of particle i at the next iteration t+1, is the velocity of particle i at the next iteration t+1, is the position of particle i at the current iteration t.

[0057] In the above technical solution, step S4 specifically includes the following steps:

[0058] S41. Use the mean square error (MSE) function as the loss function. The expression of the loss function is as follows (15):

[0059]

[0060] Where: t is the time step, T is the total number of time steps;

[0061] The loss function of the model for the B production well in each layer is expressed as follows (16):

[0062]

[0063] Assume that the actual reservoir is divided into C layers, with A injection well and B production wells. In the neural network, constraints are added to the process of splitting the transfer factor so that the flow of the injection well is completely split to the production well to ensure the correct mobility calculation. The constraint expression is as follows (17):

[0064]

[0065] Where: c is an A×B matrix, the element in the ath row and bth column is I is a vector of all 1s, with lengths A and B respectively;

[0066] Add constraints to the production splitting using the PSO algorithm in the vertical direction so that the production of the directional well meets the production balance in each layer. The expression of the vertical production splitting constraint is as follows (18):

[0067] Γ c I=I......(18)

[0068] Where: c is an A×B matrix, the element in the ath row and bth column is I is a vector of all 1s and length B;

[0069] S42, determine the objective function of the model according to the loss function, the neural network and the constraints in the vertical direction. After adding the constraints, the objective function is minimized. The expression of the objective function of the hierarchical development law dynamic prediction model is as follows (19):

[0070]

[0071] In the above technical solution, step S5 specifically includes the following steps:

[0072] S51. Use mean absolute percentage error to measure the accuracy of the prediction model;

[0073] The calculation formula of the accuracy is as follows (20):

[0074]

[0075] Where: n represents the number of samples, y i represents the actual value of the i-th sample, Represents the predicted value of the i-th sample. The smaller the MAPE value, the more accurate the prediction, and the larger the MAPE value, the greater the error in the model.

[0076] S52. Take measures to optimize oilfield production strategies based on model prediction results and regularity analysis.

[0077] In the above technical solution, the measures for optimizing the oilfield production strategy are specifically:

[0078] (a) By dynamically adjusting the water injection volume and zoning water injection;

[0079] (b) Optimize the layout of new wells and existing well locations based on reservoir geological models and production data, and use horizontal well and branch well technologies to improve recovery;

[0080] (c) Adjust production pressure difference and wellhead flow rate and adopt intermittent production strategy to achieve the best production effect;

[0081] (d) Improve the comprehensive development effect of oil reservoirs by optimizing the injection-production ratio, deploying intelligent monitoring systems and implementing multi-layer combined production technology.

[0082] The beneficial effects of the present invention are:

[0083] The present invention provides a data-driven dynamic prediction method for reservoir development laws. By combining the static data (such as geological characteristics, permeability, porosity, etc.) and dynamic data (such as production data, injection and production information, etc.) of the reservoir, the characteristics and operating status of the reservoir can be fully understood, thereby improving the accuracy and comprehensiveness of the prediction model; by constructing a layered dynamic prediction model for development laws, each reservoir layer can be predicted and analyzed more finely, dynamic monitoring and prediction of reservoir development laws can be achieved, and a more specific basis for decision-making can be provided; dynamic indicator predictions for reservoir development can be generated in real time, helping oilfield managers to timely understand the reservoir development situation, adjust development strategies in a timely manner, and improve recovery rates and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 is a flow chart of the method of the present invention;

[0085] Figure 2 It is a structural diagram of the hierarchical development law dynamic prediction model of the present invention;

[0086] Figure 3is a flow chart of the longitudinal yield splitting optimization algorithm of the present invention;

[0087] Figure 4 It is a model structure diagram of the longitudinal production splitting optimization algorithm of the present invention.

[0088] For ordinary technicians in this field, other relevant drawings can be obtained based on the above drawings without any creative work. DETAILED DESCRIPTION

[0089] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is further described below with reference to the accompanying drawings and through specific implementation methods.

