A Single-Well Production Dynamic Prediction Method Based on Long Short-Term Memory Deep Neural Network

Abstract

This invention discloses a method for dynamic production prediction of a single well based on a long short-term memory (LSTM) deep neural network, comprising: acquiring an input historical dataset X and an output historical dataset Y; extracting an input training set X1 and an input test set X2 from the input historical dataset X, and applying an LSM deep neural network to obtain a dynamic production prediction model and an output test prediction set Y2; selecting a dynamic production prediction model that meets the desired effect through a root mean square error function; acquiring single-well production parameters and constraints, and applying an optimization algorithm to generate multiple injection-production schemes; for each injection-production scheme, acquiring the historical dynamic production data corresponding to each scheme, and sequentially applying the dynamic production prediction model and the oilfield economic net present value function to obtain the net present value corresponding to each injection-production scheme, thereby selecting the optimal injection-production scheme. The dynamic production prediction method of this invention improves calculation speed and timeliness, and has high fitting accuracy.

Description

Technical Field

[0001] This invention belongs to the field of reservoir engineering, and more specifically, relates to a method, electronic equipment and medium for predicting the dynamic production of a single well based on a long short-term memory deep neural network. Background Technology

[0002] In reservoir numerical simulation, to ensure that dynamic predictions closely approximate actual conditions, history fitting is typically performed. This involves calculating the oilfield's development history using existing reservoir parameters such as permeability, porosity, and saturation, and then comparing the calculated development indicators, such as pressure, production, and water cut, with the actual dynamics of oilfield development. If the calculated results do not match the actual measured data, it indicates that the input parameters are inconsistent with reality, requiring parameter adjustment and recalculation until the calculated results are within the allowable error range. Once the model is well-fitted, it is used for production dynamic prediction, thereby guiding adjustments in on-site oilfield development.

[0003] The reservoir simulation history fitting problem is a typical inverse problem. Based on its development process and implementation methods, history fitting can be divided into two main categories: manual history fitting and automatic history fitting. Currently, the common practice is to manually perform repeated calculations to modify reservoir parameters and reduce history fitting errors. This work is arduous and tedious, relying on the experience and expertise of reservoir engineers, as well as other uncertainties. Furthermore, actual reservoirs are often highly heterogeneous with numerous parameters, while high-precision reservoir models have a huge mesh size, making it difficult to meet the needs of large reservoirs requiring long-term development.

[0004] Computers and optimization algorithms have been introduced to automatically adjust reservoir parameters, gradually leading to the development of automatic history fitting technology. Automatic history fitting compensates for the shortcomings of manual calculations, using optimization methods to automatically correct model parameters and structure, striving to reduce fitting time and achieve higher accuracy. However, due to the diverse characteristics of various automatic history fitting methods, the diversity of measured data, and the inherent ill-posedness of inverse problems, different methods yield varying results when solving the same problem.

[0005] In recent years, machine learning algorithms have shown remarkable advantages in fitting problems and have been widely applied to various data fitting and regression problems. Building production prediction models using machine learning algorithms is simple and practical. Many scholars have used machine learning methods such as BP neural networks and support vector machines to achieve dynamic prediction of oil well production, which has great application value. However, these traditional machine learning methods ignore the trend of production changes over time and the correlation between previous and subsequent data, resulting in poor prediction performance.

[0006] Therefore, it is hoped that a method for dynamic prediction of single-well production can be invented to solve the problem of poor prediction results in the use of machine learning algorithms in the prior art, which ignores the trend of production change over time and the correlation between previous and subsequent data. Summary of the Invention

[0007] The object of the present invention is to propose a single-well production dynamic prediction method to solve the problem in the prior art that when using machine learning algorithms, the prediction effect is poor due to ignoring the change trend of production over time and the correlation between front and back data.

