CN110630256B - Low-gas-production oil well wellhead water content prediction system and method based on depth time memory network - Google Patents
Low-gas-production oil well wellhead water content prediction system and method based on depth time memory network Download PDFInfo
- Publication number
- CN110630256B CN110630256B CN201910613947.2A CN201910613947A CN110630256B CN 110630256 B CN110630256 B CN 110630256B CN 201910613947 A CN201910613947 A CN 201910613947A CN 110630256 B CN110630256 B CN 110630256B
- Authority
- CN
- China
- Prior art keywords
- time
- water content
- frequency
- recursion
- characteristic
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 title claims abstract description 135
- 238000000034 method Methods 0.000 title claims abstract description 54
- 230000015654 memory Effects 0.000 title claims abstract description 48
- 239000003129 oil well Substances 0.000 title claims abstract description 28
- 238000004519 manufacturing process Methods 0.000 title claims abstract description 26
- 239000013598 vector Substances 0.000 claims abstract description 53
- 238000005259 measurement Methods 0.000 claims abstract description 19
- 238000012549 training Methods 0.000 claims abstract description 19
- 230000008859 change Effects 0.000 claims abstract description 5
- 239000011159 matrix material Substances 0.000 claims description 46
- 230000006870 function Effects 0.000 claims description 45
- 238000013528 artificial neural network Methods 0.000 claims description 35
- 229910052751 metal Inorganic materials 0.000 claims description 23
- 239000002184 metal Substances 0.000 claims description 23
- 238000009826 distribution Methods 0.000 claims description 18
- 238000000605 extraction Methods 0.000 claims description 16
- 230000008569 process Effects 0.000 claims description 15
- 230000001681 protective effect Effects 0.000 claims description 13
- 229910001220 stainless steel Inorganic materials 0.000 claims description 13
- 239000010935 stainless steel Substances 0.000 claims description 13
- 238000004458 analytical method Methods 0.000 claims description 11
- 230000005284 excitation Effects 0.000 claims description 11
- 230000002427 irreversible effect Effects 0.000 claims description 11
- 238000004445 quantitative analysis Methods 0.000 claims description 9
- 238000004364 calculation method Methods 0.000 claims description 8
- 230000007787 long-term memory Effects 0.000 claims description 7
- 230000006403 short-term memory Effects 0.000 claims description 6
- 238000010408 sweeping Methods 0.000 claims description 6
- 239000012634 fragment Substances 0.000 claims description 5
- 238000005070 sampling Methods 0.000 claims description 5
- 238000009434 installation Methods 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 4
- 238000012360 testing method Methods 0.000 claims description 4
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 239000004744 fabric Substances 0.000 claims description 2
- 210000002268 wool Anatomy 0.000 claims description 2
- 239000003990 capacitor Substances 0.000 claims 3
- 238000011156 evaluation Methods 0.000 claims 2
- 239000007788 liquid Substances 0.000 abstract description 18
- 238000003556 assay Methods 0.000 abstract description 2
- 239000003921 oil Substances 0.000 description 11
- 238000010586 diagram Methods 0.000 description 5
- 239000000284 extract Substances 0.000 description 5
- 230000035699 permeability Effects 0.000 description 5
- 125000004122 cyclic group Chemical group 0.000 description 4
- 239000010779 crude oil Substances 0.000 description 3
- 239000012530 fluid Substances 0.000 description 3
- 230000011218 segmentation Effects 0.000 description 3
- 238000012706 support-vector machine Methods 0.000 description 3
- 230000005514 two-phase flow Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 235000017166 Bambusa arundinacea Nutrition 0.000 description 1
- 235000017491 Bambusa tulda Nutrition 0.000 description 1
- 241001330002 Bambuseae Species 0.000 description 1
- RYGMFSIKBFXOCR-UHFFFAOYSA-N Copper Chemical compound [Cu] RYGMFSIKBFXOCR-UHFFFAOYSA-N 0.000 description 1
- 235000015334 Phyllostachys viridis Nutrition 0.000 description 1
- 239000011425 bamboo Substances 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000033558 biomineral tissue development Effects 0.000 description 1
- 229910052802 copper Inorganic materials 0.000 description 1
- 239000010949 copper Substances 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 239000011521 glass Substances 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000008092 positive effect Effects 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 210000000664 rectum Anatomy 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 238000003466 welding Methods 0.000 description 1
