CN110630256B - A system and method for predicting wellhead water cut of low-production gas wells based on deep long-short-term memory network - Google Patents

Abstract

The invention relates to a system and a method for predicting the water content of a wellhead of a low gas production oil well based on a depth time memory network, which are characterized in that water content fluctuation information of wellhead mixed liquid is obtained through a high-frequency double-ring type capacitance sensor, a collected water content fluctuation time sequence is divided into a plurality of time slices with time sequence change through windowing treatment on the collected water content fluctuation time sequence, and a time-frequency characteristic, a nonlinear characteristic and a time-irreversible characteristic of each time slice are extracted to form a characteristic vector, so that a characteristic vector time sequence forming the water content of the wellhead is obtained; and then, taking the extracted water content characteristic vector time sequence as the input of a depth long-short time memory network, establishing a water content prediction model based on the depth long-short time memory network and the multivariate characteristics, and training by taking a wellhead liquid production content assay value as a water content label by adopting the model to finally obtain a predicted value of the water content. Because the characteristic time sequence of the water content fluctuation signal is the accurate description of the liquid production characteristics of the wellhead, the method can effectively eliminate the influence of a small amount of gas in the wellhead on the measurement, and further improve the measurement accuracy of the water content of the liquid produced by the wellhead.

Description

Prediction of wellhead water cut in low-production gas wells based on deep long-short-term memory network System and method

technical field

The invention belongs to the field of crude oil production, and relates to the measurement of water content of liquid produced in low-yield gas wells, in particular to a system and method for predicting the water content of low-yield oil wells based on a deep long-short-term memory network.

Background technique

In the process of crude oil production, timely mastering and controlling the water cut parameters of oil well production is not only the prerequisite for reliable estimation of crude oil net production, but also the basis for correct diagnosis and maintenance of oil well problems, and is also the basis for the adjustment of reservoir production mode. Therefore, it is of great significance to the detection of water cut parameters in oil well production. At present, the ultra-high water cut characteristic of oilfield production fluid has put forward new requirements for the water cut measurement of oil well production fluid, how to accurately obtain the water cut information of high water cut oil well production fluid has become an urgent problem to be solved. At present, the detection of water cut in oil well production fluid is usually realized by specially designed sensors, and its measurement methods include ultrasonic method, optical method, ray method, imaging method, conductometric method and electrical method, etc. However, the measurement effect of existing sensors cannot meet the requirements under the condition of high water content in oil well production, which is characterized by nonlinear sensor response and low water cut resolution, and the measurement results are greatly affected by salinity; in addition, oil field The traditional assay method in the operation is greatly affected by the sampling conditions and sampling frequency, and the measurement cycle is also long, making it difficult to achieve real-time measurement. Although shallow networks such as neural networks and support vector machines are widely used for soft sensing of water cut in oil-water two-phase flow, the characteristics of shallow network structures need to be carefully designed during the application process. Layer characteristics are highly subjective, and the prediction results of the model for water content will also be greatly affected by the designed characteristics.

Through the search of published patent documents, two published patent documents similar to the purpose and technical solution of this patent application were found:

1. A method (109447342A) for predicting water cut in an ultra-low permeability sandstone reservoir oil well at the initial stage of production, the method comprising: collecting and arranging the calculation parameters of the selected ultra-low permeability sandstone reservoir; utilizing the functional relationship between effective stress and water saturation Predict the initial water cut of wells in ultra-low permeability sandstone reservoirs. The water cut prediction method of the ultra-low permeability sandstone reservoir oil wells in the early stage of production provides a theoretical basis for explaining and revealing the water cut of oil wells in this type of reservoirs and predicting the water cut of oil wells in the early stage of production. Therefore, it has certain theoretical and practical significance.

2. A time-series-based multi-model prediction method for water content of oil well oil (105631554A), characterized in that it comprises the following steps: 1) Using historical data to establish a data set of water content of oil well oil as {xi, i=1 ,2,…,N}; 2), use wavelet analysis method to preprocess the data in the oil well oil water cut data set {xi,i=1,2,…,N}; 3), use the nearest neighbor propagation to gather class algorithm to classify {xi}Wave; 4), express the data in each cluster by the following time series form: 5), establish the time series model of each cluster according to the extreme learning machine algorithm and use the time series Sequence models get predicted values. It solves the problems of time-consuming and labor-intensive manual sampling of the water content of oil in existing oil wells, and affects the real-time performance of production monitoring and oil recovery data.

Through the comparison of technical features, in comparison document 1, the reservoir calculation parameters and methods used are also fundamentally different from those in the application of the present invention; and comparison document 2, although the time series method is used to predict water cut, its water cut The rate model and method are fundamentally different from the application of the present invention, so it will not have a substantial creative impact on the application of the present invention.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies of the prior art, to provide a low-production gas well head water cut prediction system and method based on deep long-short-term memory network, through which the time-series change information of water cut can be captured, and can Realize the accurate prediction of the water cut of oil well production fluid, especially high water cut oil well production fluid.

