Flow and water content prediction method based on distributed optical fiber sound wave monitoring data

Abstract

The invention provides a flow and water content prediction method based on distributed optical fiber sound wave monitoring data, which comprises the following steps: step 1, extracting DAS monitoring data of different layer sections of a production well, and preprocessing; 2, DAS sound wave frequency band energy FBE of each layer section is calculated; step 3, performing LSTM high-precision intelligent modeling to realize liquid production capacity prediction; 4, LSTM model sensitivity analysis is carried out, and a main liquid production section is determined; 5, verifying a sensitivity analysis result by utilizing DTS temperature data; step 6, taking the liquid production capacity as a main control factor influencing the water content, and remodeling to realize short-term synchronous prediction of the flow and the water content; and 7, performing field application, and evaluating an intelligent modeling effect. According to the method, future flow prediction of the field production well can be achieved, the main liquid production section of the multi-layer (section) oil well is determined, short-term synchronous prediction of the future flow and the water content of the oil well is achieved, and intelligent management and control of oil field injection and production are facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil and gas exploitation, and particularly relates to a flow rate and water cut prediction method based on distributed optical fiber acoustic wave monitoring data. BACKGROUND

[0002] The distributed optical fiber monitoring technology breaks through the limitations of traditional electronic logging technology, such as few measuring points, short time (3-6h), large size, and difficulty in lifting and lowering. With the advantages of distributed measuring points, dynamics, multi-fields, and high temperature and pressure resistance, the distributed optical fiber monitoring technology provides a real-time, reliable, and comprehensive solution for the whole life cycle monitoring of oil and water wells, and helps the digital construction and intelligent development of oilfields. At present, the distributed acoustic sensing (DAS) technology and the distributed temperature sensing (DTS) technology are gradually applied in oil well field monitoring, and have a very broad development prospect in downhole production dynamic monitoring. However, at the same time, the huge amount of data generated by the optical fiber DAS has brought difficulties to the application and development of the DAS technology.

[0003] At present, with the rapid development of big data and machine learning, intelligent modeling technology based on machine learning algorithms is gradually applied in optical fiber DAS monitoring data processing. In the Chinese patent application with the application number CN202110216451.9, a DAS signal high-precision classification and identification method based on model fusion is involved. In the DAS signal identification, the stacking strategy is introduced to integrate the decision tree model, the random forest model, the support vector machine model, and the extreme gradient boosting algorithm model in machine learning. The original data obtained by the DAS system is preprocessed, and then the features in multiple analysis domains are extracted. The artificial neural network is used to further extract and classify the features. The logistic regression is used to re-learn the prediction results of the above models, and the final prediction of the environment state of the current input signal is obtained. The identification method has higher identification rate and shorter identification time than the traditional method, and has great significance in the pursuit of real-time optical fiber distributed sensing detection.

[0004] In the Chinese patent application with application number: CN202210588497.8, a wellbore fluid type identification method based on distributed optical fiber acoustic wave monitoring data is disclosed. The acoustic wave monitoring data collected by the DAS system is processed to obtain relevant acoustic wave data after filtering and normalization. The acoustic wave feature parameters are extracted to obtain acoustic wave features related to the wellbore fluid type. The low-variance scattering feature set a and the short-time time-frequency feature set b are fused to form a fusion feature set c. The obtained acoustic wave features are input into a classification algorithm for training to form an identification model. The acoustic wave monitoring data is input into the identification model, and the wellbore fluid type is output. The low-variance scattering feature and the short-time time-frequency feature are combined to form a fusion feature for fluid type identification. The fusion feature is a multi-source feature that integrates the information contained in the low-variance scattering feature and the short-time time-frequency feature. It provides more data sources for accurate identification of wellbore fluid type, effectively solves the problem of low accuracy of single feature identification of fluid type, and greatly improves the identification accuracy.

[0005] In the Chinese patent application with application number: CN202311401894.0, a pipeline mode identification method and system based on distributed optical fiber sensing are disclosed. The method includes acquiring optical fiber vibration images corresponding to different events, performing data denoising and data cleaning on the acquired optical fiber vibration images, constructing a network identification model based on a one-stage mode, training the network identification model using the optical fiber vibration images, and identifying the pipeline mode using the trained network identification model. By collecting vibration signals occurring around the pipeline, extracting vibration features, and identifying vibration event types, timely alarms can be given to prevent intentional or unintentional pipeline damage, ensuring the safety of pipeline energy transportation.

[0006] In summary, using intelligent modeling technology based on machine learning has great advantages in processing DAS data, but currently machine learning is mainly used for identifying DAS signals in different events, and does not involve the interpretation and prediction of DAS signals in time series data. In addition, current domestic oil well water cut testing mainly relies on manual sampling and laboratory testing, and production fluid volume calculation methods mainly rely on indicator diagram analysis, which cannot be tested simultaneously, seriously affecting the measurement accuracy of production fluid volume. Therefore, how to use intelligent modeling technology to efficiently process and interpret the DAS data collected on site, and provide technical support for the simultaneous interpretation and prediction of future oil well flow rate and water cut, is a technical problem that has been focused on in this technical field. For this reason, we have invented a new flow rate and water cut prediction method based on distributed optical fiber acoustic wave monitoring data. SUMMARY

[0007] The application aims to provide a flow rate and water cut prediction method based on distributed optical fiber acoustic wave monitoring data, which realizes short-term synchronous prediction of future flow rate and water cut of oil wells by using intelligent modeling method, guides early deployment and regulation of oil well production, and helps intelligent injection and production control of oilfields.

[0008] The application can achieve the purpose by the following technical measures: the flow rate and water cut prediction method based on distributed optical fiber acoustic wave monitoring data comprises the following steps:

[0009] Step 1, extract DAS monitoring data of different layers of production wells, and perform pretreatment;

[0010] Step 2, calculate the DAS acoustic wave frequency band energy FBE of each layer;

[0011] Step 3, perform LSTM high-precision intelligent modeling to realize liquid production prediction;

[0012] Step 4, perform sensitivity analysis of the LSTM model to determine the main liquid production section;

[0013] Step 5, use DTS temperature data to verify the result of the sensitivity analysis;

[0014] Step 6, take the liquid production as the main control factor affecting the water cut, re-model, and realize short-term synchronous prediction of the flow rate and the water cut;

[0015] Step 7, perform field application, and evaluate the intelligent modeling effect.

[0016] The application can also achieve the purpose by the following technical measures:

[0017] In step 1, extract the measured DAS monitoring data of different production layers under stable production of the field production well, and perform outlier data cleaning and normalization processing on the data.