[0008] To achieve the above object, the present invention provides a single-well production dynamic prediction method based on a long short-term memory deep neural network, including:

[0009] Obtain an input historical data set X and an output historical data set Y;

[0010] Extract an input training set X1 and an input test set X2 from the input historical data set X respectively, and apply the long short-term memory deep neural network based on the input training set X1 and the input test set X2 to obtain a production dynamic prediction model and an output test prediction set Y2;

[0011] Based on the output historical data set Y and the output test prediction set Y2, screen out a production dynamic prediction model that meets the desired effect through a root mean square error function;

[0012] Obtain single-well production parameters and single-well production parameter constraint conditions, and apply an optimization algorithm to generate multiple injection-production plans;

[0013] For each injection-production plan, obtain the historical production dynamic data corresponding to each injection-production plan, and sequentially apply the production dynamic prediction model and the oilfield economic net present value function to obtain the net present value corresponding to each injection-production plan, so as to select the optimal injection-production plan.

[0014] Optionally, the obtaining of the input historical data set X and the output historical data set Y includes:

[0015] Construct a historical data set;

[0016] Extract the data corresponding to the (t - 1)-th day, the t-th day and the (t + 1)-th day from the historical data set as input historical data, and extract the data corresponding to the (t + 2)-th day from the historical data set as output historical data, where 1 < t < Q - 1, Q is the total production time and t is an integer;

[0017] Form the input historical data set X with the input historical data, and form the output historical data set Y with the output historical data.

[0018] Optionally, the historical data set includes a reservoir static data set and a production dynamic data set, the input historical data set X includes an input historical reservoir static data set and an input historical production dynamic data set, and the output historical data set Y includes an output historical reservoir static data set and an output historical production dynamic data set.

[0019] Optionally, the process of constructing the historical dataset includes:

[0020] The historical dataset is cleaned;

[0021] The cleaned historical dataset is standardized using the following formula:

[0022]

[0023] Where x is the data in the historical dataset, mean(x) is the mean of data x, std(x) is the standard deviation of data x, and y is the standardized data.

[0024] Optionally, the step of extracting the input training set X1 and the input test set X2 from the input historical dataset X, and applying a long short-term memory deep neural network based on the input training set X1 and the input test set X2 to obtain the production dynamic prediction model and the output test prediction set Y2 includes:

[0025] A predetermined proportion of data is extracted from the input historical dataset X as the input training set X1, and the input training set X1 is used to train the long short-term memory deep neural network to obtain the production dynamic prediction model.

[0026] The remaining proportion data is extracted from the input dataset X as the input test set X2, and the input test set X2 is input into the production dynamic prediction model to obtain the output test prediction set Y2.

[0027] Optionally, for each injection-production scheme, obtaining historical production dynamic data corresponding to each injection-production scheme, and sequentially applying the production dynamic prediction model and the economic net present value function of oilfield development to obtain the net present value corresponding to each injection-production scheme, thereby selecting the optimal injection-production scheme, includes:

[0028] For each injection and extraction scheme, the production dynamic prediction model is applied to perform production dynamic prediction to obtain the predicted production dynamic data corresponding to each injection and extraction scheme.

[0029] Historical production dynamic data corresponding to each injection and production scheme is obtained, and based on the historical production dynamic data and predicted production dynamic data corresponding to each injection and production scheme, the net present value corresponding to each injection and production scheme is calculated through the oilfield economic net present value function.

[0030] The injection and extraction scheme with the highest net present value is selected as the optimal injection and extraction scheme.

[0031] Optionally, the root mean square error function is:

[0032]

[0033] Where RMSE is the error loss, M is the total number of responses extracted from the output test prediction set Y2, and K... i T is the response value extracted from the output test prediction set Y2. i K is extracted from the output historical dataset Y i The corresponding target value, where N is the total number of data in the output test prediction set Y2;

[0034] If the error loss is within a preset range, then the prediction model satisfies the expected effect.

[0035] Optionally, the net present value function of the oilfield is:

[0036]

[0037] Where NPV is the net present value, r is the discount rate, M1 represents the total production time, N1 is the total number of wells, and C d H represents the average drilling cost per unit length. i Let C be the depth of the i-th well. c For well completion costs, p o Q represents the selling price of crude oil. t,o Let p be the crude oil production on day t. wp The cost per unit of water collection, p wi The cost of water injection per unit, Q t,wp Let Q be the amount of water extracted on day t. t,wi Let t be the amount of water injected on day t.

