A well test interpretation method, medium and equipment automatically adapting to different well test models
Through the Meta-LSTM neural network model, the problem of frequent replacement of models in well test interpretation is solved, and efficient interpretation of wells of different reservoir properties is achieved, task volume and cost are reduced, and the task volume and cost are improved. The efficiency of oil and gas field development.
Patent Information
- Application Number
- CN202411059879.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-05
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-08-05
AI Technical Summary
The existing technology requires frequent replacement of the well test model in the well test interpretation, which leads to the large task volume and low efficiency when interpreting wells of different reservoir properties, making it difficult to automatically adapt to the explanation of different well test models.
Using the Meta-LSTM neural network model, the training set and test set are constructed by collecting and processing the well test model feature curve and real test well curve, and the adaptive test well interpretation model is trained, and the feature curve and explanation of the well test model are automatically identified and fitted. Parameters to reduce the amount of tasks of well test interpretation work.
Automatic adaptation to different well test models is achieved, the task volume and cost of well test interpretation work is reduced, the efficiency and accuracy of well test interpretation is improved, and greater help is provided for oil and gas field development.
Smart Images

Figure CN118606683B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of oil and gas field development, and in particular to a well testing interpretation method, medium and equipment which are automatically adapted to different well testing models. Background Art
[0002] Well test interpretation technology can quickly and accurately obtain the formation parameters of oil and gas reservoirs, providing great help for the economic and efficient development of oil and gas fields. The efficiency of manual interpretation of well test data specific to different oil and gas reservoirs is often low.
[0003] In recent years, various machine learning algorithms have been applied to well test interpretation to improve the efficiency and accuracy of well test interpretation. However, due to the large differences in reservoir properties and different fracturing parameters, a well test model is often only applicable to one type of reservoir properties. Therefore, when using existing machine learning methods to interpret well tests for wells with different reservoir properties, it is still necessary to reselect a specific well test model for repeated interpretation, and the workload of well test interpretation for multiple wells has not been substantially reduced. Therefore, establishing an interpretation method that automatically adapts to different well test models is of great significance to guiding the efficient development of oil and gas fields. Summary of the invention
[0004] The purpose of the present invention is to provide a well test interpretation method, medium and equipment that can automatically adapt to different well test models to obtain accurate well test interpretation parameters and reduce the workload of well test interpretation work.
[0005] To achieve the above object, the present invention provides the following solution.
[0006] In one aspect, the present invention provides a well test interpretation method that automatically adapts to different well test models, comprising:
[0007] Collecting different well test models of reservoir wells and corresponding well test data; the different well test models of reservoir wells are well test models of reservoir wells with different reservoir properties and different boundary conditions; the well test data include instantaneous bottom hole flow pressure data within a preset time period;
[0008] Solve each well test model to obtain a well test model characteristic curve, and plot the corresponding well test data as an actual well test curve;
[0009] Construct training set and test set based on well test model characteristic curve and actual test well curve;
[0010] A Meta-LSTM neural network model is constructed for identifying and interpreting different well test models; the Meta-LSTM neural network model is connected and integrated by an input layer, an LSTM module, a task representation module, a meta-learning neural network, a task-specific layer, and an output layer through a functional API; the input of the Meta-LSTM neural network model is an actual well test curve, and the output is a fitted well test model characteristic curve and a well test interpretation parameter;
[0011] The training set is used to train the parameters of the Meta-LSTM neural network model, and the test set is used to evaluate the performance of the Meta-LSTM neural network model. After the training is completed, an adaptive well test interpretation model is obtained;
[0012] An adaptive well test interpretation model is used to perform well test interpretation to obtain well test interpretation parameters of the reservoir well; the well test interpretation parameters include reservoir length, width and height, porosity, permeability, formation coefficient, formation flow coefficient, skin coefficient, comprehensive compression coefficient, and length, width and number of hydraulic fractures.