[0018] In step 1, the field measured DAS monitoring data reflects the phase change of the optical fiber laser backscattering Rayleigh scattering signal at any measuring point position with time, and the phase change reflects the acoustic wave intensity change of the DAS monitoring.

[0019] In step 1, the data outlier cleaning technique includes outlier identification + elimination + interpolation method, the outlier identification can adopt the box plot method, the DAS data outside the upper limit and lower limit range of the box plot is determined as an outlier, after eliminating the outliers, the adjacent average value interpolation is adopted for supplement, the purpose is to eliminate the influence of outliers to the greatest extent and improve the calculation accuracy of data; the normalization processing is to eliminate the influence of the order of magnitude and dimension of the DAS signal collected under different working conditions, so that the data has comparability and improves the calculation accuracy of the model; the selection of the normalization method can be optimized according to the effect of DAS original data outlier cleaning, mainly including maximum-minimum normalization which is very sensitive to outliers, and energy normalization and Z-Score normalization which have lower sensitivity to outliers.

[0020] In step 2, after the data preprocessing in step 1, the DAS sound wave frequency band energy FBE between specific frequency bands of each production layer section is calculated.

[0021] In step 2, FBE is the abbreviation of DAS sound wave frequency band energy, which is calculated by one-dimensional Fourier transform combined with windowed frequency shift, after a large amount of indoor experimental data processing, it is found that FBE has a certain relationship with the size of wellbore flow, which is specifically that the greater the flow in the wellbore, the stronger the interaction between the fluid and the pipe wall, the stronger the sound wave vibration, and the larger the FBE value.

[0022] In step 2, the selection of specific frequency band range also has a great influence on the calculation result of FBE, the calculation formula of FBE is shown in formula (1), [t-Δt, t+Δt] is the time interval, F(f, I) is the calculated one-dimensional Fourier amplitude value, and [f1, f2] is the selected frequency band range:

[0023]

[0024] The flow response characteristics of wellbore fluid are different in different frequency bands, the stronger the response of DAS signal in a certain frequency band, the more intense the fluid sound wave vibration in this frequency band, which also indicates that the DAS signal in this frequency band can better reflect the fluid flow characteristics; by using the dichotomy and prefix sum algorithm, a certain continuous frequency band with the maximum one-dimensional Fourier amplitude value is obtained by iterative solution, the intelligent screening of frequency band is realized, the calculation accuracy of FBE is improved, and FBE can better reflect the size of the production layer section flow.

[0025] In step 3, high-precision intelligent modeling is carried out, the FBE data of each production layer section calculated in step 2 is taken as the model input, a time sliding window is created, a LSTM long short-term memory neural network model of multiple input and single output type is established, and the liquid production rate prediction is realized.

[0026] Step 3 specifically includes:

[0027] Step 31, FBE and production data acquisition and preprocessing are performed;

[0028] Step 32, training set and test set division is performed;

[0029] Step 33, a time sliding window is created;

[0030] Step 34, LSTM high-precision intelligent model establishment and training are performed;

[0031] Step 35, LSTM model test verification is performed;

[0032] Step 36, LSTM model future short-term production prediction is performed.

[0033] In step 31, the FBE data of each production interval of the multi-layer oil well on site is calculated for a period of time [t1, t2] using the method proposed in step 2, the time interval is Δt, the selected frequency range is [f1, f2], and the selected window function is Hamming window or Hanning window when performing window frequency shift calculation; there are n production intervals, so the calculated FBE data are FBE1, FBE2, …, FBE n , and the FBE data are arranged in time sequence;

[0034] The wellhead measured production data are collected for a period of time [t1, t2], and the time interval is Δt, so the collected production data are Q1, Q2, …, Q n , and the production data are arranged in time sequence;

[0035] The calculated FBE data and the wellhead production data collected on site are standardized and pretreated, so that the data are standardized to have a mean value of 0 and a standard deviation of 1;

[0036] The standardized data are used to construct time series sample sets, including FBE feature vector sample sets and production sample sets, X t is the FBE feature vector at time t, Y t is the production at time t, each feature vector contains n features, numbered FBE1-FBE n .

[0037] In step 32, the FBE feature vector and production sample set obtained after step 31 are divided into training set and test set according to a certain proportion; the training set is a certain proportion of n production intervals of FBE data and wellhead production arranged in time sequence, and the test set is the remaining FBE data and wellhead production arranged in time sequence in the sample set.

[0038] In step 33, the time sliding window refers to predicting the data of the next time by using the training data of the previous time and the current time; specifically, the time step is set as Δt', the FBE feature vector X of the previous Δt' time is input t to predict the fluid production of the Δt'+1 time, then the time step of Δt' is moved step by step to predict the fluid production of the next Δt'+1 time, and the whole training set is iterated.

[0039] In step 34, the LSTM model structure includes input gate, output gate and forget gate, which selectively retains the previous information of the LSTM model and transmits the information throughout the LSTM network; at time t, the LSTM network processes and analyzes the input X t , the long-term hidden C t-1 and the short-term hidden H t-1 , and then generates the output Y t ; C t-1 contains the time step information before time t, H t-1 contains the previous time step information, FC is a full connection layer, X t and H t-1 are processed by FC.

[0040] The training set used by the LSTM model is composed of input time series and output time series, the input time series is the FBE feature vector X t , X t contains n features FBE1-FBE n , and the output time series is the wellhead fluid production Y t ; the input time series is input into the constructed LSTM model, and after calculation by the input gate, the forget gate and the output gate, the first layer LSTM output is obtained, the first layer LSTM output vector is used as the input vector of the second layer LSTM, and the calculation is iterated in this way, the output of each layer LSTM is the input of the next layer.

[0041] The LSTM training model can be set to have multiple hidden layers, to evaluate the training effect of the LSTM model, the mean absolute error is used as the effect evaluation criterion, and during the training of the LSTM model, the training parameters are constantly adjusted to reduce the MAE between the training prediction result and the true result, and improve the model training accuracy.

[0042] After training in steps 35 and 34, the FBE feature vectors in the test set are used as input to the LSTM model, and the wellhead production volume is used as the model output. The model prediction effect is verified by MAE. After the test, the predicted production volume is output and inversely normalized. The predicted production volume is compared with the measured production volume at the wellhead. When the predicted production volume reaches the set accuracy, the test ends. If the set accuracy is not reached, the process returns to step 34 to readjust the training parameters and train the LSTM model until the accuracy requirement is met.