[0038] An electronic device, the electronic device comprising:

[0039] Memory, which stores executable instructions;

[0040] A processor that executes the executable instructions in the memory to implement the single-well production dynamic prediction method based on a long short-term memory deep neural network.

[0041] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the single-well production dynamic prediction method based on a long short-term memory deep neural network.

[0042] The beneficial effects of this invention are as follows:

[0043] The input historical dataset X and output historical dataset Y obtained by the single-well production dynamic prediction method of this invention have time-series characteristics. A production dynamic history fitting and prediction method is established using a long short-term memory deep neural network. First, the input historical dataset X and output historical dataset Y are obtained. Then, a production dynamic prediction model is established using a long short-term memory deep neural network. Second, combined with the constraints of the input single-well production parameters, an optimization algorithm is applied to generate multiple injection and production schemes. Finally, the optimal injection and production scheme is obtained by comparing and selecting the best scheme based on the oilfield economic net present value function. The single-well production dynamic prediction method of this invention improves the calculation speed and increases the timeliness. At the same time, it has high fitting accuracy, solving the problem of poor prediction results in existing technologies when using machine learning algorithms because the trend of production changes over time and the correlation between previous and subsequent data are ignored.

[0044] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0045] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments of the invention.

[0046] Figure 1 A flowchart of a single-well production dynamic prediction method based on a long short-term memory deep neural network according to an embodiment of the present invention is shown.

[0047] Figure 2 The diagram shows the variation of the model training function with the number of training iterations for a single-well production dynamic prediction method based on a long short-term memory deep neural network according to an embodiment of the present invention.

[0048] Figure 3 The diagram shows a comparison between the actual oil production and the predicted oil production of a single well based on a long short-term memory deep neural network according to an embodiment of the present invention.

[0049] Figure 4 The diagram shows a comparison between the actual water production and the predicted water production of a single well based on a long short-term memory deep neural network according to an embodiment of the present invention.

[0050] Figure 5 The diagram illustrates a basic scheme for a single-well production dynamic prediction method based on a long short-term memory deep neural network according to an embodiment of the present invention.

[0051] Figure 6The diagram illustrates the training and testing of a production dynamic prediction model for a single-well production dynamic prediction method based on a long short-term memory deep neural network according to an embodiment of the present invention.

[0052] Figure 7 The optimal injection-production scheme of a single-well production dynamic prediction method based on a long short-term memory deep neural network according to an embodiment of the present invention is shown.

[0053] Figure 8 The diagram shows a comparison of the NPV of the basic injection-production scheme and the NPV of the optimal injection-production scheme according to an embodiment of the present invention, which is based on a long short-term memory deep neural network for single-well production dynamic prediction.

[0054] Figure 9 The diagram shows a comparison of the remaining oil saturation of the basic injection-production scheme and the optimal injection-production scheme according to an embodiment of the present invention, which is a method for predicting the dynamic production of a single well based on a long short-term memory deep neural network. Detailed Implementation

[0055] Preferred embodiments of the invention will now be described in more detail. While preferred embodiments of the invention are described below, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0056] A method for predicting the dynamic production of a single well based on a long short-term memory deep neural network according to the present invention includes:

[0057] Obtain the input historical dataset X and the output historical dataset Y;

[0058] Extract the input training set X1 and the input test set X2 from the input historical dataset X, and apply the long short-term memory deep neural network based on the input training set X1 and the input test set X2 to obtain the production dynamic prediction model and the output test prediction set Y2.

[0059] Based on the output historical dataset Y and the output test prediction set Y2, the production dynamic prediction model that meets the expected effect is selected by the root mean square error function.

[0060] Obtain the production parameters and constraints of a single well, and apply optimization algorithms to generate multiple injection and production schemes.