[0090] like Figure 1 As shown, a data-driven dynamic prediction method for reservoir development law includes the following steps:

[0091] S1. Acquire reservoir data, divide the reservoir data into layers and process the data;

[0092] The reservoir data includes historical data, static data, dynamic data and small layer information data table;

[0093] The static data includes logging interpretation data and perforation data; the logging interpretation data includes mud content, permeability, effective thickness, oil saturation, well name and layer information and interpretation conclusion; the perforation data includes boundary oblique depth, top boundary depth, perforation oblique thickness, top boundary vertical depth, lower boundary vertical depth, effective vertical thickness and well name and layer information;

[0094] The dynamic data is production data; the production data includes water injection well data and production well data; the water injection well data includes daily injection volume and production time information; the production well data includes daily liquid production, daily oil production, daily water production, gas-oil ratio, bottom hole flow pressure, pump frequency and production time;

[0095] The small layer information data table includes layer number and layer name;

[0096] The reservoir data segment division method is as follows: first, the static data and dynamic data are preliminarily segmented according to the small layer information data table; then, according to the well name segment information in the static data, the data is further segmented into multiple segments according to the naming rules, and the dynamic data will be consistent with the static data according to the segmentation result;

[0097] The data processing includes data cleaning and normalization processing;

[0098] Data cleaning is performed on all the data after division. There will be missing values, abnormal values ​​and duplicate values ​​in the collected historical data. In order to ensure the quality and accuracy of the data, the collected data set needs to be cleaned and the numerical values ​​with problems are processed; missing values ​​refer to the phenomenon that some parameters of some data are missing. This type of data will cause the model to report errors during prediction and even affect the prediction effect. Therefore, it cannot be applied to the model. Generally, a combination of deletion and filling is used to process missing values; duplicate values ​​usually refer to exactly the same row records appearing in the data set. Duplicate values ​​are usually deleted or merged based on keywords and column contents; abnormal values ​​refer to values ​​that are significantly different from the rest of the data. If it can be determined that the abnormal value is wrong and accounts for a small proportion, it can be considered to be directly deleted, or statistical methods and prediction models can be used for further correction. The present invention adopts a processing method that combines deletion and filling; for a piece of data, if there are too many missing values, the entire piece of data is directly deleted; otherwise, the mean is used to fill in the piece of data.

[0099] After data cleaning, normalization is required. Normalization methods include but are not limited to linear normalization, Z-score normalization, mean normalization, and nonlinear normalization.

[0100] In this embodiment, the data processing method of the reservoir data is linear normalization processing; the conversion function formula of the linear normalization processing is as follows (1):

[0101]

[0102] Where: x' is the normalized reservoir data, and the value range of x' is [0, 1]; x is the reservoir data; max(x) is the maximum value of the reservoir data; min(x) is the minimum value of the reservoir data;

[0103] Linear normalization can achieve the following goals:

[0104] Dimensional unification: Convert dimensional data into dimensionless data to eliminate the dimensional differences between different features;

[0105] Data standardization: Map data of different magnitudes into a unified range to eliminate magnitude differences between features;

[0106] Improve algorithm convergence: After the discrete data is normalized, the process of the model seeking the optimal solution will be smoother, which is conducive to the algorithm to converge to the optimal solution quickly.

[0107] S2. Build a hierarchical development law dynamic prediction model, divide the data into a training set and a test set according to an 8:2 ratio, and perform model training; specifically, the following steps are included:

[0108] S21. Determine the main controlling factors of development indicators at different levels during crude oil production;

[0109] The factors include but are not limited to daily fluid production and water cut at the single well level, oil change rate and injection-production ratio at the well group level, and water drive index at the block level;

[0110] Calculate the Spearman coefficient and Pearson coefficient of each factor, with the Spearman coefficient as the main factor and the Pearson coefficient as the supplement. Set the model threshold to 0.4, and select factors with both the Spearman coefficient and the Pearson coefficient threshold greater than 0.4 as the main control factors;

[0111] The calculation formula of the Spearman coefficient is as follows (2):

[0112]