[0043] In step 36, the measured DAS monitoring data of different production layers under stable production conditions of the production well are extracted, and the data are processed in steps 1, 2 and 31 to form an FBE feature vector sample set. The FBE feature vector sample set is input into the LSTM production volume prediction model trained, optimized and tested in steps 33, 34 and 35, and the production volume predicted in the short term is output.

[0044] In step 4, sensitivity analysis is performed on each segment to determine the main producing segment; the input for LSTM intelligent modeling is the FBE feature vector X. t X t It contains n features FBE1-FBE n , respectively represent the FBE calculated value corresponding to each segment, and the output is the liquid production volume; sensitivity analysis refers to the fact that when the FBE feature vector of a certain production segment is removed from the model input, the prediction accuracy of the entire LSTM network model decreases and the error MAE increases, indicating that this production segment has a significant impact on the model prediction, and also proving that this segment is the main liquid production segment.

[0045] In step 5, the DTS temperature data monitored in each production section is used to assist in verifying the accuracy of the sensitivity analysis results. DTS technology can monitor the temperature distribution changes along the wellbore. The greater the difference between the average temperature of a certain production section and the formation temperature of that section, the higher the heat carried by the fluid in that section. For ordinary oil wells, this also indicates that the flow rate in that section is greater. The DTS temperature difference data collected in each production section is used to assist in verifying the accuracy of the sensitivity analysis results in step 4. After the sensitivity analysis, if a certain production section is the main fluid-producing section, it means that after removing the FBE feature vector of this section from the LSTM model input, the MAE error of the model increases, and the DTS temperature difference is also larger.

[0046] In step 6, the production rate is taken as the main controlling factor affecting water cut, and a new LSTM neural network model is established to achieve short-term synchronous prediction of the flow rate and water cut. The measured water cut data at the wellhead are collected for a period of time [t1, t2], with a time interval of Δt. The collected water cut data are w1, w2, ..., w n; at this time, the data sample set comprises FBE data of each production layer of a multi-layer oil well in the field for a period of time, measured liquid production at the wellhead and measured water cut at the wellhead; a feature vector X input to the LSTM model t comprises n+1 features FBE1-FBE n and liquid production Q, and the result output by the LSTM model is water cut; the data set is input into the LSTM model constructed in step 3 for re-modeling, so as to realize prediction of water cut.

[0047] The object of the present application can also be achieved by the following technical measures: a flow rate and water cut prediction system based on distributed optical fiber acoustic wave monitoring data, characterized in that the flow rate and water cut prediction system based on distributed optical fiber acoustic wave monitoring data realizes short-term synchronous prediction of future flow rate and water cut of a production well in the field by using the flow rate and water cut prediction method based on distributed optical fiber acoustic wave monitoring data in any one of claims 1-18, and determines the main liquid production layer of a multi-layer oil well by using a sensitivity analysis method.

[0048] The flow rate and water cut prediction method based on distributed optical fiber acoustic wave monitoring data in the present application is an intelligent modeling method based on optical fiber DAS signals, realizes prediction of future flow rate of a production well in the field, determines the main liquid production layer of a multi-layer (segment) oil well by using a sensitivity analysis method, and has important significance for identifying potential layers, ensuring efficient production of oil wells and efficient development of oil and gas reservoirs. At the same time, the intelligent modeling method is used to realize short-term synchronous prediction of future flow rate and water cut of an oil well, guide early deployment and regulation of oil well production, and help intelligent management and control of oilfields. Compared with the prior art, the present application has the following beneficial effects:

[0049] 1. The present application provides a technology for realizing synchronous prediction of future liquid production and water cut of an oil well based on DAS monitoring data and combining LSTM intelligent modeling, which can simultaneously explain the evaluation of liquid production and water cut of an oil well on the development status of an oilfield and the adjustment of field measures.

[0050] 2. The DTS temperature difference and the mean absolute error MAE in the sensitivity analysis of the LSTM model are positively correlated, and the greater the DTS temperature difference of a production layer indicates that the layer is the main liquid production layer, which is an important discovery in engineering application.

[0051] 3. The LSTM intelligent modeling reveals the relationship between FBE data extracted from optical fiber DAS signals and liquid production and water cut through deep data mining, and the main liquid production layer of a multi-layer (segment) oil well can be determined by using the sensitivity analysis method proposed in the present application, which has important significance for ensuring fine oil production and efficient development of oil and gas reservoirs. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1A flow chart of a specific embodiment of the flow rate and water cut prediction method based on distributed optical fiber acoustic wave monitoring data of the present application;

[0053] Figure 2 A flow chart of the high-precision intelligent modeling for realizing liquid production rate prediction of the present application;

[0054] Figure 3 A flow chart of the LSTM network model training of the present application;

[0055] Figure 4 A schematic diagram of the training set liquid production rate prediction result of the LSTM model in the embodiment of the present application;

[0056] Figure 5 A schematic diagram of the test set liquid production rate prediction result of the LSTM model in the embodiment of the present application;

[0057] Figure 6 A relationship diagram of the average absolute error MAE of the DTS temperature difference and sensitivity analysis in the embodiment of the present application;

[0058] Figure 7 A schematic diagram of the future liquid production rate and water cut synchronous prediction result realized by using the LSTM to re-model in the embodiment of the present application. DETAILED DESCRIPTION

[0059] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0060] It should be noted that the terms used herein are only for the purpose of describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and furthermore, it should be understood that when the terms "comprise" and / or "include" are used in the specification, there is a presence of the features, steps, operations and / or combinations thereof.

[0061] As Figure 1 shown, Figure 1 A flow chart of the flow rate and water cut prediction method based on distributed optical fiber acoustic wave monitoring data of the present application. The flow rate and water cut prediction method based on distributed optical fiber acoustic wave monitoring data comprises:

[0062] Step 1, extract the measured DAS monitoring data of different production intervals under stable production of the field production well, and perform abnormal value data cleaning and normalization processing on the data;

[0063] Step 2, after data preprocessing in step 1, calculate the DAS sound wave frequency band energy FBE (Frequency Band Energy) between each production layer segment specific frequency band (f1-f2);

[0064] Step 3, high-precision intelligent modeling, taking the FBE data of each production layer segment calculated in step 2 as model input, creating a time sliding window, establishing a multi-input single-output type LSTM long short-term memory neural network model, and realizing liquid production rate prediction;

[0065] Step 4, sensitivity analysis is performed on each layer segment to determine the main liquid production segment;

[0066] Step 5, use the DTS temperature data monitored by each production layer segment to assist in verifying whether the sensitivity analysis result is correct;

[0067] Step 6, taking the liquid production rate as the main control factor affecting the water cut, re-establishing the LSTM neural network model to realize the future short-term synchronous prediction of the flow rate and the water cut.