Images
Classifications
-
- E—FIXED CONSTRUCTIONS
- E21—EARTH OR ROCK DRILLING; MINING
- E21B—EARTH OR ROCK DRILLING; OBTAINING OIL, GAS, WATER, SOLUBLE OR MELTABLE MATERIALS OR A SLURRY OF MINERALS FROM WELLS
- E21B49/00—Testing the nature of borehole walls; Formation testing; Methods or apparatus for obtaining samples of soil or well fluids, specially adapted to earth drilling or wells
- E21B49/08—Obtaining fluid samples or testing fluids, in boreholes or wells
- E21B49/086—Withdrawing samples at the surface
-
- E—FIXED CONSTRUCTIONS
- E21—EARTH OR ROCK DRILLING; MINING
- E21B—EARTH OR ROCK DRILLING; OBTAINING OIL, GAS, WATER, SOLUBLE OR MELTABLE MATERIALS OR A SLURRY OF MINERALS FROM WELLS
- E21B49/00—Testing the nature of borehole walls; Formation testing; Methods or apparatus for obtaining samples of soil or well fluids, specially adapted to earth drilling or wells
- E21B49/08—Obtaining fluid samples or testing fluids, in boreholes or wells
- E21B49/087—Well testing, e.g. testing for reservoir productivity or formation parameters
- E21B49/0875—Well testing, e.g. testing for reservoir productivity or formation parameters determining specific fluid parameters
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/02—Agriculture; Fishing; Forestry; Mining
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Theoretical Computer Science (AREA)
- Mining & Mineral Resources (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Geology (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Economics (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- General Business, Economics & Management (AREA)
- General Life Sciences & Earth Sciences (AREA)
- Environmental & Geological Engineering (AREA)
- Fluid Mechanics (AREA)
- Geochemistry & Mineralogy (AREA)
- Tourism & Hospitality (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Primary Health Care (AREA)
- Marine Sciences & Fisheries (AREA)
- Agronomy & Crop Science (AREA)
- Animal Husbandry (AREA)
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
Technical Field
The invention belongs to the field of crude oil production, relates to water content measurement of produced liquid of a low-gas-production oil well, and particularly relates to a system and a method for predicting water content of a wellhead of the low-gas-production oil well based on a depth time memory network.
Background
In the process of crude oil production, the water content parameters of the produced liquid of the oil well are mastered and controlled in time, so that the method is not only a premise of reliably estimating the net yield of crude oil, but also a basis for correctly diagnosing and maintaining problems of the oil well, and is also an important guide index for adjusting the exploitation mode of an oil reservoir, and therefore, the method has important significance for detecting the water content parameters of the produced liquid of the oil well. At present, the ultrahigh water content characteristic of oil field produced liquid puts forward a new requirement on the measurement of the water content of oil well produced liquid, and how to accurately acquire the water content information of the high water content oil well produced liquid becomes a problem to be solved urgently. Currently, the detection of the water content of oil well production liquid is usually realized by a specially designed sensor, and the measuring method comprises an ultrasonic method, an optical method, a ray method, an imaging method, an electric conduction method, an electric method and the like. However, the measurement effect of the existing sensor can not meet the requirement under the working condition of high water content of oil well produced liquid, and is represented by the fact that the response nonlinearity and the water content resolution ratio of the sensor are low, and the measurement result is greatly influenced by the mineralization degree; in addition, the traditional assay method in the oilfield operation is greatly influenced by the sampling condition and the sampling frequency, the measurement period is long, and the real-time measurement is difficult to realize. Although the soft measurement of the water content of the oil-water two-phase flow is widely applied through shallow networks such as a neural network or a support vector machine, characteristics of the shallow network structure need to be carefully designed in the application process, generally, the shallow characteristics have strong subjectivity, and the prediction result of the model on the water content is also greatly influenced by the designed characteristics.