[0013] Optionally, solving each well test model to obtain a well test model characteristic curve, and drawing the corresponding well test data as an actual well test curve, specifically includes:
[0014] Solve each well test model to obtain different double logarithmic characteristic curves of dimensionless pressure and pressure derivative with respect to dimensionless time as characteristic curves of the well test model;
[0015] The corresponding well test data are also plotted as a double logarithmic characteristic curve of dimensionless pressure and pressure derivative with respect to dimensionless time as the actual test well curve.
[0016] Optionally, constructing a training set and a test set based on the well test model characteristic curve and the actual test well curve specifically includes:
[0017] Identifying a real test well curve is regarded as a task, the well test model characteristic curve corresponding to each task curve is used as a label, and the task curve and the corresponding label are used as a sample pool;
[0018] For each task curve in the sample pool, noise point data is randomly added, and a data set is generated together with the data on the task curve. The data set is divided into a training set and a test set according to a preset ratio.
[0019] Optionally, the input layer of the Meta-LSTM neural network model is a multimodal input layer; the multimodal input layer includes branches of independent convolutional neural network structures and branches of recurrent neural network structures; the feature data in the input task data is passed to the task representation module through the branches of the convolutional neural network structure; the time series data in the input task data is passed to the LSTM module through the branches of the recurrent neural network structure; the input layer also directly passes the task data to the task specific layer; the task data is the actual test well curve input to the input layer; the feature data includes curve slope, curvature, extreme value, continuity, period and starting point; the time series data includes instantaneous bottom hole pressure data.
[0020] Optionally, the LSTM module is composed of multiple nested LSTM layers; each LSTM layer contains multiple LSTM units; each LSTM unit is composed of an input gate, a forget gate, an output gate and an update storage unit; the time series data information transmitted by the input layer is called a cell state, and the cell state after the current time step is updated is called a hidden state;
[0021] After the time series data is input into the LSTM module, the forget gate uses the sigmoid function to output a number between [0,1] to determine how much information in the cell state is forgotten. The input gate uses the sigmoid function and the tanh function to determine how much information is added to the cell state. The cell state is updated according to the decisions of the input gate and the forget gate. The output gate generates a hidden state based on the updated cell state and decides which information to pass to the next time step. At the last time step, the hidden state is passed to the task representation module as the time series data output by the LSTM module.
[0022] Optionally, the task representation module is composed of multiple convolutional layers and fully connected layers; the convolutional layer is used to convert the feature data from the input layer into the same representation dimension as the time series data output by the LSTM module; the fully connected layer is used to fuse the feature data and time series data of the same dimension by weighted summation and map them into a shared representation space.
[0023] Optionally, the meta-learning neural network is composed of an input layer, a hidden layer and a meta-learning layer; the input layer is used to receive fused data from a shared representation space as input data; the hidden layer is composed of multiple neurons and nonlinear activation functions used to learn abstract feature information in the input data and help the neural network clarify the differences and similarities between different tasks; the hidden layer uses a nonlinear activation function to capture the difference relationship between tasks and convert it into an abstract feature representation and pass it to the meta-learning layer; the meta-learning layer is composed of multiple fully connected layers, convolutional layers and recurrent layers connected to each other, the meta-learning layer receives the abstract feature representation passed from the hidden layer, learns how to adjust the neural network structure to adapt to different tasks by using meta-learning, and outputs the learned adaptation representation for different tasks to the task-specific layer.
[0024] Optionally, the task-specific layer is composed of a deep learning network consisting of multiple fully connected layers and recurrent layers; the task-specific layer receives task data from the input layer and adaptive representations of different tasks output by the meta-learning neural network; the task-specific layer adjusts the well test interpretation parameters through multiple fully connected layers and recurrent layers to fit the current task data, and after fitting, passes the fitted well test model characteristic curve and well test interpretation parameters to the output layer.
[0025] On the other hand, the present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the well test interpretation method that automatically adapts to different well test models is implemented.
