Intelligent regulation and control method for full-well flow of separated layer water injection of oil field

Abstract

The invention discloses a full-well flow intelligent regulation and control method for separated layer water injection of an oil field. In the initial stage of water injection of a new well, a historical expert empirical model is adopted to regulate and control the flow of each layer section of the whole well, so that the actual injection real-time flow of each layer section is matched with the set injection allocation amount, and steady-state working condition point automatic identification is carried out based on adjustment process data to extract and obtain steady-state adjustment working condition data; a full-well prediction neural network model is built through steady-state adjustment working condition data training, the prediction precision of the full-well prediction neural network model is judged in real time, and when the prediction precision reaches the set control prediction precision, the full-well prediction neural network model is combined with a feedforward controller and a feedback PID controller to achieve intelligent control over full-well flow prediction. The problems of time lag and nonlinearity of the system are solved by means of large-step coarse adjustment of the feed-forward controller and fine adjustment of the feedback PID controller, mutual interference during flow adjustment of all layers is reduced, rapid allocation of multi-layer flow is achieved, the flow adjustment rate of the whole well is increased, and the action time and frequency of an underground execution mechanism are reduced.

Description

Technical Field

[0001] The present invention relates to an intelligent control method for the whole well flow rate of stratified water injection in oil fields, belonging to the technical field of stratified injection of water injection wells in oil fields. Background Art

[0002] With the continuous development of oil fields, formation energy will be continuously consumed. At present, the vast majority of oil fields have entered the middle and late stages of water injection development. The oil field water injection production technology can maintain the reservoir pressure and reduce the decline rate of crude oil production, and is the most economical and effective technical means to achieve long-term high and stable production of oil fields. In the stratified water injection process, how to accurately and quickly obtain the required injection volume for each layer is one of the key technologies to ensure efficient water injection in oil fields. However, due to the variety of factors affecting the injection volume of each layer and the non-linear and uncertain influence between various factors, this greatly increases the difficulty of accurate downhole water injection. Another key technology for efficient water injection in oil fields is the accurate measurement and precise control of the flow rate of each layer. During the process of regulating the flow rate of stratified water injection in the oil field downhole, the flow rate regulation is affected by various adverse factors such as the change of nozzle opening, pressure change, and mutual interference between different water injection layers. To achieve precise control of the flow rate, a controller with strong anti-interference ability, excellent dynamic performance, and strong stability needs to be designed.

[0003] At present, most of the stratified water injection flow rate regulation adopts repeated manual adjustment of the nozzle opening of downhole injectors to meet the requirement of matching the actual water injection flow rate with the injection volume, which has problems such as large flow rate regulation difficulty, heavy workload, and low efficiency. With the installation of stratified water injection systems in a large number of water injection wells in oil fields, the rapid, automatic, and intelligent regulation of the total well water injection volume has become an urgent technical requirement for oil fields. To effectively improve the efficiency of multi-layer flow rate regulation, Patent CN114357852A proposes to optimize and obtain the injection flow rate of each water injection layer by using long short-term memory neural network and particle swarm algorithm, but a large amount of data is required for model training, on-site data collection is difficult, and there is a lack of automatic data acquisition and analysis process. CN110533344A provides an automatic regulation method for a stratified water injection system, which evaluates the water injection layer by using water injection parameters and determines the adjustment sequence of each water injection layer according to the evaluation results. This method improves the regulation efficiency to a certain extent, but still requires manual repeated adjustment, and the adjustment workload is relatively large. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent control method for the whole well flow rate of stratified water injection in oil fields. This method solves the problems of large workload, low efficiency, and poor effect of repeated manual adjustment during the process of stratified water injection flow rate regulation, and further improves the efficiency and accuracy of multi-layer flow rate measurement and adjustment.

[0005] The technical solution of the present invention: An intelligent control method for the whole well flow rate of stratified water injection in oil fields, comprising the following steps:

[0006] S1: In the initial stage of flow regulation in a new well, use the historical expert experience model to perform flow stratification regulation on the total well flow rate, and adjust the total well flow rate in a single-layer section step-by-step manner according to the set injection volume of each layer section;

[0007] S2: Collect the data during the adjustment process, including the actual injection real-time flow rate, tubing pressure, formation pressure, set injection volume, and injection nozzle opening of each layer section; perform adaptive identification and extraction on the adjustment process data based on the steady-state identification method to obtain the steady-state adjustment working condition data under different adjustment states;

[0008] S3: Design and construct a full-well prediction neural network model, perform data preprocessing on the steady-state adjustment working condition data obtained in S2, and use the preprocessed data to train the full-well prediction neural network model;