The purpose of the present invention is achieved like this:

A wellhead water cut prediction system for low-production gas wells based on a deep long-term and short-term memory network. The type capacitive sensor is used to obtain the water content information of the wellhead, and the high-frequency sinusoidal excitation signal source of the multivariate time-series feature extraction module of the water content generates the excitation signal, which is sent to the annular measuring electrode of the double-ring type capacitive sensor through the power divider for frequency sweeping , the water content data measured by the ring-shaped measuring electrode is excited and entered into the mixer for signal mixing, and the mixed signal is subjected to an adder and a voltage bias to obtain a time-series feature vector of the water content; The wellhead moisture content prediction network of the deep long short-term memory network splices the obtained multivariate time-series feature vectors of water content in chronological order as the input vector of the deep long short-term memory neural network. There is an LSTM unit inside the deep long short-term memory neural network. There are three functions: input gate, forget gate, and output gate. The deep long-short-term memory neural network unit has 6 layers. The Softmax classification function is used as the output function to output the predicted value.

Moreover, the double-ring capacitive sensor is composed of a stainless steel metal protective shell and an internal sensor pipe. The two ends of the stainless steel metal protective shell are provided with a left flange and a right flange, wherein the metal protective shell where the right flange and the left flange are located is threaded. Connection, the two ends of the metal protective shell are connected with the wellhead pipeline, and there are lead holes radially opened on the side wall of the stainless steel metal protective shell, and an internal sensor pipe made of woolen material is coaxially inlaid inside the stainless steel metal protective shell, and on the outer wall of the internal sensor pipe Two ring-shaped measuring electrodes are installed at intervals, and an electromagnetic shielding layer is installed outside the ring-shaped measuring electrodes, and the internal sensor pipe is compressed and sealed with the metal casing through the O-rings on both sides of the end face.

Moreover, the window function of the multivariate time-series feature extraction module for moisture content adopts non-overlapping windows with a window size of 1000 to divide the moisture signal multiple times, extract the multivariate feature sequences of moisture in different time periods, and convert the multivariate features of moisture into The fragments of the sequence are distributed by WVD to obtain the time-frequency domain matrix, and the recursive graph analysis method is used to process the signal to obtain the recursive graph matrix. The time-frequency energy and time-frequency entropy features are respectively extracted from the time-frequency domain matrix, and the recursive Rate, certainty, average diagonal length, hierarchy, and time irreversible features, the extracted feature vectors total the above seven feature parameters.

A method for predicting wellhead water cut of low-production gas wells based on deep long-short-term memory neural network, comprising the following steps:

⑴Installation and working parameter setting of double-loop high-frequency capacitive sensor:

Install the sensor in the wellhead descending pipeline, and perform a frequency sweep operation on the sensor to determine the optimal operating frequency of the sensor; when the optimal operating frequency of the sensor is determined, use a high-frequency sinusoidal excitation signal source to excite the ring-shaped measuring electrode to measure the microwave The amplitude attenuation and phase attenuation of the signal after passing through the sensor are used as the original measurement information of water cut; the double-ring high-frequency capacitive sensor adopts a continuous measurement method for the measurement of wellhead cut-up, and the sampling frequency is set at 10 times per minute, and the measurement data is typical The time series of the response holdup change of ;

⑵ Preprocessing of sensor acquisition signal

Carry out window segmentation on the signal. The window function segmentation signal sets the window size to 1000, and there is no overlapping window between the windows. The result of segmentation is the one-dimensional time series of the current time period. Multiple segmentations can extract the multivariate moisture content of different time periods. The feature sequence, the value in each serial port segmentation signal is taken out according to the time direction to obtain the multivariate feature sequence of water content; the feature extraction module performs time-frequency joint distribution and recursive graph analysis on the obtained water content fluctuation sequence segment, and obtains the time-frequency graph matrix and Recursive graph matrix, the corresponding eigenvector of each segment is obtained by calculation; the multivariate time series eigenvector contains 7 dimensions, namely time-frequency energy, time-frequency entropy, recursion rate, recursion certainty, recursion average diagonal length, recursion Hierarchy, time irreversibility; the feature extraction methods of these 7 dimensions are as follows:

First, time-frequency domain analysis is performed on the collected and processed signal, and Wigner-Ville distribution is performed on each windowed and segmented time series segment; first, the Hilbert transform is performed on the signal, and then the formula is used:

Where f is the frequency, t is the time, τ is the time delay, z(t) is the analytical form of the original signal, and the time-frequency diagrams obtained under different time segments are obtained. After that, the time-frequency energy and time-frequency are calculated for the time-frequency diagram matrix Entropy; where:

Time-frequency energy: Calculate the time-frequency distribution of the windowed time segment as P(t,f), then the time-frequency energy E can be passed

Calculated as follows:

Time-frequency entropy: Calculate the time-frequency distribution of the windowed time segment as P(t,f), and divide the time-frequency plane into N

For rectangles with equal block sizes, set the energy of each block as P _i , and the energy of the entire time-frequency plane as E, then the time-frequency entropy

Can be calculated by:

Then carry out recursive domain quantitative analysis on the collected and processed signals, and the recursive quantitative analysis indicators include recursion rate, certainty, average diagonal length, hierarchy and time irreversibility; among them:

Recursion rate: Calculate the recursion matrix RR of the windowed time segment, then the recursion rate is the percentage of the recursion points in the recursion graph plane to the total number of points that the plane can hold, which can be calculated by the following method:

It indicates the proportion of phase space points close to each other in the m-dimensional phase space to the total number of points;

Certainty: Calculate the recurrence matrix RR of the windowed time segment, then the certainty is the percentage of the recursive points constituting the line segment along the diagonal direction to all the recursive points, which can be calculated by the following method:

In the formula, is the number of line segments with length l, counting starts only when the length of the line segment in the diagonal direction is greater than the predetermined lower limit l _min , l _min is generally selected as an integer not less than 2, DET will recursively isolate the Recursive points are distinguished from organized recursive points that form continuous diagonal line segments. The more developed the line texture along the main diagonal in the recursive graph, the stronger the certainty of the system;

Average diagonal length: Calculate the recursive matrix RR of the windowed time segment. The certainty is the weighted average of the length of the line segment in the diagonal direction, which can be calculated by the following method:

The average diagonal length L represents the time length of two phase trajectories close to each other in the phase space trajectory, or expressed as the average period of the system, and the main diagonal is not included. The larger L is, the stronger the certainty of the system is;

Hierarchy: Calculate the recursion matrix RR of the windowed time segment. Hierarchy is the percentage of recursion points constituting the vertical line segment to all recursion points, which can be calculated by the following method:

Time irreversible quantity: first convert the original time series x(t) into incremental time series y(t), which represents

as follows:

y(i)=Δu(i)=x(i+1)-x(i), 1<i≤N

Then the time irreversible quantity can be calculated by the following way:

Among them, A represents the time irreversible quantity of the nonlinear dissipative system, y _i is the incremental time of the original time series

Between sequences, N is the length of the signal, H(*) is a sign function;

(3) Splicing of feature vectors and deep long short-term memory neural network prediction

① The eigenvectors of different signal segments are spliced according to the time direction to form a multivariate time-series eigenvector of water content;

②The multivariate time-series feature vector of water content is used as the training data of the deep short-term memory neural network, which is input into the network model for training; the deep long-term short-term memory neural network uses a total of 6 layers of LSTM units, and the hyperparameters of the deep long-term short-term memory neural network are set. The maximum number of iterations is 10,000 times to end the training, where the batch size is 100, the time step is 150, and the number of LSTM units is 128; there are three functions inside each LSTM unit, which are input gate function, forget gate function and output gate function. The input gate determines how much input value information at the current moment is added to the LSTM unit state, the forget gate determines how much information is discarded from the LSTM state, and the output gate determines what value needs to be output according to the current LSTM unit state; the formulas are as follows:

input _t = σ(W _i *[h _t-1 , x _t ]+b _i )

forget _t ＝σ(W _f *[h _t-1 ，x _t ]+b _f )

output _t = σ(W _O *[h _t-1 , x _t ]+b _o )

Among them, W _i , W _f and W _O represent the weight parameters corresponding to the input gate, forget gate and output gate respectively, b _i , b _f and b _o correspond to the bias items respectively, and h _t-1 is the LSTM unit at the previous moment Internal state, x _t is the input value at the current moment;

After the water content multivariate time-series feature vector t1 is input to the first layer of LSTM units, it must go through the calculation of the above three gate functions and determine the output of the LSTM; after calculating the feature sequence at the current moment, the LSTM unit moves to the next moment t2, Repeat the above process and calculate the output; after calculating the first layer of LSTM units, the output vector of the first layer is used as the input vector of the second layer of LSTM units, the process is the same as above; the output of each layer of LSTM units is the input of the next layer;

During the training process, the multi-dimensional feature time-series signals are input into the LSTM unit in the deep long short-term memory network in sequence according to time for training. During the training process, the classification value is predicted by the deep long short-term memory neural network and compared with the actual wellhead moisture test value;

③Use the Softmax function to judge, pass the judgment result back to the deep long short-term memory neural network and update the network parameters layer by layer; the Softmax function can compress a K-dimensional vector Z containing any real number to another K-dimensional real vector σ In (Z), the range of each element is between (0,1), and the sum of all elements is 1, and the Softmax form is:

Among them, j=1,...,K, i represents a category in K, and z _j represents the value of this category;

④The trained model can predict the moisture content

When predicting, after inputting the multi-dimensional feature time series signal into the deep long-short-term memory network, the output value of the Softmax function is the water content of the current signal.

Advantage and positive effect of the present invention are:

1. The double-loop capacitive sensor adopted in the system of the present invention can quickly and accurately obtain the fluctuation signal of the water content sequence; the windowing processing of the signal can extract the multiple characteristic sequences of the water content in different time periods; the time-frequency domain of the windowing signal Multivariate eigenvalues can be obtained after recursive domain analysis, which can highlight the multidimensional characteristics of the signal; deep long short-term memory (LSTM) neural network training for multidimensional feature sequences can accurately predict the water cut value of the wellhead.

2. The double-ring capacitive sensor used in the system of the present invention is installed in the wellhead descending pipeline, which can directly measure the water content of the liquid produced as soon as possible. Optimal management is of great significance. Compared with existing sensors, it has stronger stability, and the shielding layer can effectively shield microwave scattering and external electromagnetic wave interference, and lock the signal within a certain range. The sensor can effectively and accurately measure the gas-liquid flow inside the low gas production oil well pipeline.