[0026] On the other hand, the present invention provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the well test interpretation method that automatically adapts to different well test models.
[0027] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0028] In order to solve the limitation that a lot of preliminary preparation work is required for the current well test interpretation of different reservoir wells and reduce the workload of well test interpretation work, the present invention provides an interpretation method that can automatically adapt to different well test models to obtain accurate well test interpretation parameters, reduce the cost of well test interpretation work, and thus provide more help for the economical and efficient development of oil and gas fields. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0030] Figure 1 A flow chart of a well test interpretation method that automatically adapts to different well test models according to the present invention;
[0031] Figure 2 This is a schematic diagram of the structure of the Meta-LSTM neural network model of the present invention;
[0032] Figure 3 It is a schematic diagram of an actual test well curve of an example well in a specific embodiment;
[0033] Figure 4 This is a fitting curve diagram obtained by using the Meta-LSTM neural network model to interpret the well test of the example well. DETAILED DESCRIPTION
[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0035] The present invention provides a well test interpretation method, medium and equipment that automatically adapt to different well test models, which can automatically adapt to different well test models to obtain accurate well test interpretation parameters and reduce the workload of well test interpretation work.
[0036] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0037] like Figure 1 As shown, the present invention provides a well test interpretation method that automatically adapts to different well test models, including:
[0038] Step 1: Collect different well test models of reservoir wells and corresponding well test data.
[0039] The different well test models collected by the present invention are well test models of reservoir wells with different reservoir properties and different boundary conditions, which may include: analytical models, semi-analytical models and numerical models applied to homogeneous reservoirs and heterogeneous reservoirs, with closed boundaries, constant pressure boundaries and infinite formations. The well test data corresponding to different well test models include instantaneous bottom hole flow pressure data within a preset time period, which can be calculated by recording the shut-in time and bottom hole flow pressure of reservoir wells with the same reservoir properties and boundary conditions as the well test model. The instantaneous bottom hole flow pressure data is the bottom hole flow pressure corresponding to each time point (also referred to as pressure in the present invention).
[0040] Step 2: Solve each well test model to obtain the well test model characteristic curve, and plot the corresponding well test data as the actual test well curve.
[0041] Specifically, each well test model is solved using analytical, semi-analytical or numerical methods to obtain different dimensionless pressures and double logarithmic characteristic curves of pressure derivatives with respect to dimensionless time as well test model characteristic curves. The analytical or semi-analytical methods used may be Laplace transform, Stehfest numerical inversion, Green function method, Bessel function method, etc., to solve each well test model to obtain an analytical solution or semi-analytical solution of the transient pressure response. Alternatively, the numerical model can be solved by programming, and a tool such as MRST can be used to obtain a numerical solution of the transient pressure response; then the time parameters and pressure in the solution process are dimensionless and plotted as horizontal and vertical coordinates in a double logarithmic coordinate system to obtain the well test model characteristic curve.
[0042] Similarly, the actual well test data is also plotted as a double logarithmic characteristic curve of dimensionless pressure and pressure derivative with respect to dimensionless time as the actual test well curve. Specifically, firstly, the pressure data of the actual reservoir well (i.e., instantaneous bottom hole flow pressure data) is derived with respect to time to obtain pressure derivative data, and the pressure data, pressure derivative data, and time data are dimensionless, and are plotted in a double logarithmic coordinate system with dimensionless time data as the horizontal coordinate and dimensionless pressure data and dimensionless pressure derivative data as the vertical coordinate to obtain the actual test well curve of the actual reservoir well.
[0043] Among them, the dimensionless pressure formula is:
[0044] (1)
[0045] The dimensionless time formula is:
[0046] (2)
[0047] in, is the reservoir permeability; is the reservoir thickness; is the horizontal well flow rate; is the viscosity of the oil phase; is the oil phase volume coefficient; Indicates stress; is the initial pressure; is dimensionless pressure; is the reservoir porosity; is the comprehensive compression factor; is the width of the fracturing stimulation area; Indicates time; Represents dimensionless time.