[0009] S4: Evaluate the prediction accuracy of the full-well prediction neural network model. When the prediction accuracy is less than the set control prediction accuracy, use the historical expert experience model to stably regulate the actual injection real-time flow rate of each layer section, and at the same time continue to collect the steady-state adjustment working condition data of the adjustment process and continue to optimize and train the full-well prediction neural network model. When the prediction accuracy reaches the set control prediction accuracy, use the neural network intelligent control module to perform intelligent regulation on the actual injection real-time flow rate of each layer section of the whole well, and realize the combined optimization calculation of the injection nozzle opening of multiple layer sections;

[0010] S5: Roll and execute the process of S2, collect, identify, and extract the steady-state adjustment working condition data of the adjustment process in real time, judge the prediction accuracy of the full-well prediction neural network model in real time. When the prediction accuracy is less than the set control prediction accuracy, start the model rolling optimization training mechanism, introduce the latest steady-state adjustment working condition data for model training, and ensure that the real-time prediction accuracy of the model is greater than the set control prediction accuracy.

[0011] In the above intelligent regulation method for the total well flow rate of oilfield layered water injection, the step S1 specifically includes:

[0012] S11: The method of using the historical expert experience model for flow regulation is as follows: First, the wellhead adopts a constant flow control method for the total injection volume to ensure that the total injection flow rate of the whole well is the same as the sum of the set injection volumes of each layer section. Determine the under-injected layer section and over-injected layer section according to the positive and negative deviations between the actual injection real-time flow rate, set injection volume, and allowable error E of each layer section; if the sum of the set injection volume of a certain layer section and the allowable error E minus the actual injection real-time flow rate of that layer section is negative, it means that this layer section is an over-injected layer section. If the difference between the set injection volume of a certain layer section and the allowable error E minus the actual injection real-time flow rate of that layer section is positive, it means that this layer section is an under-injected layer section; record the numbers of the under-injected layer sections and over-injected layer sections of the whole well as and , determine the adjustment direction according to the sizes of the numbers of the under-injected layer sections and over-injected layer sections; when Greater than When the adjustment direction is to adjust the underfilled layer first, then adjust the overfilled layer, and vice versa. Less than When adjusting, the direction is to adjust the over-injection layer first, and then adjust the under-injection layer;

[0013] S12: According to the adjustment direction determined in step S11, the flow rate of the layer sections that meet the adjustment direction is automatically adjusted in the order of layer positions from top to bottom. When the deviation between the actual real-time injection flow rate of a layer section and the set injection amount of the layer section is less than the allowable deviation E, the adjustment of the layer section is stopped, the opening of the water injection nozzle of the layer section is maintained at the current opening, and the flow rate of the next layer section that meets the adjustment direction is automatically adjusted until the deviation between the actual real-time injection flow rate of the layer section and the set injection amount of the layer section is less than the allowable deviation E. After all layers in the adjustment direction are adjusted to be qualified, switch to another adjustment direction, that is, adjust the under-injection layer section after the over-injection layer section is adjusted to be qualified, or adjust the over-injection layer section after the under-injection layer section is adjusted to be qualified, until all layers are adjusted to be qualified, that is, the number of under-injection layers and over-injection layers in the whole well is 0.

[0014] In the aforementioned method for intelligently controlling the full-well flow rate of oilfield stratified water injection, step S2 specifically includes:

[0015] S21: Steady-state identification method is: setting the change thresholds of tubing pressure and formation pressure to be and ; The actual stable threshold of the real-time flow parameter is , the stability thresholds of tubing pressure and formation pressure parameters are , the maximum data length of the steady-state data judgment data window is M, and the stable judgment variable sliding window is N;

[0016] S22: Read the set injection volume of each layer, the actual injection flow rate, the water injection nozzle opening, the tubing pressure and the formation pressure parameters, clean the data, and eliminate abnormal zero-value data or flying point data;

[0017] S23: When the formation pressure change in a certain layer is greater than the formation pressure change threshold Or the tubing pressure change in each layer of the whole well is greater than the tubing pressure change threshold When the opening degree of the water injection nozzle in a certain layer or section changes by more than 5%, starting from the data change point, continuously read N consecutive data and put them into the variable sliding window for stability judgment. The data parameters read include the actual injection real-time flow rate, tubing pressure, and formation pressure. Calculate the mean values of the N data for each data parameter obtained respectively in the variable sliding window for stability judgment. Calculate the proportion of the number of data whose deviation from the mean value in the N data is less than the stability threshold of each parameter. When the proportion is less than the set judgment threshold, the variable sliding window for stability judgment slides backward by Z data points, where Z < N, and then make a judgment until the proportion of the number of data whose deviation from the mean value in the variable sliding window for stability judgment is less than the stability threshold is greater than the set judgment threshold;