Through the search of the patent publications, two patent publications similar to the purpose and technical scheme of the patent application are found:
1. a method for predicting the initial water content of an ultra-low permeability sandstone reservoir oil well during production operation (109447342A), which comprises the following steps: collecting, sorting and selecting ultra-low permeability sandstone oil reservoir calculation parameters; and predicting the initial water content of the ultra-low permeability sandstone reservoir oil well in the production process by using the functional relation between the effective stress and the water saturation. The method for predicting the initial water content of the ultra-low permeability sandstone reservoir oil well in the production period provides theoretical basis for explaining and revealing the initial water content of the oil reservoir oil well in the production period and predicting the initial water content of the oil well in the production period, and achieves the purpose of dynamically predicting the initial water content of the ultra-low permeability sandstone reservoir oil well in the production period, so that the method has certain theoretical and practical significance.
2. A multi-model prediction method (105631554A) for oil water content of an oil well based on a time sequence is characterized by comprising the following steps: 1) Establishing an oil well oil water content data set of { xi, i =1,2, \ 8230;, N } by using historical data; 2) Preprocessing data in an oil water content data set { xi, i =1,2, \8230;, N } by adopting a wavelet analysis method; 3) Classifying the { xi } Wave by a neighbor propagation clustering algorithm; 4) And representing the data in each cluster by the following time series form: 5) And establishing a time series model of each cluster according to an extreme learning machine algorithm and obtaining a predicted value by using the time series model. It has solved the artifical sample of current oil fluid moisture content wastes time and energy, influences the problem of production control and oil recovery data's real-time.
Through comparison of technical characteristics, the adopted oil reservoir calculation parameters and modes in the comparison document 1 are fundamentally different from those in the application of the invention; in contrast, in the comparison document 2, although the moisture content is predicted in a time series manner, the moisture content model and the manner thereof are fundamentally different from those of the present invention, and thus, the present invention is not substantially creatively affected.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides 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.
The purpose of the invention is realized by the following steps:
a low gas production oil well wellhead water content prediction system based on a depth long-short time memory network is composed of a double-ring type high-frequency capacitance sensor, a water content multi-element time sequence feature extraction module and a well wellhead water content prediction network based on the long-short time memory network, wherein the double-ring type capacitance sensor is used for acquiring well wellhead water content information, a high-frequency sine excitation signal source of the water content multi-element time sequence feature extraction module generates an excitation signal, the excitation signal is sent to an annular measuring electrode of the double-ring type capacitance sensor through a power divider to carry out frequency sweeping, the annular measuring electrode excites water content data measured by the frequency sweeping and then enters a frequency mixer to carry out signal frequency mixing, and the frequency-mixed signal is subjected to adder and voltage bias to obtain a water content time sequence feature vector; the wellhead water content prediction network based on the depth long-short time memory network splices the obtained water content multi-element time sequence feature vectors according to the time sequence, the obtained water content multi-element time sequence feature vectors are used as input vectors of the depth long-short time memory neural network, an LSTM unit is arranged in the depth long-short time memory neural network, three functions including an input gate, a forgetting gate and an output gate are arranged in the unit, the depth long-short time memory neural network has 6 layers, and a Softmax classification function is used as an output function to output a predicted value.