[0061] For each injection-production plan, obtain the historical production performance data corresponding to each injection-production plan, and successively apply the production performance prediction model and the economic net present value function of the oilfield to obtain the net present value corresponding to each injection-production plan, so as to select the optimal injection-production plan.

[0062] Specifically, the input historical data set X and the output historical data set Y obtained by applying the single-well production performance prediction method of the present invention have the characteristic of time series. A production performance historical fitting and prediction method is established by using a long short-term memory deep neural network. First, obtain the input historical data set X and the output historical data set Y. Then, apply the long short-term memory deep neural network to establish a production performance prediction model. Secondly, combine the input single-well production parameter constraints and apply an optimization algorithm to generate multiple injection-production plans. Finally, compare and optimize based on the economic net present value function of the oilfield to obtain the optimal injection-production plan; the single-well production performance prediction method of the present invention improves the calculation speed and increases the timeliness. At the same time, it has a high fitting accuracy, and solves the problem that in the prior art, when using machine learning algorithms, the prediction effect is poor due to ignoring the change trend of production over time and the correlation between front and back data.

[0063] Furthermore, through the learning of the long short-term memory deep neural network, the mutual influence between data can be obtained, so as to map the inter-well connection relationship, making the production performance prediction model have the production performance prediction function. Compare the gap between the prediction result and the actual value of the production performance prediction model on the test set to test the prediction effect of the production performance prediction model.

[0064] Furthermore, for injection wells, it is necessary to satisfy: minimum single-well injection volume < injection volume of injection well < maximum single-well injection volume; for production wells, it is necessary to satisfy: economic limit production of single well < production of production well < maximum production of single well.

[0065] In one example, obtaining the input historical data set X and the output historical data set Y includes:

[0066] Construct a historical data set;

[0067] Extract the data corresponding to the (t - 1)-th day, the t-th day and the (t + 1)-th day from the historical data set as the input historical data, and extract the data corresponding to the (t + 2)-th day from the historical data set as the output historical data, where 1 < t < Q - 1, Q is the total production time and t is an integer;

[0068] Form the input historical data set X with the input historical data and form the output historical data set Y with the output historical data.

[0069] Specifically, the historical dataset mainly includes three categories of factors affecting oil well production. The first category includes production wells' production time, bottom hole pressure, production pressure differential, water saturation, daily oil production, daily water production, and water cut. The second category includes injection wells' injection time, bottom hole pressure, and injection volume. The third category includes the reservoir's remaining recoverable reserves, surrounding reservoir porosity, and permeability. These three categories of characteristic parameters are collected and organized.

[0070] In one example, the historical dataset includes a reservoir static dataset and a production dynamic dataset. The input historical dataset X includes the input historical reservoir static dataset and the input historical production dynamic dataset, and the output historical dataset Y includes the output historical reservoir static dataset and the output historical production dynamic dataset.

[0071] Specifically, the historical dataset includes a reservoir static dataset and a production dynamic dataset. Data corresponding to day t-1, day t, and day t+1 are extracted from the reservoir static dataset as input to the historical reservoir static dataset, and data corresponding to day t+2 is extracted from the reservoir static dataset as output to the historical reservoir static dataset. Similarly, data corresponding to day t-1, day t, and day t+1 are extracted from the production dynamic dataset as input to the historical production dynamic dataset, and data corresponding to day t+2 is extracted from the production dynamic dataset as output to the historical production dynamic dataset.

[0072] Furthermore, the reservoir static dataset is static data that does not change over time; therefore, it remains unchanged at each time step.

[0073] In one example, after building the historical dataset, the following steps are included:

[0074] Clean the historical dataset;

[0075] The cleaned historical dataset is standardized using the following formula:

[0076]

[0077] Where x is the data in the historical dataset, mean(x) is the mean of data x, std(x) is the standard deviation of data x, and y is the standardized data.

[0078] Specifically, cleaning historical datasets includes removing noisy points from them, because some non-floating points in the production dynamic dataset are not under the same production conditions as other points and may be due to human factors, so these points need to be cleaned; while standardizing historical datasets is to eliminate the influence of dimensions between features.