[0018] S24: The variable sliding window for stability judgment continues to slide backward by N data points. If the deviation of all data in the variable sliding window for stability judgment from their mean value is less than the stability threshold, save and record all data in the variable sliding window for stability judgment into the steady-state data judgment data window; then continue to slide backward by N data points for judgment. When the proportion of data whose deviation from their mean value in the variable sliding window for stability judgment is less than the stability threshold is less than the set judgment threshold or the total number of data recorded in the steady-state data judgment data window is greater than M, exit the sliding judgment process of the variable sliding window for stability judgment, take the mean value of all data in the steady-state data judgment data window as the final steady-state adjustment working condition data, and store this value into the database.

[0019] In the intelligent flow rate regulation method for layered water injection in an oilfield described above, step S3 specifically includes:

[0020] S31: The full-well prediction neural network model uses a long short-term memory convolutional neural network to predict the time series data of the opening degree of the water injection nozzle and the water injection flow rate. The preprocessed steady-state adjustment working condition data is used as the input of the front-end convolutional neural network. The automatic extraction of multi-dimensional input data features is realized through the convolutional layer and pooling layer of the convolutional neural network. The feature extraction result of the convolutional neural network is used as the input of the back-end bidirectional long short-term memory network layer. The output result of the bidirectional long short-term memory network layer obtains different weights through the attention mechanism, and after normalization processing, it is output to the fully connected layer to obtain the final prediction output result. The output of the full-well prediction neural network model is the predicted values of the opening degree of the water injection nozzle and the water injection flow rate for each layer or section;

[0021] S32: The preprocessing method for the steady-state adjustment working condition data adopts linear normalization processing. Linear normalization processing is respectively performed on the steady-state adjustment working condition data of the pressure, actual injection real-time flow rate, and opening degree of the water injection nozzle for each layer or section. The processing formula is as follows:

[0022] ,

[0023] represents the original signal; Indicates the result after normalization; Indicates the minimum value of the original signal; Indicates the maximum value of the original signal.

[0024] In the above-mentioned intelligent flow regulation method for whole-well layered water injection in oilfields, the specific steps of step S4 are as follows:

[0025] S41: The prediction accuracy calculation method calculates the mean square error by taking continuous K points. When the result of the mean square error is less than the given value , it indicates that the prediction accuracy meets the training requirements;

[0026] ,

[0027] Indicates the original signal with noise, Indicates the model prediction signal, Indicates the data length, Indicates the mean square error;

[0028] S42: The neural network intelligent control module combines the whole-well prediction neural network model, the feed-forward controller, and the feedback PID controller. The product of the predicted value of the water injection nozzle opening of each layer section of the whole-well prediction neural network model and the output result of the input set value judgment module is used as the output of the feed-forward controller. The deviation between the predicted value of the water injection flow of each layer section of the whole-well prediction neural network model and the set injection volume of each layer section is used as the input value of the feedback PID controller. Calculate and output the feedback adjustment amount of the water injection nozzle opening of each layer section. The sum of the feedback adjustment amount of the water injection nozzle opening of each layer section and the output of the feed-forward controller is used as the final control amount output and controls the water injection nozzle opening of each layer section of the multi-layered water injection control object;

[0029] S43: The input set value judgment module is used to judge whether there is a step change in the set value of the injection volume of each layer section. The judgment amplitude is , that is, when the change amplitude of the set injection volume of the input value is greater than or equal to , it indicates that a step change has occurred, and the output of the input set value judgment module is 1; when the change amplitude of the input value is less than , the output of the input set value judgment module is 0.

[0030] Advantages of the present invention: Compared with the prior art, the present invention has the following advantages:

[0031] 1. The injector is placed in the wellbore for a long time. The harsh environment of high temperature and high pressure in the wellbore may cause damage to the injector sensors. However, due to the high operation cost, when the flow sensors of one or two layers in a multi-layer injection well show abnormalities, the abnormal injector will not be immediately removed for operation. The full-well prediction neural network model designed by the present invention can, under the condition of abnormal sensors in a certain layer, realize real-time prediction of the flow rate through the historical data of this layer and the flow opening data of other layers, ensuring the normal operation of the system.

[0032] 2. The present invention combines the empirical method with the artificial intelligence neural network. First, the full-well flow rate is adjusted step by step through the empirical method to meet the flow rate regulation requirements of each layer in the field. At the same time, as many steady-state regulation working condition data as possible under different regulation states are collected. Then, the full-well prediction neural network model is trained with the steady-state regulation working condition data under different regulation states. The full-well prediction neural network model is used to overcome the interference caused by formation pressure changes and inter-layer flow rate changes, further improving the full-well flow rate regulation efficiency and ensuring the accuracy and rapidity of the flow rate regulation.