[0068] Step 7, field application, evaluate the effect of intelligent modeling.

[0069] The present application provides a kind of based on DAS monitoring data, after intelligent frequency selection and sound wave frequency band energy FBE calculation, combined with the technology of realizing the future synchronous prediction of the liquid production rate and the water cut of multi-layer (segment) oil well by LSTM (long short-term memory neural network) intelligent modeling. Sensitivity analysis is performed on each layer segment by LSTM model to determine the main liquid production segment, and DTS temperature monitoring data is used to assist in verifying the sensitivity analysis result. The DTS temperature difference and the mean absolute error MAE in the sensitivity analysis of the LSTM model are positively correlated. The larger the DTS temperature difference of a certain production layer segment, the more important the layer segment is as the main liquid production segment. The LSTM intelligent modeling reveals the relationship between the FBE data extracted from the optical fiber DAS signal and the liquid production rate and the water cut through data deep mining. The main liquid production segment of multi-layer (segment) oil well can be determined by the sensitivity analysis method proposed in the present application, which is of great significance for ensuring fine oil production and efficient development of oil and gas reservoirs.

[0070] The following are several specific embodiments of the application

[0071] Example 1

[0072] In a specific embodiment 1 of the present application, the flow rate and water cut prediction method based on distributed optical fiber acoustic wave monitoring data includes the following steps:

[0073] In step 1, extract the measured DAS monitoring data of different production layers under stable production of the field production well, and perform outlier data cleaning and normalization processing on the data.

[0074] The field measurement DAS monitoring data reflects the phase change of the fiber laser backscattering Rayleigh signal of any measurement point position over time, and the phase change reflects the intensity change of the DAS monitored acoustic wave. The field fiber test is usually 0.5m or 1m per measurement point.

[0075] The data outlier cleaning technology includes outlier identification, rejection and interpolation method. The outlier identification can adopt the box plot method, and the DAS data outside the upper and lower limits of the box plot is determined as an outlier. After the outlier is rejected, the adjacent average value is interpolated to supplement, so as to eliminate the influence of the outlier to the greatest extent and improve the calculation accuracy of the data.

[0076] The normalization processing is to eliminate the influence of the DAS signal magnitude and dimension collected under different working conditions, so that the data has comparability and the calculation accuracy of the model is improved. The selection of the normalization method can be optimized according to the effect of DAS original data outlier cleaning, mainly including maximum-minimum normalization which is very sensitive to outliers, and energy normalization and Z-Score normalization which are less sensitive to outliers.

[0077] In step 2, after the data preprocessing of step 1, the DAS acoustic wave frequency band energy FBE (Frequency Band Energy) between the specific frequency band (f1-f2) of each production layer section is calculated;

[0078] The FBE is the abbreviation of DAS acoustic wave frequency band energy, which is calculated by one-dimensional Fourier transform combined with windowed frequency shift. After a large number of indoor experimental data processing, it is found that FBE has a certain relationship with the size of wellbore flow, which is specifically that the greater the flow in the wellbore, the stronger the interaction between the fluid and the pipe wall, the stronger the acoustic vibration generated, and the larger the FBE value.

[0079] The selection of the specific frequency band range also has a great influence on the FBE calculation result. The FBE calculation formula is shown in formula (1), [t-Δt, t+Δt] is the time interval, F(f, I) is the calculated one-dimensional Fourier amplitude value, and [f1, f2] is the selected frequency band range.

[0080]

[0081] The flow response characteristics of wellbore fluid are different at different frequency bands. The stronger the response of DAS signal in a frequency band, the more intense the fluid vibration at this frequency band, and the more the DAS signal in this frequency band can reflect the fluid flow characteristics. Therefore, how to select a suitable frequency band range as the basis for FBE calculation is an aspect concerned by the present application. The present application uses binary search and prefix sum algorithm to iteratively solve a certain continuous frequency band with the largest one-dimensional Fourier amplitude value, realizes intelligent screening of the frequency band, and improves the calculation accuracy of FBE, so that FBE can better reflect the flow size of the production interval.

[0082] In step 3, high-precision intelligent modeling, taking the FBE data of each production interval calculated in step 2 as the model input, creating a time sliding window, establishing a multi-input single-output type LSTM long short-term memory neural network model, and realizing production fluid rate prediction;

[0083] The LSTM is the abbreviation of long short-term memory neural network, which has great advantages in solving the interpretation and prediction of time series data. Therefore, it is feasible to process DAS data by using LSTM algorithm. Specifically, the process of realizing production fluid rate prediction by using LSTM high-precision intelligent modeling is as shown in Figure 2 .

[0084] 3-1 FBE and production fluid rate data acquisition and preprocessing

[0085] The FBE data of each production interval of the field multi-layer (segment) oil well is calculated by the method proposed in step 2 for a period of time [t1, t2], the time interval is Δt, the selected frequency band range is [f1, f2], and the selected window function is Hamming window or Hanning window when window shifting calculation is performed. Assuming that there are n production intervals, the calculated FBE data are FBE1, FBE2, …, FBE n , and the FBE data are arranged in time sequence.

[0086] The wellhead measured production fluid rate data are collected for a period of time [t1, t2], and the time interval is Δt. Then the collected production fluid rate data are Q1, Q2, …, Q n , and the production fluid rate data are arranged in time sequence.

[0087] The period of time can be any number of days, and is at least not less than 10 days. The time interval is at least not less than 6 hours.

[0088] The calculated FBE data and the collected wellhead production fluid rate data are standardized and pretreated, so that the data are standardized to data with a mean value of 0 and a standard deviation of 1.

[0089] The standardized data is used to construct a time series sample set, including an FBE feature vector sample set and a fluid production rate sample set, assuming X t is the FBE feature vector at time t, Y t is the fluid production rate at time t, each feature vector contains n features, numbered FBE1-FBE n .

[0090] 3-2 Training set and test set division

[0091] The FBE feature vector and fluid production rate sample set obtained after step 3-1 are divided into a training set and a test set according to a certain proportion. The training set is a certain proportion of n production layers FBE data and wellhead fluid production rate arranged in time sequence, and the test set is the remaining FBE data and wellhead fluid production rate arranged in time sequence in the sample set.

[0092] The certain proportion means that the training set accounts for 70%-90% of the entire data set.

[0093] 3-3 Create a time sliding window

[0094] The time sliding window refers to using the training data before the time and the current time to predict the data at the next time.