[0113] Where: ρ is the Spearman coefficient of factor x and factor y, which is used to measure the correlation between the two. Its value range is [-1,1]. The closer it is to 1, the greater the positive correlation between factor x and factor y. The closer it is to -1, the greater the negative correlation between factor x and factor y. The closer it is to 0, the smaller the correlation. i is the data of factor x, is the mean of factor x; y i is the data of factor y, is the mean of factor y;

[0114] The calculation formula of the Pearson coefficient is as follows (3):

[0115]

[0116] Where: X,Y is the Pearson coefficient of factor x and factor y, which is used to measure the linear correlation between the two. Its value range is [-1,1]. The closer it is to 1, the greater the positive linear correlation between factor x and factor y. The closer it is to -1, the greater the negative linear correlation between factor x and factor y. The closer it is to 0, the smaller the correlation. cov(X,Y) is the covariance between factor x and factor y. X , σ Y are the standard deviations of factor x and factor y, μ X , μ Y are the means of factors x and y respectively;

[0117] Aiming at the data analysis needs in the oilfield development process, the present invention proposes a data-driven feature selection method, which aims to determine the main controlling factors of development indicators at different levels and establish an oil reservoir database covering production data, well-to-well relationships and development indicators. Specifically, the method includes the following steps:

[0118] Collect and integrate multi-source reservoir data such as static data, dynamic data, well characteristics, and well interactions.

[0119] The Spearman correlation coefficient method and Pearson correlation coefficient method were used to analyze the correlation between various indicators and development strategies, and to determine the main controlling factors of the production capacity of different types of oil reservoirs.

[0120] Based on the feature selection results, important reservoir-related parameters are screened out, irrelevant parameters are removed, noise interference is reduced, and the performance of the prediction model is improved.

[0121] Construct a reservoir database covering production data, well-to-well relationships and development indicators to provide effective support for subsequent reservoir development decisions.

[0122] Among them, the Spearman correlation coefficient is a nonparametric correlation analysis method used to evaluate the monotonic relationship between two variables. It does not require a linear relationship between the variables, but focuses on whether they have the same ordering method. It is more suitable for analyzing nonlinear relationships, situations with outliers, or situations where the data distribution does not satisfy the normal distribution. It evaluates the strength of the association between the two variables by analyzing the consistency of their numerical values.

[0123] The Pearson correlation coefficient method is a commonly used linear correlation analysis method used to measure the strength and direction of the linear correlation between two continuous variables. It is calculated based on the covariance and standard deviation between the variables, reflecting the correlation between the two variables in a linear relationship. Unlike the Spearman correlation coefficient, the Pearson correlation coefficient requires the variables to meet the normal distribution assumption and is suitable for analyzing linear correlations. When there is a strong linear correlation between two variables, the Pearson correlation coefficient can more accurately reflect the degree of correlation between them.

[0124] The main controlling factor analysis can identify the key features that most significantly affect the change in reservoir productivity, which is conducive to subsequent modeling and prediction. The Spearman correlation coefficient is mainly used for feature selection because it can capture nonlinear monotonic relationships and is more suitable for complex reservoir systems. The Pearson correlation coefficient is used as a supplement to identify linearly related features and further enrich feature selection. Setting a certain correlation threshold (such as 0.4) to screen features can effectively remove redundant and irrelevant data features and reduce interference with model prediction.

[0125] S22. Construct a dynamic prediction model for hierarchical development rules: A hierarchical prediction method is proposed that embeds conduction factors in each direction. A neural network algorithm is used on the plane to calculate flow. The model input parameters are injection and production data with known observation values, and the output parameters are production data. Conduction factors are used to characterize the injection-production relationship. In the model, the relationship equation for conduction in each direction is:

[0126]

[0127] Where: C t is the total compressibility factor, in Pa -1 , V p is the drainage pore volume in m 3 ; Yes V p The average pressure in Pa; t is the time step in seconds, and X(t) is the injection rate in m 3 / s; Y(t) is the liquid production rate, in m 3 / s; when the total compression factor is less than 10 -6 Pa -1 It can be assumed that the compression effect of the fluid is minimal, resulting in almost no inflow or outflow of liquid;