3. The system of the present invention adopts a deep long-short-term memory (LSTM) neural network, which is very suitable for dealing with problems highly related to time series. Compared with traditional recognition methods, such as support vector machine (SVM), recursive neural network (RNN ), etc., can effectively avoid problems such as gradient disappearance and gradient explosion, and at the same time, the three internal gate functions can enhance the network learning ability, which can increase the prediction accuracy by about 5%-10% compared with the above network model.

4. The method of the present invention extracts the time series signal measured by the sensor to take values as features, and splices the features of each water content fluctuation sequence segment. The spliced feature is the feature vector of the signal segment, and the feature vector contains a wealth of information. Wellhead water cut information, input the time series features into the deep long short-term memory network, can capture the basic characteristics and rules of water cut change, and provide rich features for the establishment of water cut prediction model, compared with using the original signal to directly calculate the water cut Rate prediction, this feature extraction method can better obtain the feature information of the signal in different spaces, and can highlight and strengthen the feature characteristics of the signal.

5. The method of the present invention has completed the design of the signal acquisition sensor and used the deep long short-term memory (LSTM) neural network to predict the water content of the wellhead. The process is rigorous and feasible, and the predicted value of the water content is accurate. computing resources. Since the time-series characteristics of the moisture content contain abundant flow characteristics, the model proposed by the invention can achieve high prediction accuracy of the moisture content, and the prediction accuracy rate can reach more than 97%. Compared with the traditional water content prediction method, the method of the present invention has the advantages of fast speed, high accuracy, low calculation resource consumption, elimination of human subjective factors and the like.

Description of drawings

Fig. 1 is the structural diagram of the double-ring type high-frequency capacitive sensor that the present invention is used for the measurement of wellhead liquid production holdup;

Fig. 2 is the schematic diagram of electrical control of wellhead moisture content and windowing division of the present invention;

Figure 2-1 is an enlarged schematic diagram of the signal diagram in Figure 2;

Fig. 3 is a schematic diagram of feature extraction of wellhead moisture content of the present invention;

Fig. 4 is a flow chart of water content feature prediction in the present invention.

Detailed ways

The present invention is further described below in conjunction with embodiment: following embodiment is illustrative, not limiting, can not limit protection scope of the present invention with following embodiment.

A wellhead water cut prediction system for low-production gas wells based on a deep long-short-term memory network, which consists of a double-ring high-frequency capacitive sensor, a water-cut multivariate time-series feature extraction module, and a wellhead water-cut prediction network based on a long-short-term memory network.

The double-ring capacitive sensor is used to obtain wellhead water content information. Its structure is shown in Figure 1. It consists of a stainless steel metal protective shell and an internal sensor pipe 3. The two ends of the stainless steel metal protective shell are left flanges 1 with a nominal diameter of DN50. 1. The right flange 9, wherein the metal protective shell where the right flange and the left flange are located is connected by thread 8, so as to facilitate the installation of internal sensor pipes. The two ends of the metal protection shell are connected with the wellhead pipeline, and a lead hole 5 with an inner diameter of 18mm is radially opened on the side wall of the stainless steel metal protection shell, which is used for the connection line channel between the sensor electrode and the external measurement and calculation instrument; the inside of the stainless steel metal protection shell is coaxial An inner sensor pipe made of woolen material with an inner diameter of 50 mm is installed for the transmission of the oil-water mixture at the wellhead; two annular measuring electrodes 6 are installed at intervals on the outer wall of the inner sensor pipe for the water content measurement of the oil-water mixture. At the same time, an electromagnetic shielding layer 4 is installed outside the annular measuring electrode to improve the measurement effect of the sensor. The internal sensor pipe is pressed and sealed with the metal casing through the O-rings 2 on both sides of the end face to prevent the leakage of the wellhead production fluid.

In this embodiment, the distance between the flanges of the stainless steel metal protective shell is 330mm, the length of the woolen pipe in the sensor is 310mm, the diameter of the sensor pipe is 50mm, the wall thickness of the woolen pipe is 80mm, the inner diameter of the annular measuring electrode is 80mm, the outer diameter is 85mm, two The distance between the ring-shaped measuring electrodes is 50mm, the electromagnetic shielding layer is a metal copper plate with a thickness of 1mm, and it is rolled and welded into a cylindrical tube with a length of 90mm and an inner diameter of 90mm, supported by a plexiglass ring 7 between the woolen pipe and the pipe.

The multivariate time-series feature extraction module of water content is shown in Fig. 2 and Fig. 3 . In Figure 2, the high-frequency sinusoidal excitation signal source generates the excitation signal, which is sent to the ring-shaped measuring electrode of the sensor through the power divider for frequency sweeping, and the ring-shaped measuring electrode sends the water content data measured by the frequency sweep to the mixer after being excited. The signal is mixed, and the mixed signal is subjected to an adder and a voltage bias to obtain a multivariate characteristic sequence of water content. In this embodiment, the window function of the signal adopts a non-overlapping window with a window size of 1000, so that the water content signal can be segmented multiple times, and multiple feature sequences of water content in different time periods can be extracted through multiple segmentations. The time-frequency domain matrix is obtained by using the WVD distribution (joint time-frequency distribution) for the fragments of the multivariate feature sequence of water content, and the recursive graph analysis method is used to process the signal to obtain the recursive graph matrix, see Figure 3. The time-frequency energy and time-frequency entropy features are extracted for the time-frequency domain matrix, and the recurrence rate, certainty, average diagonal length, hierarchy, and time irreversible features are extracted for the recursive graph matrix. The extracted multivariate time series features of water content The vector totals the above seven feature parameters.