[0079] In one example, input training set X1 and input test set X2 are extracted from the input historical dataset X, respectively. Based on the input training set X1 and input test set X2, a long short-term memory deep neural network is applied to obtain a production dynamic prediction model and an output test prediction set Y2, including:

[0080] Extract a predetermined proportion of data from the input historical dataset X as the input training set X1, and use the input training set X1 to train the long short-term memory deep neural network to obtain a production dynamic prediction model.

[0081] Extract the remaining proportion data from the input dataset X as the input test set X2, and input the input test set X2 into the production dynamic prediction model to obtain the output test prediction set Y2.

[0082] Specifically, in practical applications, the preset ratio is preferably 75%, leaving a remaining ratio of 25%. The training output of a Long Short-Term Memory (LSTM) neural network is mainly affected by factors such as the learning rate, the number of hidden nodes, the number of training steps, and the length of the time series. The preferred range for the learning rate is 1×10⁻⁶. -5 The optimal range for the number of hidden nodes is typically 16–512, the optimal range for the number of training steps is 300–3000, and the optimal range for the length of the cyclic input time series is 3–24. By continuously adjusting the network parameters, the predicted data can achieve a high degree of consistency with the measured data.

[0083] In one example, for each injection-production scheme, historical production dynamic data corresponding to each scheme is obtained, and the production dynamic prediction model and the economic net present value function of oilfield development are applied sequentially to obtain the net present value corresponding to each injection-production scheme, thereby selecting the optimal injection-production scheme, including:

[0084] For each injection and extraction scheme, a production dynamic prediction model is applied to predict production dynamics and obtain the predicted production dynamic data corresponding to each injection and extraction scheme.

[0085] The historical production dynamic data corresponding to each injection and production plan is obtained, and based on the historical production dynamic data and predicted production dynamic data corresponding to each injection and production plan, the net present value corresponding to each injection and production plan is calculated through the oilfield economic net present value function.

[0086] The injection and extraction scheme with the highest net present value is selected as the optimal injection and extraction scheme.

[0087] Specifically, in practical applications, for each injection-production scheme, the parameters that need to be designed include the control frequency, the production volume of each production well during each control, and the injection volume of each injection well during each control.

[0088] In one example, the root mean square error function is:

[0089]

[0090] Where RMSE is the error loss, M is the total number of responses extracted from the output test prediction set Y2, and K... i T is the response value extracted from the output test prediction set Y2. i K is extracted from the output historical dataset Y i The corresponding target value, where N is the total number of data in the output test prediction set Y2;

[0091] If the error loss is within the preset range, the prediction model meets the expected results.

[0092] Specifically, in practical applications, the smaller the error loss, the better the prediction effect of the prediction model.

[0093] In one example, the net present value function for an oil field is:

[0094]

[0095] Where NPV is the net present value, r is the discount rate, M1 represents the total production time, N1 is the total number of wells, and C d H represents the average drilling cost per unit length. i Let C be the depth of the i-th well. c For well completion costs, p o Q represents the selling price of crude oil. t,o Let p be the crude oil production on day t. wp The cost per unit of water collection, p wi The cost of water injection per unit, Q t,wp Let Q be the amount of water extracted on day t. t,wi Let t be the amount of water injected on day t.

[0096] Specifically, for each injection-production scheme, based on historical production dynamics data and predicted production dynamics data, the net present value (NPV) of each scheme is calculated using the oilfield economic NPV function. The injection-production scheme with the highest NPV is selected as the optimal scheme. The higher the NPV, the better the injection-production scheme and the better the investment returns.

[0097] An electronic device, the electronic device comprising:

[0098] Memory, which stores executable instructions;

[0099] A processor that executes the executable instructions in the memory to implement the single-well production dynamic prediction method based on a long short-term memory deep neural network.

[0100] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the single-well production dynamic prediction method based on a long short-term memory deep neural network.