[0128] The model infers the injection-production relationship between the injection well and the production well through the transmission factor. The input parameter data is defined as follows after the transmission factor:

[0129]

[0130] Where: ε ij is the injection-production relationship matrix σ ij , which is the independent variable of the conduction factor;

[0131] Due to the multiple solutions of injection-production relationship analysis, uniformly distributed random values ​​are used to initialize model parameters. By calculating the inverse of the matrix elements, we can get:

[0132]

[0133] Where: A is the number of injection wells, B is the number of production wells, ρ ij is the inverse of a random value uniformly distributed between the injection well and the production well;

[0134] In the layered development law dynamic prediction model, a series of fully connected layers are used to approximate the fluid change rate in the control volume; in actual reservoirs, linear models cannot accurately predict the nonlinear flow characteristics in reservoirs; therefore, the nonlinear mapping ability of neural networks combined with multiple nonlinear layers can better simulate the target output; therefore, nonlinear mapping is used to represent nonlinear relationships:

[0135] Q=N(mP)......(7)

[0136] Where: Q represents production, m is a constant, and P is pressure. In actual oil fields, m is usually not available, so the differential form of the nonlinear relationship is:

[0137]

[0138] At constant pressure, this can be simplified to:

[0139]

[0140] Multiply both sides of the above equation by C Vj We can get:

[0141]

[0142] For production well j, considering B production wells, the fully connected layer neural network is used for approximation:

[0143]

[0144] Where: Net represents the connection layer in the network. The flow characteristics in the reservoir are simulated through nonlinear mapping relationships and multiple connection layers. The nonlinear ability of the neural network is used to better reflect the complex dynamic behavior of the reservoir.

[0145] A production module is used to calculate the output of the production well. The output of the production well is obtained based on the mobility calculation of the neural network model and the hierarchical prediction method of the conduction factors in each direction. That is, the expression of the hierarchical development law dynamic prediction model is:

[0146]

[0147] Where: represents the predicted data of production well j, T j represents the real data of production well j, Net(Q j ) represents the data predicted by the connected layers in the network;

[0148] S23, dividing the data processed in step S1 into a training set (80%) and a test set (20%) according to the layer segments.

[0149] S3, using the particle swarm optimization algorithm to split the production of the directional well vertically, and using the multi-layer optimization algorithm to calculate the splitting coefficient of each small layer;

[0150] The particle velocity update formula is:

[0151]

[0152] Where: is the velocity of particle i at the next iteration t+1, is the velocity of particle i at the current iteration t, pbest i The individual best position found by particle i so far, gbest is the best position found by the entire particle group so far, is the position of particle i at the current iteration t.

[0153] The position update formula is:

[0154]

[0155] Where: The position of particle i at the next iteration t+1, is the velocity of particle i at the next iteration t+1, is the position of particle i at the current iteration t.

[0156] There is a variable in the particle velocity update formula, such as inertia weight w, individual learning factor c 1 and c 2 , and the random number r 1 and r 2 Among them, the inertia weight w controls the ability of the particle to maintain the current flight direction. A larger w value can enhance the global search ability. The individual learning factor determines the degree to which the particle is affected by the individual and the group. The two factors jointly affect the speed of particle update. The random number r 1 and r 2 This is to increase the randomness of the algorithm so that particles can search all areas of the world.

[0157] The present invention uses a particle swarm optimization (PSO) algorithm to regard each particle as a potential solution for the splitting coefficient. The particle continuously updates its speed and position based on its own experience (individual optimality) and group experience (global optimality), and finally converges to the global optimal splitting coefficient combination.

[0158] The PSO algorithm is simple and easy to implement for vertical production splitting of directional wells. It does not have too many parameters and does not require gradient information. It is an effective optimization tool for nonlinear continuous optimization problems and has been widely used in neural network training.