[0033] 3. The model has the self-learning ability. As the process data accumulates continuously, the full-well prediction neural network model is optimized through rolling training, and its prediction accuracy will be higher and higher, realizing "one-to-one" precise intelligent control of a single well, effectively improving the full-well layered water injection efficiency, reducing the adjustment times of downhole injectors, and increasing the service life of downhole instruments. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a block diagram of the full-well intelligent regulation method system;

[0035] Figure 2 It is a logic flow chart for automatic identification and extraction of steady-state data;

[0036] Figure 3 It is a structural diagram of the full-well flow neural network prediction model;

[0037] Figure 4 It is a block diagram of the composition of the neural network intelligent control module system. SPECIFIC EMBODIMENTS

[0038] The present invention will be further described below in conjunction with the drawings and embodiments, but it shall not be used as a basis for limiting the present invention.

[0039] Embodiments of the present invention: An intelligent control method for the whole well flow rate in stratified water injection in oil fields. The control method includes establishing a multi-step approximation adjustment method for the whole well flow rate based on a historical expert experience model. First, at the initial stage of water injection in a new well, the historical expert experience model is used to control the flow rate of each layer section of the whole well, so that the actual injection real-time flow rate of each layer section matches the set injection allocation rate. Collect the data of the whole well adjustment process, automatically identify the steady-state working condition points based on the adjustment process data to extract the steady-state adjustment working condition data, use the steady-state adjustment working condition data to train and construct a whole well prediction neural network model, and judge the prediction accuracy of the whole well prediction neural network model in real time. When the prediction accuracy of the whole well prediction neural network model reaches the set control prediction accuracy, the whole well prediction neural network model is combined with a feedforward controller and a feedback PID controller to realize intelligent control of the whole well flow rate prediction. The feedforward controller is used for large-step rough adjustment and the feedback PID controller is used for fine adjustment to handle the time delay and nonlinear problems of the system, reduce the mutual interference during the adjustment of each layer flow rate, realize the rapid allocation of multi-layer flow rates, improve the adjustment rate of the whole well flow rate, reduce the action time and frequency of downhole actuators, and extend the service life of downhole instruments.

[0040] Specifically, it includes the following steps:

[0041] S1: At the initial stage of flow rate control in a new well, use the historical expert experience model to conduct flow rate stratified control on the whole well flow rate, and adjust the whole well flow rate in a single-layer section step-by-step manner according to the set injection allocation rate of each layer section;

[0042] S2: Collect the adjustment process data, including the actual injection real-time flow rate, tubing pressure, formation pressure, set injection allocation rate, and water injection nozzle opening of each layer section; based on the steady-state recognition method, adaptively identify and extract the adjustment process data to obtain the steady-state adjustment working condition data under different adjustment states;

[0043] S3: Design and construct a whole well prediction neural network model, use the steady-state adjustment working condition data obtained in S2 for data preprocessing, and use the preprocessed data to train the whole well prediction neural network model;

[0044] S4: Evaluate the prediction accuracy of the whole well prediction neural network model. When the prediction accuracy is less than the set control prediction accuracy, use the historical expert experience model to stably control the actual injection real-time flow rate of each layer section, and at the same time continue to collect the steady-state adjustment working condition data of the adjustment process, and continue to optimize and train the whole well prediction neural network model. When the prediction accuracy reaches the set control prediction accuracy, use the neural network intelligent control module to intelligently control the actual injection real-time flow rate of each layer section of the whole well, realize the combined optimization calculation of the water injection nozzle opening of multiple layers, and achieve a fast and stable adjustment effect;

[0045] S5: Roll and execute the process of S2, collect, identify, and extract the steady-state regulation condition data of the regulation process in real time, and judge the prediction accuracy of the full-well prediction neural network model in real time. When the prediction accuracy is less than the set control prediction accuracy, start the model rolling optimization training mechanism, introduce the latest steady-state regulation condition data for model training, and ensure that the real-time prediction accuracy of the model is greater than the set control prediction accuracy.