Extracting the time-frequency domain matrix and recursive graph matrix for quantitative analysis is highly advanced. It can analyze signals in different dimensions and extract features of different dimensions. Compared with directly using signals as data sources, it not only increases Its dimensions, but also highlight and enhance the characteristics of the signal.

The wellhead moisture prediction network based on the deep long-short-term memory network splices the multivariate time-series feature vectors of the obtained moisture content in chronological order as the input vector of the deep long-short-term memory (LSTM) neural network, and its structure is shown in Figure 4 shown. There are LSTM units inside the deep long short-term memory (LSTM) neural network, and there are three functions inside the unit: input gate, forget gate, and output gate. There are 6 layers of LSTM units. The Softmax classification function is used as the output function to output the predicted value. The predicted true value is the test value of water cut at the wellhead, which is used to reversely correct the internal parameters of the deep long-short-term memory (LSTM) network to achieve the purpose of prediction.

A method for predicting wellhead water cut in low-production gas wells based on deep long-short-term memory (LSTM) neural network, comprising the following steps:

⑴Installation and working parameter setting of double-loop high-frequency capacitive sensor

Install the sensor on the downpipe at the wellhead and connect it to the pipeline through a DN50 flange. A frequency sweep is then performed on the sensor to determine the optimum operating frequency for the sensor. Set the scanning frequency band of the sensor to 0.8Ghz-10GHz, which is the microwave band. When the optimum working frequency of the sensor is determined, the ring-shaped measuring electrode is excited at this frequency, and the amplitude attenuation and phase attenuation of the microwave signal after passing through the sensor are measured as the original measurement information of the water content. The double-ring high-frequency capacitive sensor adopts a continuous measurement method for the measurement of the wellhead holdup, and the sampling frequency is set at 10 times per minute. The measurement data is a typical time series of response to the change of holdup, and the sensor measurement time series value can be transmitted through wireless Upload to the server for storage and analysis operations.

⑵ Preprocessing of sensor acquisition signal

Carry out window segmentation on the signal, as shown on the right side of Figure 2, the window function segmentation signal sets the window size to 1000, and there is no overlapping window between the windows. The result of segmentation is the one-dimensional time series of the current time period. For the multivariate feature sequence of water content in the time period, the value in each serial port segmentation signal is taken out according to the time direction to obtain the multivariate feature sequence of water content. The feature extraction module performs time-frequency joint distribution and recursive graph analysis on the obtained water content fluctuation sequence segments, as shown in Figure 3, and obtains the time-frequency graph matrix and the recursive graph matrix, and calculates the corresponding feature vector of each segment through the corresponding formula; The multivariate time series feature vector contains 7 dimensions, which are time-frequency energy, time-frequency entropy, recursion rate, recursion certainty, recursion average diagonal length, recursion hierarchy, and time irreversibility. The feature extraction method of these 7 dimensions is as follows:

First, time-frequency domain analysis is performed on the collected and processed signals, and Wigner-Ville distribution (WVD) is performed on each time series segment after windowing and segmentation. First perform Hilbert transform on the signal, and then pass the formula:

(where f is the frequency, t is the time, τ is the time delay, and z(t) is the analytical form of the original signal)

The time-frequency diagrams under different time segments are obtained, and then the time-frequency energy and time-frequency entropy are calculated for the time-frequency diagram matrix.

in:

Time-frequency energy: Calculate the time-frequency distribution of the windowed time segment as P(t,f), then the time-frequency energy E can be passed

Calculated as follows:

2. Time-frequency entropy: Calculate the time-frequency distribution of the windowed time segment as P(t,f), and divide the time-frequency plane into

Assuming that the energy of each block is P _i and the energy of the entire time-frequency plane is E, then the time

Frequency entropy can be calculated by:

For the quantitative analysis of the joint time-frequency distribution characteristics of low-velocity high-water-cut vertical oil-water two-phase flow, the quadratic time-frequency distribution can reflect the fluid characteristics more reasonably and intuitively, and the time-frequency energy and time-frequency entropy can directly reflect the characteristics of the time-frequency diagram. is an important feature of the time-frequency distribution.

Subsequently, the recursive domain quantitative analysis is carried out on the collected and processed signals, and the recursive quantitative analysis indicators include recursive rate, certainty, average diagonal length, hierarchy and time irreversible quantity. in:

Recursion rate: Calculate the recursion matrix RR of the windowed time segment, then the recursion rate is the percentage of the recursion points in the recursion graph plane to the total number of points that the plane can hold, which can be calculated by the following method:

It indicates the proportion of phase space points close to each other in the m-dimensional phase space to the total number of points;

Certainty: Calculate the recurrence matrix RR of the windowed time segment, then the certainty is the percentage of the recursive points constituting the line segment along the diagonal direction to all the recursive points, which can be calculated by the following method:

where is the number of line segments with length l. Counting starts only when the length of the line segment in the diagonal direction is greater than the predetermined lower limit l _min . l _min is generally selected as an integer not less than 2. DET distinguishes isolated recurrence points in a recurrence graph from those organized to form continuous diagonally oriented segments. The more developed the line texture along the main diagonal in the recurrence graph, the stronger the certainty of the system;

Average diagonal length: Calculate the recursive matrix RR of the windowed time segment. The certainty is the weighted average of the length of the line segment in the diagonal direction, which can be calculated by the following method:

The average diagonal length L represents the time length of two phase trajectories close to each other in the phase space trajectory, or expressed as the average period of the system, and the main diagonal is not included. The larger L is, the stronger the certainty of the system is.