[0159] S4. Determine the objective function and optimization method of the hierarchical development law dynamic prediction model, and use the optimization algorithm to optimize the model, which specifically includes the following steps:

[0160] S41. Use the mean square error (MSE) function as the loss function. The expression of the loss function is as follows (15):

[0161]

[0162] Where: t is the time step, T is the total number of time steps;

[0163] The loss function of the model for the B production well in each layer is expressed as follows (16):

[0164]

[0165] Assume that the actual reservoir is divided into C layers, with A injection well and B production wells. In the neural network, constraints are added to the process of splitting the transfer factor so that the flow of the injection well is completely split to the production well to ensure the correct mobility calculation. The expression of the constraints added to the process of splitting the transfer factor is as follows (17):

[0166]

[0167] Where: c is an A×B matrix, the element in the ath row and bth column is I is a vector of all 1s, with lengths A and B respectively;

[0168] Add constraints to the production splitting using the PSO algorithm in the vertical direction so that the production of the directional well meets the production balance in each layer. The expression of the vertical production splitting constraint is as follows (18):

[0169] Γ c I=I......(18)

[0170] Where: c is an A×B matrix, the element in the ath row and bth column is I is a vector of all 1s and length B;

[0171] S42, determine the objective function of the model according to the loss function, the neural network and the constraints in the vertical direction. After adding the constraints, the objective function is minimized. The expression of the objective function of the hierarchical development law dynamic prediction model is as follows (19):

[0172]

[0173] During the model training process, the optimization algorithm Adam is used to update and calculate the parameters in the model to minimize the loss function. The Adam algorithm combines the ideas of momentum method and adaptive learning rate. It calculates the first-order moment estimate and second-order moment estimate of the gradient at each iteration and adjusts the learning rate based on them. In deep learning, the Adam algorithm is widely used in the model training process, which can effectively accelerate training and achieve better performance.

[0174] S5. Calculate and analyze the data indicators predicted by the model to provide real-time and reliable dynamic indicator prediction for the reservoir, which specifically includes the following steps:

[0175] S51. Use the mean absolute percentage error (MAPE) to measure the accuracy of the prediction model, which is used to determine the relative error between the prediction results and the actual data;

[0176] The calculation formula of the accuracy is as follows (20):

[0177]

[0178] Where: n represents the number of samples, y i represents the actual value of the i-th sample, Represents the predicted value of the i-th sample. The smaller the MAPE value, the more accurate the prediction, and the larger the MAPE value, the greater the error in the model.

[0179] S52. Based on the model prediction results and regularity analysis, a series of measures are taken to optimize the oilfield production strategy to improve the recovery rate and economic benefits.

[0180] Firstly, the water injection rate is dynamically adjusted and divided into different zones to ensure that the water injection rate matches the absorptive capacity of the reservoir, and the optimal water injection rate is determined using a water injection optimization model.

[0181] Secondly, based on the reservoir geological model and production data, the layout of new wells and the location of existing wells are optimized, and horizontal well and branch well technologies are used to improve the recovery rate.

[0182] Again, adjust the production pressure difference and wellhead flow, and adopt an intermittent production strategy to achieve the best production effect.

[0183] In addition, the comprehensive development effect of the oil reservoir can be improved by optimizing the injection-production ratio, deploying an intelligent monitoring system and implementing multi-layer combined production technology.

[0184] The above measures achieve real-time monitoring and adjustment through scientific management and technical means, ensuring the flexibility and adaptability of production strategies, thereby maximizing the development effect and economic benefits of the oil field.

[0185] The present invention proposes a data-driven dynamic prediction method for reservoir development laws. In this machine learning-based method, a data set including geological parameters, engineering parameters, production data and other information is first established, and a development dynamic prediction model is constructed using a machine learning algorithm to predict reservoir dynamic indicators. The machine learning model can be continuously updated and optimized based on real-time or historical data to achieve dynamic prediction of reservoir development laws.

[0186] The applicant declares that the above is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention shall fall within the protection scope and disclosure scope of the present invention.