And, two ring type capacitance sensor comprises stainless steel metal protection shell and inside sensor pipeline, stainless steel metal protection shell both ends are provided with left flange, right flange, wherein the metal protection shell at right flange and left flange place is threaded connection, metal protection shell both ends and well head pipe connection, radially open the pin hole that has at stainless steel metal protection shell lateral wall, the inside coaxial inside sensor pipeline that inlays and be equipped with a woollen goods material of stainless steel metal protection shell, interval installation has two cyclic annular measuring electrode on inside sensor pipeline outer wall, install the electromagnetic shield layer in cyclic annular measuring electrode outside, inside sensor pipeline compresses tightly sealedly with metal casing through the O type circle of both sides terminal surface.
And the window function of the water content multivariate timing characteristic extraction module adopts a non-overlapping window with the window size of 1000 to divide the water content signal for multiple times, extract the water content multivariate characteristic sequences of different time periods, distribute the segments of the water content multivariate characteristic sequences by adopting WVD to obtain a time-frequency domain matrix, process the signal by adopting a recursive graph analysis method to obtain a recursive graph matrix, respectively extract time-frequency energy and time-frequency entropy characteristics from the time-frequency domain matrix, respectively extract the characteristics of recursion rate, certainty, average diagonal length, hierarchy and time irreversibility from the recursive graph matrix, and the extracted characteristic vectors totally account for the seven characteristic parameters.
A prediction method of low gas production oil well wellhead water content based on a depth long-term and short-term memory neural network comprises the following steps:
the method comprises the following steps of double-ring type high-frequency capacitance sensor installation and working parameter setting:
installing a sensor on a wellhead descending pipeline, and carrying out frequency sweeping operation on the sensor so as to determine the optimal working frequency of the sensor; after the optimal working frequency of the sensor is determined, exciting the annular measuring electrode by adopting a high-frequency sinusoidal excitation signal source, and taking the amplitude attenuation and the phase attenuation of the measured microwave signal after passing through the sensor as the original measurement information of the water content; the dual-ring high-frequency capacitance sensor adopts a continuous measurement mode for measuring the wellhead content, the sampling frequency is set to be 10 times per minute, and the measurement data is a typical time sequence of reaction content change;
preprocessing of sensor acquisition signals
Windowing and dividing the signals, setting the window size of window function dividing signals to be 1000, enabling the windows to be free of overlapping windows, obtaining a one-dimensional time sequence of the current time period by dividing the signals, extracting water content multi-element characteristic sequences of different time periods by multiple times of division, and taking out numerical values in serial port dividing signals according to the time direction to obtain the water content multi-element characteristic sequences; the characteristic extraction module carries out time-frequency joint distribution and recursive graph analysis on the obtained water content fluctuation sequence fragments to obtain a time-frequency graph matrix and a recursive graph matrix, and corresponding characteristic vectors of each fragment are obtained through calculation; the multivariate time sequence feature vector comprises 7 dimensions, namely time-frequency energy, time-frequency entropy, recursion rate, recursion certainty, recursion average diagonal length, recursion hierarchy and time irreversible capacity; the 7-dimensional feature extraction method is as follows:
firstly, carrying out time-frequency domain analysis on the collected and processed signals, and carrying out Wigner-Ville distribution on each windowed and segmented time sequence segment; firstly, performing Hilbert transform on a signal, and then performing Hilbert transform on the signal by using a formula:
wherein f is frequency, t is time, tau is time delay, z (t) is an analytic form of an original signal, time-frequency graphs under different time segments are obtained, and then time-frequency energy and time-frequency entropy are calculated for a time-frequency graph matrix; wherein:
time-frequency energy: calculating the time-frequency distribution of the windowed time segment as P (t, f), the time-frequency energy E can pass through
Calculated as follows:
time-frequency entropy: calculating the time-frequency distribution of the windowed time segment as P (t, f), and dividing the time-frequency plane into N
The energy of each block is P in the rectangle with the same block size i The energy of the whole time-frequency plane is E, the time-frequency entropy
Can be calculated in the following way:
then carrying out recursive domain quantitative analysis on the collected and processed signals, wherein the recursive quantitative analysis indexes comprise a recursion rate, a certainty, an average diagonal length, a hierarchy and a time irreversible amount; wherein:
the recursion rate: calculating a recursion matrix RR of the windowed time segment, wherein the recursion rate is the percentage of recursion points in the recursion graph plane to the total number of receivable points of the plane, and can be calculated in the following way:
it shows the proportion of phase space points close to each other in the m-dimensional phase space to the total number of points;
certainty: calculating the recursion matrix RR of the windowed time segment, the certainty is the percentage of all recursion points of the recursion points forming the line segment along the diagonal direction, and can be calculated by the following method:
wherein the length of the diagonal line segment is greater than a predetermined lower limit l min Counting is started only when the counter reaches a predetermined value, < i > l min Typically chosen to be an integer no less than 2, DET distinguishes isolated recursion points in the recursion map from organized recursion points forming segments of successive diagonal directions, the more developed the line texture along the main diagonal in the recursion map, the stronger the certainty of the system;
average diagonal length: calculating a recursive matrix RR of the windowed time segments, wherein the certainty is a weighted average of the lengths of the diagonal line segments, and the method can be calculated by the following steps:
the average diagonal length L represents the time length of two phase trajectories close to each other in the phase space trajectory, or represents the average period of the system, without accounting for the main diagonal. The larger L, the more deterministic the system is;
layering: calculating a recursion matrix RR of the windowing time segment, wherein the hierarchy is the percentage of recursion points forming a vertical line segment to all recursion points, and the recursion matrix RR can be calculated in the following mode:
time irreversible capacity: the original time series x (t) is first converted into an incremental time series y (t), which represents
The following were used:
y(i)＝Δu(i)＝x(i+1)-x(i)，1＜i≤N
the time irreversibility can be calculated by:
wherein A represents the time irreversibility of the nonlinear dissipative system, y i In increments of the original time series
An inter-sequence, N being the length of the signal, H (—) being a sign function;
thirdly, splicing characteristic vectors and predicting depth long-time and short-time memory neural network
(1) Splicing the feature vectors of different signal segments according to the time direction to form a water content multi-element time sequence feature vector;
(2) the water content multi-element time sequence feature vector is used as training data of a deep long-short time memory neural network and is input into a network model for training; the depth long-short term memory neural network adopts 6 layers of LSTM units, the ultra-parameters of the depth long-short term memory neural network are set, the training is finished through the maximum iteration times of 10,000 times, wherein the batch size is 100, the time step is 150, and the number of the LSTM units is 128; three functions are arranged in each LSTM unit and are respectively an input gate function, a forgetting gate function and an output gate function, wherein the input gate determines how much input value information at the current moment is added into the state of the LSTM unit, the forgetting gate determines how much information is discarded from the state of the LSTM unit, and the output gate determines what value needs to be output according to the current state of the LSTM unit; the formulas are respectively 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 )
wherein W i 、W f And W O Respectively represent the corresponding weight parameters of the input gate, the forgetting gate and the output gate, b i 、 b f And b o Respectively corresponding to the bias terms, h t-1 Internal state of LSTM cell, x, at the previous moment t Is the input value at the current moment;
after the water content multi-element time sequence feature vector t1 is input into the first layer of LSTM unit, all the three kinds of gate functions are calculated, and the LSTM output is determined; after calculating the characteristic sequence of the current moment, the LSTM unit moves to the next moment t2, repeats the above process and calculates the output; after the first layer of LSTM units are calculated, the output vector of the first layer is used as the input vector of the second layer of LSTM units, and the process is the same as the above; the output of each layer of LSTM units is the input of the next layer;
in the training process, multi-dimensional characteristic time sequence signals are sequentially input into LSTM units in a depth long-short time memory network according to time for training, and the classification values are predicted through the depth long-short time memory neural network in the training process and are compared with the actual wellhead water content test values;
(3) judging through a Softmax function, reversely transmitting a judgment result to the depth long-time memory neural network, and updating network parameters layer by layer; the Softmax function compresses a K-dimensional vector Z containing arbitrary real numbers into another K-dimensional real vector σ (Z) such that each element ranges between (0, 1) and the sum of all elements is 1, the form of Softmax:
wherein j =1, \8230;, K, i represents a certain class in K, z j A value representing the classification;
(4) the trained model can be used for predicting the water content
During prediction, after the multi-dimensional characteristic time sequence signals are input into the depth long-short time memory network, the output value of the Softmax function is the water content of the current signals.