[0101] Example 1

[0102] like Figure 1 As shown, a method for predicting the dynamic production of a single well based on a long short-term memory deep neural network includes:

[0103] Obtain the input historical dataset X and the output historical dataset Y;

[0104] Extract the input training set X1 and the input test set X2 from the input historical dataset X, and apply the long short-term memory deep neural network based on the input training set X1 and the input test set X2 to obtain the production dynamic prediction model and the output test prediction set Y2.

[0105] Based on the output historical dataset Y and the output test prediction set Y2, the production dynamic prediction model that meets the expected effect is selected by the root mean square error function.

[0106] Obtain the production parameters and constraints of a single well, and apply optimization algorithms to generate multiple injection and production schemes.

[0107] For each injection and production scheme, historical production dynamic data corresponding to each scheme are obtained, and the production dynamic prediction model and the oilfield economic net present value function are applied sequentially to obtain the net present value corresponding to each injection and production scheme, thereby selecting the optimal injection and production scheme.

[0108] The specific implementation method is as follows:

[0109] Taking a fractured carbonate reservoir as an example, the original formation pressure was 65.0 MPa, comprising two systems: fractures and matrix, at a depth of 5200–5700 m. The matrix is ​​the main oil-bearing space, with a porosity of 0.01%–9.5%. The fracture pore space is smaller, with a porosity of 0.002%–1.2%, and serves as the main seepage channel. Historical datasets were constructed using the production time, bottom hole pressure, oil production, and water production characteristics of five individual wells.

[0110] In this embodiment, the Long Short-Term Memory (LSTM) deep neural network consists of an input layer, two LSTM layers, and an output layer. 75% of the historical dataset is used as the training set, and 25% as the test set. A production dynamic prediction model is constructed through training and testing. The relevant parameters are: the first LSTM layer has 120 neurons, the second LSTM layer has 90 neurons, the learning rate is 0.1, the number of training iterations is 300, and the length of the cyclic input time series is 12.

[0111] The established production dynamics prediction model was used to fit the historical production dynamics data of well T101 to predict the water production and oil production (output layer) of well T101. The change of the model loss function with the number of training iterations during the training process is as follows: Figure 2 As shown, the training error decreases with increasing training iterations. The fitting results for oil production and water production are as follows: Figures 3-4 As shown, the actual data and the predicted data are in high agreement.

[0112] Example 2

[0113] like Figure 1 As shown, a method for predicting the dynamic production of a single well based on a long short-term memory deep neural network includes:

[0114] Obtain the input historical dataset X and the output historical dataset Y;

[0115] Extract the input training set X1 and the input test set X2 from the input historical dataset X, and apply the long short-term memory deep neural network based on the input training set X1 and the input test set X2 to obtain the production dynamic prediction model and the output test prediction set Y2.

[0116] Based on the output historical dataset Y and the output test prediction set Y2, the production dynamic prediction model that meets the expected effect is selected by the root mean square error function.

[0117] Obtain the production parameters and constraints of a single well, and apply optimization algorithms to generate multiple injection and production schemes.

[0118] For each injection and production scheme, historical production dynamic data corresponding to each scheme are obtained, and the production dynamic prediction model and the oilfield economic net present value function are applied sequentially to obtain the net present value corresponding to each injection and production scheme, thereby selecting the optimal injection and production scheme.

[0119] The specific implementation method is as follows:

[0120] Reservoir L is a deep-water turbidite reservoir driven by edge water, with a burial depth of 4279m, a reservoir pressure of 49.2MPa, a saturation pressure of 17.4MPa, and a reservoir temperature of 74℃, belonging to a normal temperature and pressure system. The reservoir has an average porosity of 17%, an average permeability of 446mD, and a formation crude oil viscosity of 0.89mPa·s, classifying it as a medium-high porosity, high-permeability, thin oil reservoir. A production dynamic scheme is planned for reservoir L, aiming to maximize the net present value (NPV) function of the oilfield under this scheme after 20 years of continuous production. The basic scheme maintains the production dynamics unchanged at the last moment and continues production for 20 years, with the production regime as follows: Figure 5 As shown.