[0048] The double logarithmic characteristic curve is a curve showing the relationship between dimensionless pressure and dimensionless pressure derivative and dimensionless time in double logarithmic coordinates, with dimensionless time as the abscissa and dimensionless pressure and dimensionless pressure derivative as the ordinate. Therefore, the characteristics of the well test model characteristic curve and the actual test well curve are the same. When the geological parameters of the well test model are the same as the geological parameters of the actual reservoir well, the two curves (the well test model characteristic curve and the corresponding actual test well curve) should be consistent within the allowable error range.
[0049] Solving different types of well test models can obtain different types of characteristic curves. The reservoir properties and boundary conditions of the well corresponding to the characteristic curve can be determined by the type of characteristic curve of the well test model. For example, for a dual-porosity medium reservoir, its pressure derivative curve has a "concave" feature, and both curves have horizontal line features; for a homogeneous infinite reservoir, there is a radial flow stage, in which the pressure derivative curve is a horizontal line with a value of 0.5, the slope of the pressure derivative curve in the linear flow stage is 0.5, and the slope of the curve in the bilinear flow stage is 0.25; when the reservoir is a radial closed boundary, after the "concave" feature of the pressure derivative curve ends, the slope of the pressure derivative curve becomes 1; when the reservoir is in a quasi-steady-state flow state under closed boundary conditions, there is a straight line with a slope of 1 on the pressure derivative curve; when the reservoir is a linear closed boundary, the pressure curve and the pressure derivative curve are two horizontal lines with different values; when the reservoir is a constant pressure boundary, the pressure derivative curve will have the characteristic of continuously decreasing and tending to zero.
[0050] Step 3: Construct training set and test set based on the well test model characteristic curve and the actual test well curve.
[0051] Specifically, identifying an actual test well curve is regarded as a task, so the actual test well curve is also called a task curve. All collected task curves are put into a set to identify all curve types in the set as a task set. The well test model characteristic curve corresponding to each task curve is used as a label, and the task curve and the corresponding label are used as a sample pool.
[0052] For each task curve in the sample pool, noise point data is randomly added, and a data set is generated together with the data on the task curve. The data set is divided into a training set and a test set in a ratio of 7:3. The training set is used to train the neural network model parameters, and the test set is used to evaluate the performance of the neural network model.
[0053] Step 4: Construct a Meta-LSTM neural network model for identifying and interpreting different well test models.
[0054] To establish a well test interpretation method that is suitable for different well test models, a neural network model based on a meta-learning algorithm is required. The present invention proposes a Meta-LSTM neural network model to identify the pressure characteristic curves of different types of reservoir wells and perform well test interpretation. Specifically, Figure 2 As shown, the Meta-LSTM neural network model is connected and integrated by an input layer, an LSTM module, a task representation module, a meta-learning neural network, a task-specific layer, and an output layer through a functional API. The input of the Meta-LSTM neural network model is the actual test well curve, and the output is the fitted well test model characteristic curve and the well test interpretation parameters. Both the actual test well curve and the well test model characteristic curve contain characteristic data and time series data, wherein the characteristic data include curve slope, curvature, extreme value, continuity, cycle and starting point; the time series data includes instantaneous bottom hole pressure data. Therefore, the input task curve is also referred to as task data in the present invention.
[0055] Specifically, the input layer of the Meta-LSTM neural network model is a multimodal input layer; the multimodal input layer includes independent branches of convolutional neural network structure and branches of recurrent neural network structure. The feature data in the input task data is passed to the task representation module through the branches of the convolutional neural network structure; the time series data in the input task data is passed to the LSTM module through the branches of the recurrent neural network structure. The input layer also directly passes the task data to the task specific layer.