The invention has the advantages and positive effects that:
1. the double-ring type capacitance sensor adopted by the system can quickly and accurately obtain the water content sequence fluctuation signal; windowing the signals, and extracting water content multivariate characteristic sequences of different time periods; the time-frequency domain and the recursion domain of the windowed signal are analyzed to obtain a multivariate characteristic value, so that the multidimensional characteristic of the signal can be highlighted; the water content value of the wellhead can be accurately predicted by training the multidimensional characteristic sequence by a depth long-time memory (LSTM) neural network.
2. The dual-ring type capacitance sensor adopted by the system is arranged on a wellhead descending pipeline, the water content of the produced liquid can be directly measured as soon as possible, the measured value can truly reflect the liquid production condition of the measured oil well, and the system has important significance for guiding the optimization management of the oil field. Compared with the existing sensor, the sensor has stronger stability, the shielding layer can effectively shield the scattering of microwaves and the interference of external electromagnetic waves, and the signal is locked within a certain range. The sensor can effectively and accurately measure the gas-liquid flowing condition in the low-gas-production oil well pipeline.
3. The system adopts a depth long-time memory (LSTM) neural network, is very suitable for processing the problem highly related to the time sequence, can effectively avoid the problems of gradient disappearance, gradient explosion and the like compared with the traditional identification mode, such as a Support Vector Machine (SVM), a Recurrent Neural Network (RNN) and the like, and can enhance the network learning capability by three gate functions in the system, thereby being capable of improving the prediction accuracy by about 5-10 percent compared with the network model.
4. The method extracts the sensor measurement time sequence signal to take value as the characteristic, splices the characteristic of each water content fluctuation sequence segment, the spliced characteristic is the characteristic vector of the signal segment, the characteristic vector contains rich wellhead water content information, inputs the time sequence characteristic into a depth time memory network, can capture the basic characteristic and rule of the water content change, provides rich characteristics for the establishment of a water content prediction model, and can better obtain the characteristic information of the signal in different spaces compared with the method for directly predicting the water content by using the original signal, and can highlight and strengthen the characteristic characteristics of the signal.
5. The method of the invention predicts the water content of the well mouth by using the depth long-time memory (LSTM) neural network from the design of the signal acquisition sensor, the flow is rigorous and feasible, the obtained water content predicted value is accurate, and the network model is smaller, thereby reducing the calculation resources. Because the time sequence characteristics of the water content contain rich flow characteristics, the model provided by the invention can achieve higher water content prediction precision, and the prediction accuracy can reach more than 97%. Compared with the traditional water content prediction method, the method has the advantages of high speed, high accuracy, low computing resource consumption, elimination of artificial subjective factors and the like.
Drawings
FIG. 1 is a schematic diagram of a dual ring high frequency capacitive sensor for wellhead fluid production content measurement according to the present invention;
FIG. 2 is a schematic view of electrical control and windowed segmentation of wellhead water content in accordance with the present invention;
FIG. 2-1 is an enlarged schematic view of the signal diagram of FIG. 2;
FIG. 3 is a schematic diagram of the wellhead water content feature extraction of the present invention;
FIG. 4 is a flow chart of water cut feature prediction according to the present invention.