Claims

1. A data-driven dynamic prediction method for reservoir development law, characterized by: The following steps are involved: S1. Acquire reservoir data, divide the reservoir data into layers and process the data; S2. Build a hierarchical development law dynamic prediction model, divide the data into a training set and a test set in an 8:2 ratio, and perform model training; S3, using the particle swarm optimization algorithm to split the production of the directional well vertically, and using the multi-layer optimization algorithm to calculate the splitting coefficient of each small layer; S4. Determine the objective function and optimization method of the hierarchical development law dynamic prediction model, and use the optimization algorithm to optimize the model; S5. Calculate and analyze the data indicators based on the model prediction to provide real-time and reliable dynamic indicator prediction for the reservoir.

2. The data-driven dynamic prediction method for reservoir development law according to claim 1 is characterized by: The reservoir data includes historical data, static data, dynamic data and small layer information data table; the static data includes logging interpretation data and perforation data; the logging interpretation data includes mud content, permeability, effective thickness, oil saturation, well name layer information and interpretation conclusions; the perforation data includes boundary oblique depth, top boundary depth, perforation oblique thickness, top boundary vertical depth, lower boundary vertical depth, effective vertical thickness and well name layer information; the dynamic data is production data; the production data includes water injection well data and production well data; the water injection well data includes daily injection volume and production time information; the production well data includes daily liquid production, daily oil production, daily water production, gas-oil ratio, bottom hole flow pressure, pump frequency and production time; the small layer information data table includes layer number and layer name.

3. The data-driven dynamic prediction method for reservoir development law according to claim 1 is characterized in that: The layer segment division method of the reservoir data is as follows: first, according to the small layer information data table, the static data and dynamic data are preliminarily divided into layers; then, according to the well name layer segment information in the static data, the data is further divided into multiple segments according to the naming rules, and the dynamic data will be consistent with the static data according to the division results.

4. The data-driven dynamic prediction method for reservoir development rules according to claim 1 is characterized by: The data processing includes data cleaning and normalization processing; the data cleaning specifically processes missing values, outliers and duplicate values ​​in the data.

5. The data-driven dynamic prediction method for reservoir development rules according to claim 4 is characterized by: The normalization processing method is a linear normalization processing; the conversion function formula of the linear normalization processing is as follows (1): In the formula: x' is the normalized data, the data range of x' is [0, 1]; x is the data; max(x) is the maximum value of the data; min(x) is the minimum value of the data.

6. The data-driven dynamic prediction method for reservoir development law according to claim 1 is characterized by: The step S2 specifically includes the following steps: S21. Determine the main controlling factors of development indicators at different levels during crude oil production Calculate the Spearman coefficient and Pearson coefficient of each factor, with the Spearman coefficient as the main factor and the Pearson coefficient as the supplement. Set the model threshold to 0.4, and select factors with both the Spearman coefficient and the Pearson coefficient greater than the model threshold as the main control factors; The calculation formula of the Spearman coefficient is as follows (2): Where: ρ is the Spearman coefficient of factor x and factor y, ranging from [-1,1]; x i is the data of factor x, is the mean of factor x; y i is the data of factor y, is the mean of factor y; The calculation formula of the Pearson coefficient is as follows (3): Where: X,Y is the Pearson coefficient of factor x and factor y, and its value range is [-1,1]; cov(X,Y is the covariance of factor x and factor y, σ X , σ Y are the standard deviations of factor x and factor y, μ X , μ Y are the means of factor x and factor y respectively; S22. A hierarchical prediction method with embedded conduction factors in each direction is proposed to build a dynamic prediction model for hierarchical development rules. A neural network algorithm is used to calculate flow on the plane. The model input parameters are injection and production data with known observation values, and the output parameters are production data. The conduction factors are used to characterize the injection and production relationship. In the model, the relationship equation for conduction in each direction is: Where: C t is the total compressibility factor, in Pa -1 , V p is the drainage pore volume in m 3 ; Yes V p The average pressure in Pa; t is the time step in seconds, and X(t) is the injection rate in m 3 / s; Y(t) is the liquid production rate, in m 3 / s; when the total compression factor is less than 10 -6 Pa -1 It can be assumed that the compression effect of the fluid is minimal, resulting in almost no inflow or outflow of liquid; The model infers the injection-production relationship between the injection well and the production well through the transmission factor. The input parameter data is defined as follows after the transmission factor: Where: ε ij is the injection-production relationship matrix σ ij , which is the independent variable of the conduction factor; The model parameters are initialized with uniformly distributed random values, and the inverse of the matrix elements is calculated to obtain: Where: A is the number of injection wells, B is the number of production wells, ρ ij is the inverse of a random value uniformly distributed between the injection well and the production well; Nonlinear relationships are expressed using nonlinear mappings. The specific expressions are: Q=N(mP)......(7) Where: Q represents production; m is a constant; P is pressure; in actual oil fields, m is usually not available, so the differential form of the nonlinear relationship is: At constant pressure, this simplifies to: Multiply both sides of the above equation by C Vj have to: For production well j, considering B production wells, a fully connected neural network is used for approximation: Where: Net represents the connection layer in the network; A production module is used to calculate the output of the production well. According to the mobility calculation of the neural network model and the hierarchical prediction method of the conduction factors in each direction, the output of the production well can be expressed as: Where: represents the predicted data of production well j, T j represents the real data of production well j, Net(Q j ) represents the data predicted by the connected layers in the network; S23, dividing the data processed in step S1 into a training set (80%) and a test set (20%) according to the layer segments.