[0056] Specifically, the LSTM module is composed of multiple nested LSTM layers; each LSTM layer contains multiple LSTM units; each LSTM unit is composed of an input gate, a forget gate, an output gate and an update storage unit. The time series data information transmitted by the recorded input layer is called the cell state, and the cell state updated at the current time step is called the hidden state. The hidden state is a zero vector in the initial state. After the time series data is input into the LSTM module, the forget gate uses the sigmoid function to output a number between [0,1] to determine how much information in the cell state is forgotten. The input gate uses the sigmoid function and the tanh function to determine how much information is added to the cell state. The cell state is updated according to the decision of the input gate and the forget gate. The output gate generates a hidden state based on the updated cell state, determines which information is passed to the next time step, and passes the hidden state as the time series data output by the LSTM module to the task representation module in the last time step.
[0057] Specifically, the task representation module is composed of multiple convolutional layers connected to fully connected layers. The convolutional layer is used to convert the feature data from the input layer into the same representation dimension as the time series data output by the LSTM module; the fully connected layer is used to fuse the feature data and time series data of the same dimension through weighted summation and map them into a shared representation space.
[0058] The weighted sum expression of feature data and time series data is as follows:
[0059] (3)
[0060] in, Represents the fused data after weighted summation; Indicates Characteristic data; Indicates the weight of feature data, percentage; Indicates Time series data; Indicates the weight of time series data, in percentage.
[0061] Mapping is to convert the input data in the original data space into a vector representation in the shared representation space. The shared representation space is an abstract feature space shared by different tasks, where each task can learn some common features or similar features.
[0062] Specifically, the meta-learning neural network is composed of an input layer, a hidden layer and a meta-learning layer. The input layer is used to receive fused data from the shared representation space. As input data. The hidden layer consists of multiple neurons and nonlinear activation functions used to learn abstract feature information in the input data and help the neural network clarify the similarities and differences between different tasks. The hidden layer uses nonlinear activation functions to capture the difference relationship between tasks and convert it into abstract feature representation and pass it to the meta-learning layer. The meta-learning layer is composed of multiple fully connected layers, convolutional layers and recurrent layers connected to each other. The meta-learning layer receives the abstract feature representation passed by the hidden layer, learns how to adjust the neural network structure to adapt to different tasks by using meta-learning, and outputs the learned adaptive representation for different tasks to the task-specific layer. Abstract feature representation is a general feature representation that is independent of the task and has generalization ability. Learning adaptive representations for different tasks means identifying the current actual test well curve and judging the reservoir properties and boundary condition types. After identification, the meta-learning layer passes the adaptive representation to the task-specific layer, and the task-specific layer gives the corresponding well test model characteristic curve.
[0063] The nonlinear activation function formula uses the ReLU function as an example, and its expression is as follows:
[0064] (4)
[0065] in, Represents the output of the ReLU function; Represents the input data passed by the input layer, and max represents the maximum value.
[0066] Specifically, the task-specific layer is composed of a deep learning network composed of multiple fully connected layers and recurrent layers. The task-specific layer receives the task curve from the input layer and the adaptation representation of different tasks output by the meta-learning neural network. The task-specific layer adjusts the well test interpretation parameters through multiple fully connected layers and recurrent layers to fit the current task curve, and after fitting, the well test model characteristic curve and well test interpretation parameters are passed to the output layer.
[0067] Well test interpretation parameters may include: reservoir length, width and height, porosity, permeability, formation coefficient, formation flow coefficient, skin coefficient, comprehensive compression coefficient, hydraulic fracture length, width and number, etc. Well test interpretation parameters are important parameters in the well test model. By changing the well test interpretation parameters, the pressure of the well test model can be changed, thereby changing the curve morphology characteristics and the objective function value. The well test interpretation parameters are changed multiple times to make the objective function value meet the error requirements. At this time, the fitted well test model characteristic curve coincides with the actual test well curve.