[0046] The specific steps of step S1 include:

[0047] S11: The method of using the historical expert experience model for flow regulation is as follows: First, the wellhead adopts a constant flow control method for the total injection volume to ensure that the total injection flow rate of the whole well is the same as the sum of the set injection volumes of each interval. Determine the under-injected intervals and over-injected intervals according to the positive and negative deviations between the actual real-time injection flow rate, the set injection volume, and the allowable error E of each interval; if the sum of the set injection volume of a certain interval and the allowable error E minus the actual real-time injection flow rate of this interval is negative, it means that this interval is an over-injected interval. If the difference between the set injection volume of a certain interval and the allowable error E minus the actual real-time injection flow rate of this interval is positive, it means that this interval is an under-injected interval; record the numbers of the under-injected intervals and over-injected intervals of the whole well as and , and determine the regulation direction according to the magnitudes of the numbers of the under-injected intervals and over-injected intervals; when is greater than , the regulation direction is to first regulate the under-injected intervals and then the over-injected intervals. On the contrary, when is less than , the regulation direction is to first regulate the over-injected intervals and then the under-injected intervals;

[0048] S12: According to the regulation direction determined in step S11, automatically regulate the flow rate of the intervals that meet this regulation direction in the order from top to bottom of the interval layer positions. When the deviation between the actual real-time injection flow rate of a certain interval and the set injection volume of this interval is less than the allowable deviation E, stop the regulation of this interval, and keep the opening of the injection nozzle of this interval at the current opening. Automatically regulate the flow rate of the next interval that meets this regulation direction until the deviation between the actual real-time injection flow rate of this interval (i.e., the interval being regulated) and the set injection volume of this interval (i.e., the interval being regulated) is less than the allowable deviation E. After all the intervals in this regulation direction are regulated qualifiedly, switch to another regulation direction, that is, after the over-injected intervals are regulated qualifiedly, regulate the under-injected intervals, or after the under-injected intervals are regulated qualifiedly, regulate the over-injected intervals, until all the intervals are regulated qualifiedly, that is, the numbers of the under-injected intervals and over-injected intervals of the whole well are both 0.

[0049] The specific steps of step S2 include:

[0050] S21: The steady-state identification method is: Set the change thresholds of the tubing pressure and the formation pressure to be and ; The actual stable threshold of the real-time flow parameter is , the stability thresholds of tubing pressure and formation pressure parameters are , the maximum data length of the steady-state data judgment data window is M, and the stable judgment variable sliding window is N;

[0051] S22: Read the set injection volume of each layer, the actual injection flow rate, the water injection nozzle opening, the tubing pressure and the formation pressure parameters, clean the data, and eliminate abnormal zero-value data or flying point data;

[0052] S23: When the formation pressure change in a certain layer is greater than the formation pressure change threshold Or the tubing pressure change in each layer of the whole well is greater than the tubing pressure change threshold Or when the opening of the water injection nozzle in a certain layer section changes by more than 5%, N continuous data are continuously read with the data change point as the starting point and put into the stability judgment variable sliding window. The read data parameters include the actual real-time injection flow rate, tubing pressure and formation pressure. The mean of each data parameter obtained in the N data in the stability judgment variable sliding window is calculated respectively, that is, the mean of N real-time injection flow rate data, the mean of N tubing pressure data and the mean of N formation pressure data are calculated. The proportion of the number of data in the N data whose deviation from the mean is less than the stability threshold of each parameter is calculated. When the proportion is less than the set judgment threshold, the stability judgment variable sliding window slides back Z data points, and Z <N,再进行判断,直到稳定判断可变滑动窗中数据均值的偏差小于稳定阈值的数据数量占比大于设定判断阈值;

[0053] For example, when N injected real-time traffic data are divided into A1, A2, A3...A N , and the mean of N injected real-time traffic data is A 均 When A1-A 均 、A2-A 均 、A3-A 均 ...A N -A 均 The absolute value of , thus obtaining N absolute values, and then observing how many of the N absolute values are smaller than the stable threshold of the injected real-time traffic data , thus calculating the stability threshold value of the deviation of N data from the mean value less than the real-time traffic data The same technical approach is used to analyze the data on pipeline pressure and formation pressure.

[0054] S24: The stable judgment variable sliding window continues to slide backward by N data points. If the deviation of all data within the stable judgment variable sliding window from its mean value is less than the stable threshold, all data within the stable judgment variable sliding window are saved to the steady-state data judgment data window; then continue to slide backward by N data points for judgment. When the proportion of data within the stable judgment variable sliding window whose deviation from its mean value is less than the stable threshold is less than the set judgment threshold or the total amount of data recorded in the steady-state data judgment data window is greater than M, the sliding judgment process of the stable judgment variable sliding window is exited, and the mean value of all data in the steady-state data judgment data window is used as the final steady-state regulation working condition data, and this value is stored in the database.