7. The data-driven dynamic prediction method for reservoir development law according to claim 1 is characterized by: The speed update formula of the particles in the particle swarm optimization algorithm in step S3 is: Where: is the velocity of particle i at the next iteration t+1, is the velocity of particle i at the current iteration t, pbest i The individual best position found by particle i so far, gbest is the best position found by the entire particle group so far, is the position of particle i at the current iteration t; The position update formula is: Where: The position of particle i at the next iteration t+1, is the velocity of particle i at the next iteration t+1, is the position of particle i at the current iteration t.

8. The data-driven dynamic prediction method for reservoir development rules according to claim 1 is characterized by: The step S4 specifically comprises the following steps: S41, using the mean square error function as the loss function, the loss function is expressed as follows (15): Where: t is the time step, T is the total number of time steps; The loss function of the model for the B production well in each layer is expressed as follows (16): Assume that the actual reservoir is divided into C layers, with A injection well and B production wells. In the neural network, constraints are added to the process of splitting the transfer factor so that the flow of the injection well is completely split to the production well. The expression of the constraints added to the process of splitting the transfer factor is as follows (17): Where: c is an A×B matrix, the element in the ath row and bth column is I is a vector of all 1s, with lengths A and B respectively; Add constraints to the production splitting using the PSO algorithm in the vertical direction so that the production of the directional well meets the production balance in each layer. The expression of the vertical production splitting constraint is as follows (18): Γ c I=I......(18) Where: c is an A×B matrix, the element in the ath row and bth column is I is a vector of all 1s and length B; S42, determine the objective function of the model according to the loss function, the neural network and the constraints in the vertical direction. After adding the constraints, the objective function is minimized. The expression of the objective function of the hierarchical development law dynamic prediction model is as follows (19):

9. The data-driven dynamic prediction method for reservoir development law according to claim 1 is characterized by: The step S5 specifically The following steps are involved: S51. Use mean absolute percentage error to measure the accuracy of the prediction model; The calculation formula of the accuracy is as follows (20): Where: MAPE represents accuracy; n represents the number of samples; y i represents the actual value of the i-th sample; Represents the predicted value of the i-th sample; S52. Take measures to optimize oilfield production strategies based on model prediction results and regularity analysis.

10. The data-driven dynamic prediction method for reservoir development rules according to claim 9, characterized in that: The measures for optimizing oilfield production strategy are specifically: (a) By dynamically adjusting the water injection volume and zoning water injection; (b) Optimize the layout of new wells and existing well locations based on reservoir geological models and production data, and use horizontal well and branch well technologies to improve recovery; (c) Adjust production pressure difference and wellhead flow rate and adopt intermittent production strategy to achieve the best production effect; (d) Improve the comprehensive development effect of oil reservoirs by optimizing the injection-production ratio, deploying intelligent monitoring systems and implementing multi-layer combined production technology.

Images