[0068] Objective function is used to quantify the characteristic curve of the well test model Compared with the actual test well curve Set an error limit. When the objective function is less than the error limit, the fitting is considered successful. The objective function expression is:
[0069] (5)
[0070] in, represents the objective function; is the data point on the curve; is the number of data points. The objective function error is the mean error of all data points on the entire curve. The output fitting characteristic curve is used to check whether there is a phenomenon of excessive error in some data, so as to further adjust and optimize the neural network model.
[0071] The output layer of the Meta-LSTM neural network model is a fully connected layer, which receives the output of the task-specific layer and outputs the fitted well test model characteristic curve and well test interpretation parameters as the output of the Meta-LSTM neural network model.
[0072] Step 5: Use the training set to train the Meta-LSTM neural network model parameters, and use the test set to evaluate the performance of the Meta-LSTM neural network model. After the training is completed, the adaptive well test interpretation model is obtained.
[0073] The two double logarithmic characteristic curves collected by the present invention, namely the well test model characteristic curve and the actual test well curve, are used to construct a data set. The well test model characteristic curve is used to train the neural network model to identify the characteristics of the input task curve, and the actual test well curve is used to determine the characteristics of this type of well and fit the well test model characteristic curve during the model training process. The established Meta-LSTM neural network model is meta-trained using the training set divided in step 3, and is quickly adapted to different tasks by using a meta-gradient descent algorithm. The meta-gradient descent algorithm includes an inner gradient descent and an outer gradient descent. The gradient update of the algorithm is performed on the gradient of the task performance, rather than on the gradient of the loss function.
[0074] The meta-training steps include:
[0075] (5.1.1) Initialize the parameters of the meta-learning network, including: learning rate, number of hidden layers, number of neural units in each hidden layer, activation function, optimization algorithm, and meta-learning objective function weight;
[0076] (5.1.2) Using the well test model characteristic curve corresponding to each task as the label, and the data and label of the task as the sample pool, a task is selected from the training set using the simple random sampling method, and some data is randomly selected from the task sample pool as samples to input into the neural network model, and forward propagation is performed according to the established model structure, and the task curve is identified in the meta-learning layer to complete the task; the simple random sampling method is to randomly select a task in each task set with equal probability, and each sampling is independent of each other;
[0077] (5.1.3) Each task uses the inner gradient descent algorithm to learn the parameters that are suitable for the task and update the parameters of the neural network model so that the model can better adapt to the current task; the inner gradient descent algorithm is to perform the traditional gradient descent process within each task to update the weights;
[0078] (5.1.4) After the task is completed, select a new task and repeat the above steps for iteration. When all tasks are completed, use the outer gradient descent algorithm to update the initial parameters of the neural network model for multiple tasks to make it better adapt to different tasks; the outer gradient descent algorithm is to perform a gradient descent process on multiple tasks to update the weights.
[0079] After the training is completed, use the test set divided in step 3 to meta-test the trained Meta-LSTM neural network model and analyze the accuracy and generalization ability of the model, including:
[0080] (5.2.1) Initialize the parameters of the neural network model to the final parameters learned in the meta-training phase. Use a simple random sampling method to select a task in the test set, randomly select some samples from the task sample pool as input, and use a small number of steps of gradient descent in a similar way to the meta-training phase to ensure the adaptability of the model to the new task;
[0081] (5.2.2) Select a set of test samples that are different from the input samples in the task sample pool for testing and calculate the loss function. Collect the loss functions of the entire test set tasks to comprehensively evaluate the neural network model.
[0082] The loss function uses the cross entropy loss function, which is expressed as:
[0083] (6)
[0084] in, is the number of data points; is the value of the label data point; To adjust the data point values; is the cross entropy loss.
[0085] (5.2.3) Back propagation is performed based on the test and evaluation results to further adjust or optimize the neural network model to improve the generalization and accuracy of the model. Back propagation is to input the updated parameters of multiple tasks and the calculated task losses, and output the parameters that need to be adjusted for the neural network model. The neural network model is adjusted according to the output of back propagation.