[0055] The specific steps of step S3 are as follows:

[0056] S31: The full-well prediction neural network model uses a long short-term memory convolutional neural network to predict the water injection nozzle opening and water injection flow time series data. The preprocessed steady-state regulation working condition data is used as the input of the front-end convolutional neural network. The automatic extraction of multi-dimensional input data features is realized through the convolutional layer and pooling layer of the convolutional neural network. The feature extraction result of the convolutional neural network is used as the input of the back-end bidirectional long short-term memory network layer. The output result of the bidirectional long short-term memory network layer obtains different weights through the attention mechanism, and after normalization, it is output to the fully connected layer to obtain the final prediction output result. The output of the full-well prediction neural network model is the predicted values of the water injection nozzle opening and water injection flow for each layer section. The bidirectional long short-term memory network layer enables the model to selectively retain, forget, and output valid information by setting an input gate, a forget gate, an output gate, and a memory unit, ensuring the real-time performance and accuracy of the full-well prediction neural network model;

[0057] S32: The preprocessing method of the steady-state regulation working condition data uses linear normalization processing. Linear normalization processing is respectively performed on the steady-state regulation working condition data of the pressure, actual injection real-time flow, and water injection nozzle opening for each layer section. The processing formula is as follows:

[0058] ,

[0059] represents the original signal; represents the result after normalization; represents the minimum value of the original signal; represents the maximum value of the original signal.

[0060] The specific steps of step S4 are as follows:

[0061] S41: The prediction accuracy calculation method calculates the mean square error by taking continuous K points. When the mean square error result is less than the given value , it indicates that the prediction accuracy meets the training requirements;

[0062] ,

[0063] represents the noisy original signal, represents the model prediction signal, K represents the data length, represents the mean square error;

[0064] S42: The neural network intelligent control module combines the full-well prediction neural network model, the feedforward controller, and the feedback PID controller. The product of the predicted value of the injection nozzle opening of each section of the full-well prediction neural network model and the output result of the input set value judgment module is used as the output of the feedforward controller. The deviation between the predicted value of the injection flow rate of each section of the full-well prediction neural network model and the set injection volume of each section is used as the input value of the feedback PID controller, and the feedback adjustment amount of the injection nozzle opening of each section is calculated and output. The sum of the feedback adjustment amount of the injection nozzle opening of each section and the output of the feedforward controller is used as the final control amount and output to control the injection nozzle opening of each section of the multi-layer stratified injection control object;

[0065] S43: The input set value judgment module is used to judge whether there is a step change in the set value of the injection volume of each section. The judgment amplitude is , that is, when the change amplitude of the set injection volume of the input value is greater than or equal to , it indicates that a step change has occurred, and the output of the input set value judgment module is 1; when the change amplitude of the input value is less than , the output of the input set value judgment module is 0.

[0066] In the specific implementation process, a field test application was carried out on a certain four-layer new well in the Third Factory of Daqing Oilfield. According to the full-well intelligent regulation method shown in Figure 1 , first, the historical expert experience model was used to regulate the flow rate of the whole well in a stratified manner. According to the injection volume of each layer, the flow rate of the whole well was adjusted in a single-layer step-by-step manner. At the same time, the real-time adjustment process data was collected, and the steady-state adjustment condition data was automatically identified using the steady-state data automatic identification and extraction logic shown in Figure 2 . The identified data was automatically added to the model database. A total of 763 effective data samples were collected in the model database. 500 of these effective sample data were used as the model training input data set, and 263 data were used as the model verification data set. Based on the model structure shown in Figure 3 , the full-well prediction neural network model was trained. The model after 1000 iterations of training was tested three times using the verification data set. The test results showed that the prediction accuracies were 96.08%, 96.73%, and 98.1% respectively. The control prediction accuracy was set to 96%. When the prediction accuracy of the full-well prediction neural network model reached the control prediction accuracy, the method shown in Figure 4The neural network intelligent control module shown performs intelligent regulation on the actual real-time injection flow rate of the whole well. Due to the change of formation characteristics, the actual real-time injection flow rate and the set injection allocation requirements of each interval of this well are shown in Table 1: The single-layer step-by-step adjustment method of the historical expert experience model and the intelligent regulation method of the neural network intelligent control module are respectively used for the whole-well flow rate regulation. The actual real-time injection flow rate and the total adjustment time after adjustment are shown in Table 2:

[0067] Table 1 Injection Allocation Requirements

[0068] Horizon First layer Second layer Third layer Fourth layer <![CDATA[Actual real-time injection flow rate before adjustment (m 3 / d)]]> 19.6 34.3 15.7 44.0 <![CDATA[Set dispensing volume (m 3 / d)]]> 40 15 30 30

[0069] Table 2 Adjustment Effect Table

[0070] Regulation method First layer Second layer Third layer ​ ​ ​ 38.2 14 32.6 28.4 ​ ​ 39.3 14.7 29.8 29.1 ​

[0071] It can be seen from Table 2 that the automatic regulation method of the whole-well flow rate using the neural network intelligent control module has a shorter adjustment time and higher regulation efficiency than the adjustment method using the historical expert model experience method. In addition, the regulation accuracy is better than the experience method.