[0121] The calculation parameters of the production dynamic prediction model are shown in Tables 1 and 2. Table 1 shows the single-well production system constraint parameters and NPV-related parameters. 75% of the historical dataset was used as the training set, and 25% as the test set. The Long Short-Term Memory (LSTM) deep neural network used a two-layer LSTM network, with 32 neurons in the first layer and 64 neurons in the second layer. The learning rate was 0.1, the training iterations were 3000, and the cyclic input time series length was 3. The prediction results for the training and test sets are shown below. Figure 6 As shown, the overall accuracy of production dynamics prediction is 0.92. An injection and extraction scheme is generated based on the particle swarm optimization algorithm. A production dynamics model established using a long short-term memory deep neural network is applied to calculate the output. The production regime is adjusted multiple times during the production period, and the final optimized result is taken as the optimal scheme.

[0122] Table 1. Constraints on Production System for Single Wells

[0123] <![CDATA[Maximum injection rate, m 3 / d]]> 4050 <![CDATA[Minimum injection speed, m 3 / d]]> 0 <![CDATA[Maximum oil production rate, m 3 / d]]> 3000 <![CDATA[Minimum oil production rate, m 3 / d]]> 0 <![CDATA[Maximum liquid production rate, m 3 / d]]> 3000 <![CDATA[Minimum liquid production rate, m 3 / d]]> 0

[0124] Table 2 NPV related parameters

[0125] Decrease rate, decimal 0.1 Oil price, USD / cubic meter 385 Water injection price, USD / cubic meter 20 Wastewater treatment cost, USD / cubic meter 40 Total drilling cost, USD / cubic meter 37500

[0126] The optimal production system is as follows: Figure 7 As shown, adjustments are made monthly, with a predicted total of 235 adjustments over 20 years. A comparison of the production dynamics results of the optimized scheme with the basic scheme is presented below. Figure 8 and Figure 9 As shown, the NPV gain was 340 million RMB, cumulative oil production increased by 290.3 × 10⁴ m³, cumulative water production decreased by 238.7 × 10⁴ m³, recovery rate increased by 7.7%, and water cut decreased by 4.5%.

[0127] The single-well production dynamic prediction method of this invention takes 2.6 hours to calculate 3,000 schemes, with an average of 3.12 seconds per scheme. In contrast, conventional numerical simulation technology takes 1 to 2 hours. Therefore, the single-well production dynamic prediction method of this invention greatly improves the timeliness of reservoir development scheme prediction.

[0128] Example 3

[0129] This disclosure provides an electronic device comprising: a memory storing executable instructions; and a processor executing the executable instructions in the memory to implement the aforementioned method for dynamic prediction of single-well production based on a long short-term memory deep neural network.

[0130] An electronic device according to an embodiment of the present disclosure includes a memory and a processor.

[0131] This memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0132] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory.

[0133] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0134] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0135] Example 4

[0136] This disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for dynamic prediction of single-well production based on a long short-term memory deep neural network.

[0137] A computer-readable storage medium according to embodiments of the present disclosure stores non-transitory computer-readable instructions. When these non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the methods described in the foregoing embodiments of the present disclosure are performed.