[0095] Specifically, the time step is set to Δt', the FBE feature vector X t in the previous Δt' time is input to predict the fluid production rate at time Δt'+1, then the time step Δt' is moved step by step to predict the fluid production rate at the next time Δt'+1, and the entire training set is iterated.

[0096] 3-4 LSTM high-precision intelligent model establishment and training

[0097] The LSTM model structure includes input gate, output gate and forget gate, and the role of these gates is to selectively retain previous information in the LSTM model and pass this information through the entire LSTM network. Specifically, the LSTM network model training process is as shown in Figure 3 , the LSTM network processes and analyzes the input X t , the long-term hidden C t-1 and the short-term hidden H t-1 at time t, and then generates the output Y t . C t-1 contains the time step information before time t, H t-1 contains the previous time step information, FC is a fully connected layer, X t and H t-1 are processed by FC.

[0098] The training set used by the LSTM model is composed of an input time series and an output time series, the input time series is an FBE feature vector X t , X t contains n features FBE1-FBE n , and the output time series is wellhead fluid production Y t . The input time series is input into the constructed LSTM model, and after calculation by the input gate, the forgetting gate and the output gate, the first layer LSTM output is obtained, the first layer LSTM output vector is used as the input vector of the second layer LSTM, and the iteration calculation is repeated, and the output of each layer LSTM is the input of the next layer.

[0099] The preferred LSTM training model can set multiple hidden layers, the number of neuron nodes of the hidden layer ranges from 100 to 1000, the number of training cycles ranges from 500 to 5000, and the batch size ranges from 10 to 50. In order to evaluate the training effect of the LSTM model, the mean absolute error (MAE, mean absolute error) is used as the effect evaluation criterion, and during the training of the LSTM model, the training parameters are continuously adjusted to reduce the MAE between the training prediction result and the true result, and improve the model training accuracy.

[0100] 3-5 LSTM model test and verification

[0101] After step 3-4, the FBE feature vector in the test set is used as the input of the LSTM model, and the wellhead fluid production is used as the output of the model, and the effect of the model prediction is verified by MAE.

[0102] After the test is completed, the predicted fluid production is output, and the inverse normalization processing is performed, the predicted fluid production is compared with the measured wellhead fluid production, when the predicted fluid production reaches the set accuracy, the requirement is met and the test is ended, and when the set accuracy is not reached, the training parameters are adjusted again in step 3-4, the LSTM model is trained, and the accuracy requirement is met.

[0103] The accuracy requirement is greater than 85%.

[0104] 3-6 LSTM model fluid production future short-term prediction

[0105] Under stable production of the field production well, the measured DAS monitoring data of different production intervals is extracted, and the data is processed according to steps 1, 2 and 3-1 to form an FBE feature vector sample set.

[0106] The FBE feature vector sample set is input into the LSTM fluid production prediction model trained and optimized in steps 3-3, 3-4 and 3-5 and tested, and the future short-term predicted fluid production is output.

[0107] In step 4, sensitivity analysis is performed on each layer section to determine the main liquid production section;

[0108] The input of the LSTM intelligent modeling is the FBE feature vector X t , which contains n features FBE1-FBE n , representing the FBE calculation value of each layer section, and the output is the liquid production. t The sensitivity analysis refers to that when the FBE feature vector of a production layer section is removed from the model input, the prediction accuracy of the entire LSTM network model decreases and the error MAE increases, indicating that this production layer section has a greater impact on the model prediction, and proving that this layer section is the main liquid production section.

[0109] The sensitivity analysis is of great significance for identifying the main liquid production section and ensuring fine oil production and efficient development of oil and gas reservoirs.

[0110] In step 5, DTS temperature data collected from each production layer section is used to assist in verifying whether the sensitivity analysis result is correct;

[0111] DTS technology can monitor the temperature distribution along the wellbore. The greater the difference between the average temperature of a production layer section and the formation temperature of that layer section, the higher the heat carried by the fluid in the wellbore of that layer section, and for ordinary oil wells, the greater the flow of that layer section.

[0112] DTS temperature difference data collected from each production layer section is used to assist in verifying whether the sensitivity analysis result in step 4 is correct. After sensitivity analysis, if a production layer section is the main liquid production section, it means that the error MAE of the model increases and the DTS temperature difference is larger when the FBE feature vector of this layer section is removed from the LSTM model input.

[0113] In step 6, liquid production is taken as the main control factor affecting water cut, and an LSTM neural network model is re-established to realize the future short-term synchronous prediction of flow and water cut.

[0114] The wellhead measured water cut data is collected for a period of time [t1, t2] with a time interval of Δt, and the collected water cut data is w1, w2, …, w n At this time, the data sample set includes the FBE data of each production layer section of the multi-layer (section) oil well, the wellhead measured liquid production, and the wellhead measured water cut.

[0115] The input feature vector X t of the LSTM model contains n+1 features FBE1-FBE n and liquid production Q, and the output of the LSTM model is the water cut. The data set is input into the LSTM model established in step 3 to re-model and realize the prediction of water cut.

[0116] Step 3 predicts the fluid production, and step 6 predicts the water cut mainly controlled by the FBE data and the fluid production, so as to achieve the purpose of synchronous prediction of the fluid production and the water cut at the same time in the future.

[0117] Step 7, field application, evaluation of the effect of intelligent modeling.

[0118] Example 2

[0119] The example well is a multi-layer production straight and inclined well in a certain reservoir, which contains 7 production intervals. The distributed optical fiber DAS+DTS logging operation is successfully carried out in the oil well, and the data recording time lasts for about 35 days. The first 5 days are shut-in and opening test data, and from the 6th day, the DAS+DTS monitoring data in the stable production stage are recorded.

[0120] Using the DAS and DTS data monitored in real time in the oil field, the LSTM intelligent modeling technology proposed in the application is used to realize the synchronous prediction of the fluid production and the water cut, and the steps are as follows.

[0121] Step 101, extract the measured DAS monitoring data of the 7 production intervals of the oil well in the stable production of the field production well. Since the wellhead sampling is once every 6 hours, 120 groups of DAS monitoring data consistent with the sampling time are extracted, and the data are subjected to abnormal value data cleaning and Z-Score normalization processing. Through intelligent frequency selection, the frequency band range is determined to be 600-1000Hz, and the DAS sound wave frequency band energy FBE of each production interval in a specific frequency band at different times is calculated.