[0072] In order to solve the problems of large workload, low efficiency and poor effect of repeated manual adjustment in the process of layered water injection flow rate regulation, the intelligent regulation method of the whole-well flow rate of layered water injection proposed by the present invention is of great significance for improving the efficiency and accuracy of multi-layer flow rate measurement and adjustment.

Claims

1. An intelligent regulation method for the total well flow rate of stratified water injection in oil fields, characterized in that: The following steps are involved: S1: In the initial stage of flow control in a new well, the historical expert experience model is used to perform flow stratification control on the entire well flow. The flow rate of the entire well is adjusted by a single layer step-by-step adjustment method based on the set injection volume of each layer. S2: Collecting regulation process data, including the actual real-time injection flow rate of each layer, tubing pressure, formation pressure, set injection volume, and water injection nozzle opening; adaptively identifying and extracting the regulation process data based on the steady-state identification method to obtain steady-state regulation condition data under different regulation states; S3: Design and construct a full-well prediction neural network model, use the steady-state regulation condition data obtained in S2 to perform data preprocessing, and use the preprocessed data to train the full-well prediction neural network model; S4: Evaluate the prediction accuracy of the full-well prediction neural network model. When the prediction accuracy is less than the set control prediction accuracy, use the historical expert experience model to stably regulate the actual real-time injection flow of each layer. At the same time, continue to collect steady-state regulation data during the regulation process and continue to optimize and train the full-well prediction neural network model. When the prediction accuracy reaches the set control prediction accuracy, use the neural network intelligent control module to intelligently regulate the actual real-time injection flow of each layer in the whole well, and realize the combined optimization calculation of the multi-layer injection nozzle opening; S5: Execute the S2 process in a rolling manner to collect, identify and extract the steady-state adjustment condition data of the adjustment process in real time, and judge the prediction accuracy of the full-well prediction neural network model in real time. When the prediction accuracy is less than the set control prediction accuracy, start the model rolling optimization training mechanism and introduce the latest steady-state adjustment condition data for model training to ensure that the real-time prediction accuracy of the model is greater than the set control prediction accuracy.

2. The intelligent flow rate regulation method for the whole well of oilfield layered water injection according to claim 1, characterized in that: The step S1 specifically includes: S11: The method of flow rate adjustment using the historical expert experience model is as follows: First, the constant flow control mode of the total injection volume is adopted at the wellhead to ensure that the total injection flow rate of the whole well is the same as the sum of the set injection volumes of each interval. Determine the under-injected interval and over-injected interval according to the positive and negative deviations between the actual real-time injection flow rate, the set injection volume, and the allowable error E of each interval. If the sum of the set injection volume of a certain interval and the allowable error E minus the actual real-time injection flow rate of this interval is negative, it means that this interval is an over-injected interval. If the difference between the set injection volume of a certain interval and the allowable error E minus the actual real-time injection flow rate of this interval is positive, it means that this interval is an under-injected interval. Record the numbers of the under-injected intervals and over-injected intervals of the whole well as and , respectively, and determine the adjustment direction according to the magnitudes of the numbers of the under-injected intervals and over-injected intervals. When is greater than , the adjustment direction is to adjust the under-injected intervals first and then the over-injected intervals. On the contrary, when is less than , the adjustment direction is to adjust the over-injected intervals first and then the under-injected intervals. S12: According to the adjustment direction determined in step S11, the flow rate of the layer sections that meet the adjustment direction is automatically adjusted in the order of layer positions from top to bottom. When the deviation between the actual real-time injection flow rate of a layer section and the set injection amount of the layer section is less than the allowable deviation E, the adjustment of the layer section is stopped, the opening of the water injection nozzle of the layer section is maintained at the current opening, and the flow rate of the next layer section that meets the adjustment direction is automatically adjusted until the deviation between the actual real-time injection flow rate of the layer section and the set injection amount of the layer section is less than the allowable deviation E. After all layers in the adjustment direction are adjusted to be qualified, switch to another adjustment direction, that is, adjust the under-injection layer section after the over-injection layer section is adjusted to be qualified, or adjust the over-injection layer section after the under-injection layer section is adjusted to be qualified, until all layers are adjusted to be qualified, that is, the number of under-injection layers and over-injection layers in the whole well is 0.