[0138] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0139] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A method for predicting the dynamic production of a single well based on a long short-term memory deep neural network, characterized in that, Including: Obtain an input historical data set X and an output historical data set Y; Extract an input training set X1 and an input test set X2 from the input historical data set X respectively, and based on the input training set X1 and the input test set X2, apply a long short-term memory deep neural network to obtain a production dynamics prediction model and an output test prediction set Y2; Based on the output historical data set Y and the output test prediction set Y2, screen out a production dynamics prediction model that meets the expected effect through a root mean square error function; the root mean square error function is: ， Where RMSE is the error loss, M is the total number of responses extracted from the output test prediction set Y2, and K... i T is the response value extracted from the output test prediction set Y2. i K is extracted from the output historical dataset Y i The corresponding target value, N is the total number of data in the output test prediction set Y2; if the error loss is within a preset range, the production dynamic prediction model meets the expected effect; Obtain single-well production parameters and single-well production parameter constraint conditions, and apply an optimization algorithm to generate multiple injection-production plans; For each injection-production plan, obtain the historical production dynamics data corresponding to each injection-production plan, and sequentially apply the production dynamics prediction model and the oilfield economic net present value function to obtain the net present value corresponding to each injection-production plan, so as to select the optimal injection-production plan; the oilfield economic net present value function is: ， Where NPV is the net present value, r is the discount rate, M1 represents the total production time, N1 is the total number of wells, and C d H represents the average drilling cost per unit length. i Let C be the depth of the i-th well. c For well completion costs, p o Q represents the selling price of crude oil. t,o Let p be the crude oil production on day t. wp The cost per unit of water collection, p wi The cost of water injection per unit, Q t,wp Let Q be the amount of water extracted on day t. t,wi Let t be the amount of water injected on day t.

2. The single-well production dynamics prediction method based on a long short-term memory deep neural network according to claim 1, wherein The obtaining the input historical data set X and the output historical data set Y includes: Construct a historical data set; Extract the data corresponding to the (t - 1)-th day, the t-th day, and the (t + 1)-th day from the historical data set as input historical data, and extract the data corresponding to the (t + 2)-th day from the historical data set as output historical data, where 1 < t < Q - 1, Q is the total production time and t is an integer; Form the input historical data set X with the input historical data, and form the output historical data set Y with the output historical data.

3. The single-well production dynamics prediction method based on a long short-term memory deep neural network according to claim 2, wherein The historical data set includes a reservoir static data set and a production dynamics data set, the input historical data set X includes an input historical reservoir static data set and an input historical production dynamics data set, and the output historical data set Y includes an output historical reservoir static data set and an output historical production dynamics data set.

4. The single-well production dynamics prediction method based on a long short-term memory deep neural network according to claim 2, wherein After constructing the historical data set, it includes: Clean the historical data set; ， Normalize the cleaned historical data set using the following formula, the formula is: where x is the data in the historical data set, mean(x) is the average value of the data x, std(x) is the standard deviation of the data x, and y is the normalized data.

5. The single-well production dynamics prediction method based on a long short-term memory deep neural network according to claim 1, wherein The extracting the input training set X1 and the input test set X2 from the input historical data set X respectively, and based on the input training set X1 and the input test set X2, applying a long short-term memory deep neural network to obtain a production dynamics prediction model and an output test prediction set Y2 includes: A predetermined proportion of data is extracted from the input historical dataset X as the input training set X1, and the input training set X1 is used to train the long short-term memory deep neural network to obtain the production dynamic prediction model. The remaining proportion data is extracted from the input historical dataset X as the input test set X2, and the input test set X2 is input into the production dynamic prediction model to obtain the output test prediction set Y2.

6. The method for predicting the dynamic production of a single well based on a long short-term memory deep neural network according to claim 1, characterized in that, For each injection and production scheme, historical production dynamic data corresponding to each scheme is obtained, and the production dynamic prediction model and the oilfield economic net present value function are applied sequentially to obtain the net present value corresponding to each scheme, thereby selecting the optimal injection and production scheme, including: For each injection and extraction scheme, the production dynamic prediction model is applied to perform production dynamic prediction to obtain the predicted production dynamic data corresponding to each injection and extraction scheme. Historical production dynamic data corresponding to each injection and production scheme is obtained, and based on the historical production dynamic data and predicted production dynamic data corresponding to each injection and production scheme, the net present value corresponding to each injection and production scheme is calculated through the oilfield economic net present value function. The injection and extraction scheme with the highest net present value is selected as the optimal injection and extraction scheme.

7. An electronic device, characterized in that, The electronic device includes: Memory, which stores executable instructions; A processor that executes the executable instructions in the memory to implement the single-well production dynamic prediction method based on a long short-term memory deep neural network according to any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the single-well production dynamic prediction method based on a long short-term memory deep neural network as described in any one of claims 1-6.

Images