[0122] Step 102, collect the wellhead measured fluid production data for 30 days, with a time interval of 6 hours. The collected fluid production data is 120 groups. The FBE data calculated in step 101 and the wellhead fluid production data collected in the field are subjected to standardization preprocessing, so that the data are standardized to have a mean value of 0 and a standard deviation of 1. The time series sample set is constructed by using the standardized data, including the FBE feature vector sample set and the fluid production sample set. It is assumed that X t is the FBE feature vector at time t, Y t is the fluid production at time t, and each feature vector contains 7 features, numbered FBE1-FBE7.

[0123] Step 103, divide the first 100 groups of data of the 120 groups of FBE feature vectors and fluid production data arranged in time sequence into a training set, and 10 groups of data into a test set, and the remaining 10 groups of data are used for short-term prediction of fluid production in the future.

[0124] Step 104, create a time sliding window, set the window step size Δt as 10. Build the LSTM model, the parameters used to train the LSTM model are: set two hidden layers in the LSTM network, the number of neuron nodes in the first hidden layer is 256, the number of neuron nodes in the second hidden layer is 512, the number of training cycles is 1000, the batch size is 10, use the mean absolute error MAE as the effect evaluation criterion, set the MAE as 5%, end the training within the error requirement range, and the result is as shown in Figure 4 .

[0125] Step 105, substitute the 10 groups of data in the test set into the trained and optimized LSTM model for verification, meet the accuracy requirement, after the test is completed, output the predicted fluid production, and do the inverse normalization processing, and the test result is as shown in Figure 5 .

[0126] Step 106, substitute the remaining 10 groups of data into the LSTM model trained and optimized and verified to predict the future short-term fluid production, and the future 60-hour fluid production is predicted.

[0127] Step 107, use sensitivity analysis to determine the main fluid production interval, and use the DTS temperature difference data collected from each production interval to assist in verifying the sensitivity analysis effect. Table 1 shows the prediction error results of the whole model after removing the FBE data of a certain production interval in the LSTM model. From the table, it can be seen that after removing the FBE data of the intervals 1332-1337m and 1391-1397m in the LSTM model, the model error MAE is larger, and the DTS temperature difference monitored in the field is also larger, which indicates that these two intervals are the main fluid production intervals. The DTS temperature difference is positively correlated with the error MAE in the sensitivity analysis of the LSTM model, and the result is as shown in Figure 6 .

[0128] Step 108, collect the wellhead measured water cut data for 30 days, the time interval is 6 hours, and the collected water cut data is 120 groups. At this time, the data sample set includes the FBE data of each production interval of the oil well, the wellhead measured fluid production and the wellhead measured water cut. The feature vector X t input into the LSTM model includes 8 features FBE1-FBE7 and fluid production Q, and the result output by the LSTM model is the water cut. Substitute the data set into the built LSTM model to re-model, realize the synchronous prediction of fluid production and water cut, and the synchronous prediction result of the future 60-hour fluid production and water cut is as shown in Figure 7 , and the black circles at the end of the graph are the predicted data.

[0129] Table 1 Sensitivity analysis to determine the main fluid production interval

[0130]

[0131] Example 3

[0132] An example well is a steam huff and puff horizontal well in a certain oil reservoir, which forms six horizontal production well sections after being sealed by multiple packers. The distributed optical fiber DAS+DTS logging operation is successfully carried out in the horizontal well, and the data recording duration lasts for about 40 days, of which the first 5 days are shut-in and open-hole test data, and from the 6th day, the DAS+DTS monitoring data in the stable production stage are recorded.

[0133] Using the DAS and DTS data monitored in real time on the oil well site, the LSTM intelligent modeling technology proposed in the application is used to realize the synchronous prediction of the liquid production and the water cut, and the steps are as follows.

[0134] Step 201, extract the measured DAS monitoring data of the six production well sections of the horizontal well in the stable production of the field production well. Since the wellhead sampling is once every 6 hours, 140 groups of DAS monitoring data consistent with the sampling time are extracted, and the data are subjected to abnormal value data cleaning and Z-Score normalization processing. Through intelligent frequency selection, the frequency band range is determined to be 500-800Hz, and the DAS sound wave frequency band energy FBE between different times in each production well section is calculated.

[0135] Step 202, collect the wellhead measured liquid production data for 35 days, with a time interval of 6 hours, and the collected liquid production data is 140 groups. The FBE data calculated in step 201 and the wellhead liquid production data collected on site are subjected to standardization preprocessing, so that the data are standardized to have a mean value of 0 and a standard deviation of 1. The data after standardization are used to construct time series sample sets, including FBE feature vector sample sets and liquid production sample sets. Assuming that X t is the FBE feature vector at time t, Y t is the liquid production at time t, and each feature vector contains 6 features, numbered FBE1-FBE6.

[0136] Step 203, divide the 140 groups of FBE feature vectors and the first 120 groups of liquid production data arranged in time sequence into a training set, and the 10 groups of data into a test set, and the remaining 10 groups of data are used for future short-term prediction of liquid production.

[0137] Step 204, create a time sliding window, set the window step size Δt as 5. Build the LSTM model, the parameters used to train the LSTM model are: set three hidden layers in the LSTM network, the number of neurons in the first hidden layer is 128, the number of neurons in the second hidden layer is 256, the number of neurons in the third hidden layer is 512, the number of training cycles is 2000, the batch size is 5, the mean absolute error MAE is used as the effect evaluation criterion, the MAE is set to 5%, within the error requirement range, end the training.

[0138] Step 205, substitute 10 groups of data in the test set into the trained and optimized LSTM model for verification, meet the accuracy requirement, after the test is completed, output the predicted fluid production, and do the inverse normalization processing.

[0139] Step 206, substitute the remaining 10 groups of data into the trained and optimized LSTM model for future short-term fluid production prediction, and the future 60-hour fluid production is predicted.

[0140] Step 207, use sensitivity analysis to determine the main fluid production section of the horizontal well, and use the DTS temperature difference data collected in each fluid production section to assist in verifying the sensitivity analysis effect. Table 2 shows the prediction error results of the whole model after removing the FBE data of a certain horizontal fluid production section in the LSTM model. From the table, it can be seen that after removing the FBE data of the horizontal well sections 1978-1998m and 2023-2043m close to the tail section in the LSTM model, the model error MAE is larger, and the DTS temperature difference monitored on site is also larger, which indicates that these two sections are the main fluid production sections. The DTS temperature difference and the error MAE in the sensitivity analysis of the LSTM model also show a positive correlation.