Hierarchy: Calculate the recursion matrix RR of the windowed time segment. Hierarchy is the percentage of recursion points constituting a vertical line segment to all recursion points, which can be calculated by the following method:

In the formula, P(v) is the number of line segments with length v. Counting starts only when the length of the line segment in the diagonal direction is greater than the predetermined lower limit v _min . v _min is generally selected as an integer not less than 2. LAM represents the probability of recursive points in the hierarchical state in the system. When there are more isolated recursive points in the recursive graph than the vertical line segment structure, LAM will decrease. The present invention sets v _min as 2.

Time irreversible quantity: first convert the original time series x(t) into incremental time series y(t), which represents

as follows:

y(i)=Δu(i)=x(i+1)-x(i), 1<i≤N

Then the time irreversible quantity can be calculated by the following method:

Among them, A represents the time irreversible quantity of the nonlinear dissipative system, y _i is the incremental time series of the original time series, N is the length of the signal, and H(*) is a sign function.

Recursive graph quantitative analysis is a useful supplement and exploration for revealing the flow regime transformation mechanism of two-phase flow which is complex, uncertain and difficult to describe accurately with mathematical models.

(3) Splicing of feature vectors and deep long-short-term memory (LSTM) neural network prediction

① The eigenvectors of different signal segments are spliced according to the time direction (t ₁ , t ₂ ... t _n ) to form a multivariate time-series eigenvector of water content, as shown on the left side of Figure 4. Because the feature vector is spliced in the time direction, it retains the time dimension feature, and secondly, the vector can intuitively reflect the sequence feature.

②The multivariate time-series feature vector of water content is then used as the training data of the deep long-short-term memory (LSTM) neural network and input into the network model for training. The deep long short-term memory (LSTM) neural network uses a total of 6 layers of LSTM units, and the hyperparameters of the deep long short-term memory (LSTM) neural network are set. The training ends after the maximum number of iterations is 10,000 times. The number of units is 128. There are three functions inside each LSTM unit, which are the input gate function, the forget gate function and the output gate function. The input gate determines how much input value information at the current moment is added to the LSTM unit state, and the forget gate determines how much information is added to the LSTM state from the LSTM state. How much information is discarded, and the output gate determines what value needs to be output according to the current LSTM cell state. The formulas are as follows:

input _t = σ(W _i *[h _t-1 , x _t ]+b _i )

forget _t ＝σ(W _f *[h _t-1 ，x _t ]+b _f )

output _t ＝σ(W _o *[h _t-1 ，x _t ]+b _o )

Among them, W _i , W _f and W _O represent the weight parameters corresponding to the input gate, forget gate and output gate respectively, b _i , b _f and b _o correspond to the bias items respectively, and h _t-1 is the LSTM unit at the previous moment Internal state, x _t is the input value at the current moment.

After the water content multivariate time-series feature vector t ₁ is input to the first layer of LSTM units, it must go through the calculation of the above three gate functions and determine the output of the LSTM. After calculating the feature sequence at the current moment, the LSTM unit moves to the next moment _t2 , repeats the above process and calculates the output. After calculating the LSTM unit of the first layer, the output vector of the first layer is used as the input vector of the LSTM unit of the second layer, and the process is the same as above. The output of each layer of LSTM unit is the input of the next layer.

During the training process, the multi-dimensional feature time series signals are sequentially input into the LSTM unit in the deep long short-term memory network according to time (t ₁ , t ₂ ... t _n ) for training, and the training process is predicted by the deep long short-term memory neural network Classified values, and compared with the actual wellhead moisture test values.

③The softmax function is used to judge, and the judgment result is reversely transmitted back to the deep long short-term memory neural network and the network parameters are updated layer by layer. The Softmax function can "compress" a K-dimensional vector Z containing any real number into another K-dimensional real vector σ(Z), so that the range of each element is between (0, 1), and the values of all elements The sum is 1, and the Softmax form is:

Among them, j=1,...,K, j represents a category in K, and z _j represents the value of this category.

④The trained model can predict the moisture content.

When predicting, after inputting the multi-dimensional feature time series signal into the deep long-short-term memory network, the output value of the Softmax function is the water content of the current signal.