[0086] The Meta-LSTM neural network model that has passed the meta-training stage and the meta-testing stage is applied as an adaptive well test interpretation model. The adaptive well test interpretation model can automatically select the corresponding well test model of the reservoir well for well test interpretation.
[0087] Step 6: Use the adaptive well test interpretation model to perform well test interpretation and obtain the well test interpretation parameters of the reservoir well.
[0088] The field well test data of the reservoir well to be interpreted is plotted as a real well test curve, which is input into the adaptive well test interpretation model for forward propagation. The task data is converted into a vector representation in the shared representation space in the task representation module; the meta-learning neural network identifies the reservoir properties and boundary condition types of the well according to the vector representation, and gives the corresponding well test model characteristic curve; the input task data is fitted with the well test model characteristic curve in the task specific layer, and the well test model interpretation parameters are adjusted to minimize the objective function, and the fitting characteristic curve and well test interpretation parameters of the well are obtained in the output layer. The well test interpretation parameters include the length, width and height of the reservoir, porosity, permeability, formation coefficient, formation flow coefficient, skin coefficient, comprehensive compression coefficient, and the length, width and number of hydraulic fractures.
[0089] The method of the present invention is used to interpret the well test of an example well. The method of the present invention is used to interpret the well test data of a well on site. The Meta-LSTM neural network model is constructed according to step 4, and the actual well test data of the well is collected and the actual test well curve is drawn as shown in FIG. Figure 3 The actual test well curve is input into the Meta-LSTM neural network model, and the model automatically identifies that the reservoir where the well is located is a homogeneous reservoir, and the boundary condition is a closed boundary. The well test interpretation parameters of the well are shown in Table 1.
[0090] Table 1 Well test interpretation parameters of the example well
[0091]
[0092] Figure 4 This is the fitting curve obtained by using the Meta-LSTM neural network model to interpret the well test of the example well; Figure 4 It can be seen that the actual well test data points of the well are highly coincident with the data points on the pressure curve of the well test model after fitting, which indicates that the well test interpretation method of the present invention is stable and reliable.
[0093] In some embodiments, the present invention further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the well test interpretation method that automatically adapts to different well test models is implemented.
[0094] In some embodiments, the present invention further provides a computer device, which includes a processor, a memory, an input / output interface (I / O for short) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store pending transactions. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the well test interpretation method that automatically adapts to different well test models can be implemented.
[0095] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A well test interpretation method that automatically adapts to different well test models, characterized in that: include: Collecting different well test models of reservoir wells and corresponding well test data; the different well test models of reservoir wells are well test models of reservoir wells with different reservoir properties and different boundary conditions; the well test data include instantaneous bottom hole flow pressure data within a preset time period; Solve each well test model to obtain a well test model characteristic curve, and plot the corresponding well test data as an actual well test curve; Construct training set and test set based on well test model characteristic curve and actual test well curve; A Meta-LSTM neural network model is constructed for identifying and interpreting different well test models; the Meta-LSTM neural network model is connected and integrated by an input layer, an LSTM module, a task representation module, a meta-learning neural network, a task-specific layer, and an output layer through a functional API; the input of the Meta-LSTM neural network model is an actual well test curve, and the output is a fitted well test model characteristic curve and a well test interpretation parameter; The input layer of the Meta-LSTM neural network model is a multimodal input layer; the multimodal input layer includes a branch of an independent convolutional neural network structure and a branch of a recurrent neural network structure; the feature data in the input task data is transferred to the task representation module through the branch of the convolutional neural network structure; The time series data in the input task data is passed to the LSTM module through the branch of the recurrent neural network structure; the input layer also directly passes the task data to the task specific layer; the task data is the actual test well curve input to the input layer; the characteristic data includes the curve slope, curvature, extreme value, continuity, cycle and starting point; the time series data includes instantaneous bottom hole flow pressure data; The task representation module is composed of multiple convolutional layers and fully connected layers; the convolutional layer is used to convert the feature data from the input layer into the same representation dimension as the time series data output by the LSTM module; the fully connected layer is used to fuse the feature data and time series data of the same dimension by weighted summation and map them into a shared representation space; The task-specific layer is composed of a deep learning network consisting of multiple fully connected layers connected to a recurrent layer; the task-specific layer receives task data from the input layer and adaptive representations of different tasks output by the meta-learning neural network; The task-specific layer adjusts the well test interpretation parameters through multiple fully connected layers and loop layers to fit the current task data. After fitting, the well test model characteristic curve and well test interpretation parameters are passed to the output layer. The training set is used to train the parameters of the Meta-LSTM neural network model, and the test set is used to evaluate the performance of the Meta-LSTM neural network model. After the training is completed, an adaptive well test interpretation model is obtained; An adaptive well test interpretation model is used to perform well test interpretation to obtain well test interpretation parameters of the reservoir well; the well test interpretation parameters include reservoir length, width and height, porosity, permeability, formation coefficient, formation flow coefficient, skin coefficient, comprehensive compression coefficient, and length, width and number of hydraulic fractures.