[0141] Step 208, collect 35 days of wellhead measured water cut data with a time interval of 6 hours, and the collected water cut data is 140 groups. At this time, the data sample set includes FBE data of each fluid production section of the horizontal well, wellhead measured fluid production and wellhead measured water cut. The feature vector X input by the LSTM model t contains 7 features FBE1-FBE6 and fluid production Q, and the result output by the LSTM model is the water cut. Input the data set into the built LSTM model for re-modeling, which can realize the short-term synchronous prediction of fluid production and water cut.

[0142] Table 2 Sensitivity analysis to determine the main fluid production section

[0143]

[0144] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will appreciate that modifications can be made to the technical solutions described in the foregoing embodiments, or some of the technical features thereof can be replaced equivalently, without departing from the spirit and principle of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

[0145] All that is not described in the specification is known to those skilled in the art.

Claims

1. A method for flow rate and water cut prediction based on distributed fiber optic acoustic monitoring data, characterized in that, The flow rate and water cut prediction method based on distributed optical fiber acoustic wave monitoring data comprises: Step 1, extract the DAS monitoring data of different layers of the production well, and pretreat; Step 2, calculate the DAS acoustic frequency band energy FBE of each layer; Step 3, perform LSTM high-precision intelligent modeling to realize liquid production prediction; Step 4, perform sensitivity analysis of the LSTM model to determine the main liquid production section; Step 5, use DTS temperature data to verify the results of the sensitivity analysis; Step 6, take the liquid production as the main control factor affecting the water cut, re-model, and realize short-term synchronous prediction of the flow rate and the water cut; Step 7, perform field application and evaluate the intelligent modeling effect.

2. The method of flow rate and water cut prediction based on distributed fiber optic acoustic monitoring data of claim 1, wherein, In step 1, extract the measured DAS monitoring data of different production layers of the field production well under stable production, and perform abnormal value data cleaning and normalization processing on the data.

3. The method of flow rate and water cut prediction based on distributed optical fiber acoustic monitoring data according to claim 2, characterized in that, In step 1, the field measured DAS monitoring data reflects the phase change of the fiber laser backscattering Rayleigh signal at any measurement point position with time, and the phase change reflects the acoustic intensity change of the DAS monitoring.

4. The method of flow rate and water cut prediction based on distributed fiber optic acoustic monitoring data of claim 2, wherein, In step 1, the data abnormal value cleaning technology includes abnormal value identification + elimination + interpolation method, the abnormal value identification can adopt the box plot method, the DAS data outside the upper and lower limits of the box plot is determined as abnormal, after eliminating the abnormal values, the adjacent average value interpolation is used for supplement, the purpose is to eliminate the influence of abnormal values to the greatest extent and improve the calculation accuracy of the data; The normalization processing is to eliminate the influence of the magnitude and dimension of the DAS signal collected under different working conditions, so that the data has comparability and improves the calculation accuracy of the model; the selection of the normalization method can be optimized according to the abnormal value cleaning effect of the DAS original data, mainly including maximum-minimum normalization which is very sensitive to abnormal values, and energy normalization and Z-Score normalization which are less sensitive to abnormal values.

5. The method of flow rate and water cut prediction based on distributed fiber optic acoustic monitoring data of claim 1, wherein, In step 2, after the data pretreatment of step 1, the DAS acoustic frequency band energy FBE between specific frequency bands of each production layer is calculated.

6. The method of flow rate and water cut prediction based on distributed optical fiber acoustic monitoring data according to claim 5, characterized in that, In step 2, FBE is the abbreviation of DAS acoustic frequency band energy, which is calculated by one-dimensional Fourier transform combined with windowed frequency shift, and after a large number of indoor experimental data processing, it is found that FBE has a certain relationship with the size of the wellbore flow rate, which is specifically that the larger the flow rate in the wellbore, the stronger the interaction between the fluid and the oil pipe wall, the stronger the acoustic vibration generated, and the larger the FBE value.

7. The method of flow rate and water cut prediction based on distributed fiber optic acoustic monitoring data of claim 5, wherein, In step 2, the selection of the specific frequency band range also has a great influence on the calculation result of FBE, the calculation formula of FBE is shown as formula (1), [t-Δt, t+Δt] is the time interval, F(f, I) is the calculated one-dimensional Fourier amplitude value, and [f1, f2] is the selected frequency band range: The flow response characteristics of wellbore fluid are different at different frequency bands. The stronger the response of DAS signal in a frequency band, the more intense the fluid vibration at this frequency band, and the more the DAS signal in this frequency band can reflect the fluid flow characteristics. The one-dimensional Fourier amplitude value of a certain continuous frequency band is obtained by iteration using the dichotomy and prefix sum algorithm, the intelligent screening of the frequency band is realized, and the calculation accuracy of FBE is improved, so that FBE can better reflect the flow rate of the production interval.

8. The method of flow rate and water cut prediction based on distributed optical fiber acoustic monitoring data according to claim 1, characterized in that, In step 3, high-precision intelligent modeling is performed, the FBE data of each production interval calculated in step 2 is taken as the model input, a time sliding window is created, a LSTM long short-term memory neural network model of the multi-input single-output type is established, and the liquid production prediction is realized.

9. The method of flow rate and water cut prediction based on distributed optical fiber acoustic monitoring data according to claim 8, characterized in that, Step 3 specifically includes: Step 31, FBE and liquid production data acquisition and preprocessing are performed; Step 32, the training set and the test set are divided; Step 33, a time sliding window is created; Step 34, a LSTM high-precision intelligent model is established and trained; Step 35, the LSTM model is tested and verified; Step 36, the LSTM model is used to predict the future short-term liquid production.

10. The method of flow rate and water cut prediction based on distributed optical fiber acoustic monitoring data according to claim 9, characterized in that, In step 31, the FBE data of each production interval of the multi-layer oil well on site is calculated for a period of time [t1, t2] by the method proposed in step 2, the time interval is Δt, the selected frequency range is [f1, f2], and the selected window function for windowed frequency shift calculation is Hamming window or Hanning window; there are n production intervals, and the calculated FBE data are FBE1, FBE2, …, FBE n , respectively; the FBE data are arranged in chronological order. In step 32, the FBE data of each production interval of the multi-layer oil well on site is calculated for a period of time [t1, t2] by the method proposed in step 2, the time interval is Δt, the selected frequency range is [f1, f2], and the selected window function for windowed frequency shift calculation is Hamming window or Hanning window; there are n production intervals, and the calculated FBE data are FBE1, FBE2, …, FBE n , respectively; the FBE data are arranged in chronological The wellhead actual measured liquid production data is collected for a period of time [t1, t2], and the time interval is Δt, so the collected liquid production data are Q1, Q2, …, Q n , and the liquid production data are arranged in time sequence; The calculated FBE data and the collected wellhead liquid production data are standardized and pretreated, so that the data are standardized to have a mean value of 0 and a standard deviation of 1; A time series sample set is constructed using the standardized data, including an FBE feature vector sample set and a fluid production rate sample set, X t is the FBE feature vector at time t, Y t is the fluid production rate at time t, and each feature vector contains n features, numbered FBE1-FBE n .