Claims (2)

1. A low-production gas wellhead water cut prediction system based on a deep long-short-term memory network, characterized in that it consists of a double-ring high-frequency capacitive sensor, a water-cut multivariate time-series feature extraction module, and wellhead water-cut prediction based on a long-short-term memory network Network composition, the double-loop high-frequency capacitive sensor is used to obtain wellhead water content information, the high-frequency sinusoidal excitation signal source of the water content multivariate time-series feature extraction module generates an excitation signal, and sends it to the double-loop capacitive sensor through a power divider The ring-shaped measuring electrode is used for frequency sweeping, and the water content data measured by the ring-shaped measuring electrode is excited and sent to the mixer for signal mixing. After the mixed signal is passed by the adder and voltage bias, the water content is obtained Multivariate time-series feature of rate; the wellhead water content forecasting network based on long-short-term memory network splices the multivariate time-series feature vectors of water-cut obtained according to time order, as the input vector of deep long-short-term memory neural network, deep long-short-term memory neural network There is an LSTM unit inside, and there are three functions in the unit: input gate, forget gate, and output gate. The deep long-short-term memory neural network unit has 6 layers. The Softmax classification function is used as the output function to output the predicted value; The double-ring high-frequency capacitive sensor is composed of a stainless steel metal protective shell and an internal sensor pipe. The two ends of the stainless steel metal protective shell are provided with a left flange and a right flange, and the metal protective shell where the right flange and the left flange are located is threaded. Connection, the two ends of the metal protective shell are connected with the wellhead pipeline, and there are lead holes in the radial direction on the side wall of the stainless steel metal protective shell. Inside the stainless steel metal protective shell, an internal sensor pipe made of woolen material is coaxially inlaid, and an interval is placed on the outer wall of the internal sensor pipe. Two ring-shaped measuring electrodes are installed, and an electromagnetic shielding layer is installed on the outside of the ring-shaped measuring electrodes. The inner sensor pipe is pressed and sealed with the metal casing through the O-rings on both sides of the end face; The window function of the multivariate time-series feature extraction module of water content adopts a non-overlapping window with a window size of 1000 to divide the water content signal multiple times, extract the multivariate feature sequences of water content in different time periods, and divide the multivariate time-series feature sequences of water content The time-frequency domain matrix is obtained by WVD distribution of the segment, and the recursive graph analysis method is used to process the signal to obtain the recursive graph matrix. The time-frequency energy and time-frequency entropy features are extracted from the time-frequency domain matrix, and the recurrence rate is extracted from the recursive graph matrix. , certainty, average diagonal length, hierarchy, and time irreversible features, the extracted feature vectors total the above seven feature parameters. 2. a method for predicting the water content of the low-yield gas well head based on the deep long-short-term memory neural network by using the system as claimed in claim 1, is characterized in that: comprises the steps: ⑴Installation and working parameter setting of double-loop high-frequency capacitive sensor: Install the sensor in the wellhead descending pipeline, and perform a frequency sweep operation on the sensor to determine the optimal operating frequency of the sensor; when the optimal operating frequency of the sensor is determined, use a high-frequency sinusoidal excitation signal source to excite the ring-shaped measuring electrode to measure the microwave The amplitude attenuation and phase attenuation of the signal after passing through the sensor are used as the original measurement information of water cut; the double-ring high-frequency capacitive sensor adopts a continuous measurement method for the measurement of wellhead cut-up, and the sampling frequency is set at 10 times per minute, and the measurement data is typical The time series of the response holdup change of ; ⑵ Preprocessing of sensor acquisition signal The signal is segmented by adding a window, and the window function is used to segment the signal. The window size is set to 1000, and there is no overlapping window between the windows. The result of the segmentation is the one-dimensional time series of the current time period. Multiple segmentations extract the multiple time series of moisture content in different time periods. Feature sequence, the value in each serial port segmentation signal is taken out according to the time direction to obtain a multivariate time-series feature sequence of water content; the feature extraction module performs time-frequency joint distribution and recursive graph analysis on the obtained multivariate time-series feature sequence of water content to obtain time-frequency Graph matrix and recursive graph matrix, the corresponding eigenvector of each segment is obtained by calculation; the multivariate time-series eigenvector of water content contains 7 dimensions, which are time-frequency energy, time-frequency entropy, recursion rate, recursive certainty, and recursive average pair Dimensional line length, recursive hierarchy, irreversible time; the feature extraction methods of these 7 dimensions are as follows: First, time-frequency domain analysis is performed on the collected and processed signal, and Wigner-Ville distribution is performed on each windowed and segmented time series segment; first, the Hilbert transform is performed on the signal, and then the formula is used:

Where WVD(t,f) is the multivariate time-series feature of water content, f is frequency, t is time, τ is time delay, z(t) is the analytical form of the original signal, z ^* (t) is the total of z(t) In the form of yoke function, the time-frequency diagrams under different time segments are obtained, and then the time-frequency energy and time-frequency entropy are calculated for the time-frequency diagram matrix; where: Time-frequency energy: Calculate the time-frequency distribution of the windowed time segment as P(t,f), then the time-frequency energy E is calculated as follows:

Time-frequency entropy: Calculate the time-frequency distribution of the windowed time segment as P(t,f), divide the time-frequency plane into N blocks of equal size rectangles, set the energy of each block as P _i , and the energy of the entire time-frequency plane is E, then the time-frequency entropy is calculated by the following method:

Then carry out recursive domain quantitative analysis on the collected and processed signals, and the recursive domain quantitative analysis indicators include recursion rate, certainty, average diagonal length, hierarchy and time irreversibility; among them: Recursion rate: Calculate the recursion matrix RR of the windowed time segment, then the recursion rate is the percentage of the recursion points in the recursion graph plane to the total number of points that can be accommodated in the plane, calculated by the following method:

In the above formula, RR represents the recursive matrix, R _i,j is the element value of row i and column j in the recursive matrix, i is the index of row i of the recursive matrix, j is the index of column j of the recursive matrix, and N is the dimension of the recursive matrix ; The RR recurrence rate indicates the proportion of phase space points close to each other in the m-dimensional phase space to the total number of points; Certainty: To calculate the recurrence matrix RR of the windowed time segment, the certainty is the percentage of the recursion points constituting the line segment along the diagonal direction to all the recursion points, which is calculated by the following method:

In the formula, P(l) is the number of line segments with a length of l, and the counting starts only when the length of the line segment in the diagonal direction is greater than the predetermined lower limit l _min ; l _min is selected as an integer not less than 2; DET will recursively The isolated recursive points in the recursive graph are distinguished from the recursive points organized to form a continuous diagonal line segment; the more developed the line texture along the main diagonal in the recursive graph, the stronger the determinism of the system is; Average diagonal length: Calculate the recursive matrix RR of the windowed time segment. The average diagonal length is the weighted average of the length of the line segment in the diagonal direction, which is calculated by the following method:

The average diagonal length L represents the time length of two phase trajectories close to each other in the phase space trajectory, or expressed as the average period of the system, the main diagonal is not included, the larger the L, the more certain the system is stronger Hierarchy: Calculate the recurrence matrix RR of the windowed time segment. Hierarchy is the ratio of the recursion points that constitute the vertical line segment to all the recursion points, which is calculated by the following method:

In the formula, P(v) is the number of line segments with length v, counting starts only when the length of the line segment in the diagonal direction is greater than the predetermined lower limit v _min , v _min is selected as an integer not less than 2, and LAM represents the system The probability of recursive points in the middle layer state, when there are more isolated recursive points in the recursive graph than the vertical line segment structure, LAM will decrease; Time irreversible quantity: first convert the original time series x(t) into incremental time series y(t), which is expressed as follows: y(i)=Δu(i)=x(i+1)-x(i), 1<i≤N Then the time irreversible quantity is calculated by the following method:

Among them, A represents the time irreversible quantity of the nonlinear dissipative system, y _i is the incremental time series of the original time series, N is the length of the signal, and H(*) is a sign function; (3) Splicing of feature vectors and deep long short-term memory neural network prediction ① The eigenvectors of different signal segments are spliced according to the time direction to form a multivariate time-series eigenvector of water content; ②The multivariate time-series feature vector of water content is used as the training data of the deep short-term memory neural network, which is input into the network model for training; the deep long-term short-term memory neural network uses a total of 6 layers of LSTM units, and the hyperparameters of the deep long-term short-term memory neural network are set. The maximum number of iterations is 10,000 to end the training, where the batch size is 100, the time step is 150, and the number of LSTM units is 128; there are three functions inside each LSTM unit, which are input gate function, forget gate function and output gate function , where the input gate determines how much input value information at the current moment is added to the state of the LSTM unit, the forget gate determines how much information is discarded from the LSTM state, and the output gate determines what value needs to be output according to the current state of the LSTM unit; the formulas are as follows: input _t = σ(W _i *[h _t-1 , x _t ]+b _i ) forget _t ＝σ(W _f *[h _t-1 ，x _t ]+b _f ) output _t ＝σ(W _o *[h _t-1 ，x _t ]+b _o ) Among them, W _i , W _f and W _o represent the weight parameters corresponding to the input gate, forget gate and output gate respectively, b _i , b _f and b _o correspond to the bias items respectively, and h _t-1 is the LSTM unit at the previous moment Internal state, x _t is the input value at the current moment; After the water content multivariate time-series feature vector t1 is input to the first layer of LSTM units, it must go through the calculation of the above three gate functions and determine the output of the LSTM; after calculating the feature sequence at the current moment, the LSTM unit moves to the next moment t2, Repeat the above process and calculate the output; after calculating the first layer of LSTM units, the output vector of the first layer is used as the input vector of the second layer of LSTM units, the process is the same as above; the output of each layer of LSTM units is the input of the next layer; During the training process, the multiple time-series feature signals of water cut are input into the LSTM unit in the deep long short-term memory network in sequence according to time for training. During the training process, the classification value is predicted by the deep long short-term memory neural network and compared with the actual wellhead water cut test value ; ③Use the Softmax function to judge, pass the judgment result back to the deep long short-term memory neural network and update the network parameters layer by layer; the Softmax function can compress a K-dimensional vector Z containing any real number to another K-dimensional real vector σ( Z), so that the range of each element is between (0,1), and the sum of all elements is 1, the Softmax form is:

Among them, σ(z) is the output value of the Softmax function in the water content prediction network, and is also the predicted value of the water content; j=1,...,K, i represents a certain classification in K, and zj represents the value of this classification; ④The trained model predicts the moisture content When predicting, after inputting the multivariate time-series characteristic signal of water content into the deep long-short-term memory network, the output value of the Softmax function is the water content of the current signal.

Images