2. The well test interpretation method automatically adapting to different well test models according to claim 1 is characterized in that: The method of solving each well test model to obtain a well test model characteristic curve and drawing the corresponding well test data into an actual well test curve specifically includes: Solve each well test model to obtain different double logarithmic characteristic curves of dimensionless pressure and pressure derivative with respect to dimensionless time as characteristic curves of the well test model; The corresponding well test data are also plotted as a double logarithmic characteristic curve of dimensionless pressure and pressure derivative with respect to dimensionless time as the actual test well curve.
3. The well test interpretation method automatically adapting to different well test models according to claim 2 is characterized in that: The construction of the training set and the test set based on the well test model characteristic curve and the actual test well curve specifically includes: Identifying a real test well curve is regarded as a task, the well test model characteristic curve corresponding to each task curve is used as a label, and the task curve and the corresponding label are used as a sample pool; For each task curve in the sample pool, noise point data is randomly added, and a data set is generated together with the data on the task curve. The data set is divided into a training set and a test set according to a preset ratio.
4. The well test interpretation method automatically adapting to different well test models according to claim 3 is characterized in that: The LSTM module is composed of multiple nested LSTM layers; each LSTM layer contains multiple LSTM units; each LSTM unit is composed of an input gate, a forget gate, an output gate and an update storage unit; the time series data information transmitted by the input layer is called the cell state, and the cell state after the current time step is updated is called the hidden state; After the time series data is input into the LSTM module, the forget gate uses the sigmoid function to output a number between [0,1] to determine how much information in the cell state is forgotten. The input gate uses the sigmoid function and the tanh function to determine how much information is added to the cell state. The cell state is updated according to the decisions of the input gate and the forget gate. The output gate generates a hidden state based on the updated cell state and decides which information to pass to the next time step. At the last time step, the hidden state is passed to the task representation module as the time series data output by the LSTM module.
5. The well test interpretation method automatically adapting to different well test models according to claim 4 is characterized in that: The meta-learning neural network is composed of an input layer, a hidden layer and a meta-learning layer; the input layer is used to receive fused data from a shared representation space as input data; the hidden layer is composed of a plurality of neurons and nonlinear activation functions used to learn abstract feature information in the input data and help the neural network to clarify the similarities and differences between different tasks; the hidden layer uses a nonlinear activation function to capture the difference relationship between tasks and converts it into an abstract feature representation and transmits it to the meta-learning layer; The meta-learning layer is composed of multiple fully connected layers, convolutional layers and recurrent layers connected to each other. The meta-learning layer receives the abstract feature representation passed by the hidden layer, learns how to adjust the neural network structure to adapt to different tasks by using meta-learning, and outputs the learned adaptation representation for different tasks to the task-specific layer.
6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the well test interpretation method for automatically adapting to different well test models as claimed in claim 1 is implemented.
7. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the well test interpretation method for automatically adapting to different well test models as described in claim 1.