3. The intelligent regulation method for the whole well flow rate of stratified water injection in an oilfield according to claim 1, wherein: The step S2 specifically includes: S21: The steady state recognition method is as follows: Set the change thresholds of the tubing pressure and formation pressure to be and respectively; The stability threshold of the actual injection real-time flow rate parameter is , the stability threshold of the tubing pressure and formation pressure parameters is , the maximum data length of the steady state data judgment data window is M, and the stable judgment variable sliding window is N; S22: Read the set injection volume of each layer, the actual injection flow rate, the water injection nozzle opening, the tubing pressure and the formation pressure parameters, clean the data, and eliminate abnormal zero-value data or flying point data; S23: When the formation pressure change of a certain interval is greater than the formation pressure change threshold or the tubing pressure change of each interval in the whole well is greater than the tubing pressure change threshold or when the opening change of the water injection nozzle of a certain interval is greater than 5%, starting from the data change point, continuously read N consecutive data and put them into the stable judgment variable sliding window. The read data parameters include the actual injection real-time flow rate, tubing pressure, and formation pressure. Calculate the mean values of the N data of each data parameter obtained respectively in the stable judgment variable sliding window. Calculate the proportion of the number of data with a deviation less than the stable threshold of each parameter among the N data. When the proportion is less than the set judgment threshold, the stable judgment variable sliding window slides backward by Z data points, where Z < N, and then make a judgment until the proportion of the number of data with a deviation less than the stable threshold of the data mean value in the stable judgment variable sliding window is greater than the set judgment threshold; S24: The stability judgment variable sliding window continues to slide backward by N data points. If the deviation between all the data within the stability judgment variable sliding window and its mean value is less than the stability threshold, then all the data within the stability judgment variable sliding window is saved and recorded into the steady-state data judgment data window; then continue to slide backward by N data points for further judgment. When the proportion of the data within the stability judgment variable sliding window whose deviation from its mean value is less than the stability threshold is less than the set judgment threshold or the total amount of all the data recorded in the steady-state data judgment data window is greater than M, exit the sliding judgment process of the stability judgment variable sliding window, take the mean value of all the data in the steady-state data judgment data window as the final steady-state regulation working condition data, and store this value into the database.

4. An intelligent flow rate control method for the whole well of stratified water injection in an oilfield according to claim 1, characterized in that: The specific steps of step S3 include: S31: The full-well prediction neural network model uses a long short-term memory convolutional neural network to predict the water injection nozzle opening and water injection flow time series data. The preprocessed steady-state regulation working condition data is used as the input of the front-end convolutional neural network. The automatic extraction of multi-dimensional input data features is realized through the convolutional layer and pooling layer of the convolutional neural network. The feature extraction result of the convolutional neural network is used as the input of the back-end bidirectional long short-term memory network layer. The output result of the bidirectional long short-term memory network layer obtains different weights through the attention mechanism, and after normalization processing, it is output to the fully connected layer to obtain the final prediction output result. The output of the full-well prediction neural network model is the predicted values of the water injection nozzle opening and water injection flow for each layer section; S32: The preprocessing method of the steady-state regulation working condition data adopts linear normalization processing. Linear normalization processing is respectively carried out on the steady-state regulation working condition data of the pressure, actual injection real-time flow, and water injection nozzle opening of each layer section. The processing formula is expressed as follows: ， represents the original signal; represents the result after normalization; represents the minimum value of the original signal; represents the maximum value of the original signal.

5. The intelligent regulation method for the whole-well flow rate of stratified water injection in oil fields according to claim 1, wherein: The specific steps of step S4 include: S41: The prediction accuracy calculation method calculates the mean square error by taking consecutive K points. When the result of the mean square error is less than the given value , it indicates that the prediction accuracy meets the training requirements; ， represents the noisy original signal, represents the model prediction signal, represents the data length, represents the mean square error; S42: The neural network intelligent control module combines the full-well prediction neural network model, the feed-forward controller, and the feedback PID controller. The product of the predicted value of the water injection nozzle opening of each layer section of the full-well prediction neural network model and the output result of the input set value judgment module is used as the output of the feed-forward controller. The deviation between the predicted value of the water injection flow of each layer section of the full-well prediction neural network model and the set injection volume of each layer section is used as the input value of the feedback PID controller. The feedback adjustment amount of the water injection nozzle opening of each layer section is calculated and output. The sum of the feedback adjustment amount of the water injection nozzle opening of each layer section and the output of the feed-forward controller is used as the final control amount output and controls the water injection nozzle opening of each layer section of the multi-layer hierarchical water injection control object; S43: The input set value judgment module is used to judge whether there is a step change in the injection volume set value of each layer section, and the judgment amplitude is , that is, when the change amplitude of the injection volume set by the input value is greater than or equal to , it indicates that a step change has occurred, and the output of the input set value judgment module is 1; when the change amplitude of the input value is less than , the output of the input set value judgment module is 0.

Images