11. The method of flow rate and water cut prediction based on distributed optical fiber acoustic monitoring data according to claim 9, characterized in that, In step 32, the FBE feature vector and the liquid production sample set obtained after step 31 are divided into a training set and a test set according to a certain proportion; the training set is a certain proportion of FBE data and wellhead liquid production of n production intervals arranged in time sequence, and the test set is the remaining FBE data and wellhead liquid production in the sample set arranged in time sequence.

12. The method of flow rate and water cut prediction based on distributed optical fiber acoustic monitoring data according to claim 9, characterized in that, In step 33, the time sliding window refers to predicting the data of the next time by using the training data of the previous time and the current time; specifically, setting the time step as Δt', inputting the FBE feature vector X of the previous Δt' time t to predict the fluid production of the Δt'+1 time, then moving the time step of Δt' gradually to predict the fluid production of the next Δt'+1 time, and iterating the entire training set.

13. The method of flow rate and water cut prediction based on distributed optical fiber acoustic monitoring data according to claim 9, characterized in that, At step 34, the LSTM model structure includes input gates, output gates and forget gates, which have the effect of selectively retaining previous information by the LSTM model and carrying this information throughout the LSTM network, at time t the LSTM network processes the analysis input X t , long-term hidden C t-1 and short-term hidden H t-1 , then generates output Y t ; C t-1 contains the information of the time step before t, H t-1 contains the information of the previous time step, FC is a fully connected layer, X t and H t-1 are processed by FC; The training set used by the LSTM model is composed of an input time series and an output time series, the input time series is an FBE feature vector X t , X t contains n features FBE1-FBE n , and the output time series is wellhead fluid production Y t ; the input time series is input into the constructed LSTM model, and after calculation by the input gate, the forgetting gate and the output gate, the first layer LSTM output is obtained, the first layer LSTM output vector is used as the input vector of the second layer LSTM, and the like, and the calculation is iterated continuously, and the output of each layer LSTM is the input of the next layer. The LSTM training model can be set to have multiple hidden layers. To evaluate the training effect of the LSTM model, the mean absolute error is used as the effect evaluation criterion. During the training of the LSTM model, the training parameters are continuously adjusted to reduce the MAE between the training prediction result and the true result and improve the model training accuracy.

14. The method of flow rate and water cut prediction based on distributed optical fiber acoustic monitoring data according to claim 9, wherein, In step 35, after the training in step 34 is completed, the FBE feature vector in the test set is taken as the input of the LSTM model, and the wellhead liquid production is taken as the output of the model. The prediction effect of the model is verified by MAE. After the test is completed, the predicted liquid production is output and is subjected to inverse normalization processing. The predicted liquid production is compared with the measured wellhead liquid production. When the predicted liquid production reaches the set accuracy, the requirement is met and the test is ended. If the set accuracy is not reached, the training parameters are adjusted again in step 34, the LSTM model is trained, and the process is repeated until the accuracy requirement is met.

15. The method of flow rate and water cut prediction based on distributed optical fiber acoustic monitoring data according to claim 9, wherein, In step 36, the measured DAS monitoring data of different production intervals under stable production of the field production well are extracted, the data are processed according to steps 1, 2 and 31 to form an FBE feature vector sample set, the FBE feature vector sample set is input into the LSTM liquid production prediction model trained and optimized in steps 33, 34 and 35 and completed the test, and the future short-term predicted liquid production is output.

16. The method of flow rate and water cut prediction based on distributed optical fiber acoustic monitoring data according to claim 1, characterized in that, In step 4, sensitivity analysis is performed on each segment to determine the main producing segment; the input for LSTM intelligent modeling is the FBE feature vector X. t X t It contains n features FBE1-FBE n , respectively represent the FBE calculated value corresponding to each segment, and the output is the liquid production volume; sensitivity analysis refers to the fact that when the FBE feature vector of a certain production segment is removed from the model input, the prediction accuracy of the entire LSTM network model decreases and the error MAE increases, indicating that this production segment has a significant impact on the model prediction, and also proving that this segment is the main liquid production segment.

17. The method of flow rate and water cut prediction based on distributed optical fiber acoustic monitoring data according to claim 1, characterized in that, In step 5, the DTS temperature data of each production interval is used to verify whether the sensitivity analysis result is correct; the DTS technology can monitor the temperature distribution change along the wellbore, and the greater the difference between the average temperature of a production interval and the formation temperature of the interval, the higher the heat carried by the wellbore fluid of the interval, and for ordinary oil wells, the greater the flow of the interval; the DTS temperature difference data collected in each production interval is used to verify whether the sensitivity analysis result in step 4 is correct; after sensitivity analysis, if a production interval is a main liquid production interval, it means that after removing the FBE feature vector of the interval from the input of the LSTM model, the error MAE of the model becomes larger, and the DTS temperature difference is also larger.

18. The method of flow rate and water cut prediction based on distributed optical fiber acoustic monitoring data according to claim 1, characterized in that, In step 6, the liquid production rate is taken as the main influencing factor of water cut, and the LSTM neural network model is re-established to realize the future short-term synchronous prediction of the flow rate and water cut; the measured water cut data at the wellhead is collected for a period of time [t1, t2], and the time interval is Δt, so the collected water cut data are w1, w2, …, w n ; At this time, the data sample set includes the FBE data of each production interval of the multi-layer oil well, the measured liquid production rate at the wellhead, and the measured water cut at the wellhead for a period of time; the feature vector X t input by the LSTM model contains n+1 features FBE1-FBE n and the liquid production rate Q, and the output result of the LSTM model is the water cut; the data set is input into the LSTM model constructed in step 3 for re-modeling to realize the prediction of the water cut.

19. A flow and water cut prediction system based on distributed fiber optic acoustic monitoring data, characterized in that, The flow rate and water cut prediction system based on distributed optical fiber acoustic wave monitoring data realizes short-term synchronous prediction of future flow rate and water cut of a field production well by using the flow rate and water cut prediction method based on distributed optical fiber acoustic wave monitoring data in any one of claims 1-18, and determines the main liquid production interval of a multi-layer oil well by using a sensitivity analysis method.

Images