Detailed Description
The invention is further illustrated by the following examples: the following examples are illustrative and not intended to be limiting, and are not intended to limit the scope of the invention.
A low gas production oil well wellhead water content prediction system based on a depth long-time memory network is composed of a double-ring high-frequency capacitance sensor, a water content multi-element time sequence feature extraction module and a well wellhead water content prediction network based on the long-time memory network.
The double-ring type capacitance sensor is used for acquiring wellhead water content information, is structurally shown in figure 1 and comprises a stainless steel metal protective shell and an internal sensor pipeline 3, wherein two ends of the stainless steel metal protective shell are a left flange 1 and a right flange 9 with a nominal diameter DN50, and the right flange is in threaded connection with the metal protective shell where the left flange is located so as to facilitate installation of the internal sensor pipeline. Two ends of the metal protective shell are connected with a wellhead pipeline, and the side wall of the stainless steel metal protective shell is radially provided with a lead hole 5 with the inner diameter of 18mm for connecting a sensor electrode with an external measuring and calculating instrument; an inner sensor pipeline made of wool fabric with the inner diameter of 50mm is coaxially embedded in the stainless steel metal protective shell and used for transmitting oil-water mixed liquid at a well head; two annular measuring electrodes 6 are installed on the outer wall of the internal sensor pipeline at intervals and used for measuring the water content of the oil-water mixed liquid. Meanwhile, an electromagnetic shielding layer 4 is arranged on the outer side of the annular measuring electrode so as to improve the measuring effect of the sensor. The internal sensor pipeline is tightly pressed and sealed with the metal shell through the O-shaped rings 2 on the end surfaces of the two sides, so that leakage of liquid produced at the wellhead is prevented.
In this embodiment, stainless steel metal protective housing flange interval is 330mm, woollen goods pipeline length in the sensor is 310mm, sensor pipeline latus rectum is 50mm, woollen goods pipeline wall thickness 80mm, cyclic annular measuring electrode internal diameter 80mm, external diameter 85mm, two cyclic annular measuring electrode intervals 50mm, the electromagnetic shield layer is 1 mm's metal copper plate for thickness, the roll welding is a cylinder section of thick bamboo, length is 90mm, the internal diameter is 90mm, and it supports to have organic glass ring 7 between the woollen goods pipeline.
The water content multivariate time sequence feature extraction module is shown in fig. 2 and 3. In fig. 2, a high-frequency sinusoidal excitation signal source generates an excitation signal, the excitation signal is sent to an annular measuring electrode of a sensor through a power divider to sweep frequency, the annular measuring electrode excites moisture content data measured by the sweep frequency and then enters a mixer to perform signal mixing, and the signal after the signal mixing is subjected to voltage bias through an adder to obtain a moisture content multivariate characteristic sequence. In this embodiment, the window function of the signal adopts a non-overlapping window with a window size of 1000, so that the moisture content signal can be segmented for multiple times, and the multiple segmentation can extract the moisture content multivariate feature sequences of different time periods. And (3) carrying out WVD distribution (time-frequency joint distribution) on the fragments of the water content multi-element characteristic sequence to obtain a time-frequency domain matrix, and processing the signals by adopting a recursive graph analysis method to obtain a recursive graph matrix, which is shown in figure 3. And respectively extracting time-frequency energy and time-frequency entropy characteristics from the time-frequency domain matrix, respectively extracting characteristics of recursion rate, determinacy, average diagonal length, hierarchy and time irreversible capacity from the recursion diagram matrix, and totaling the seven characteristic parameters from the extracted water content multivariate time sequence characteristic vectors.
The time-frequency domain matrix and the recursion diagram matrix are extracted for quantitative analysis, the method has strong advancement, the signals can be analyzed in different dimensions, and the features of the different dimensions are extracted, so that compared with the method that the signals are directly used as data sources, the method not only increases the dimensions, but also highlights and enhances the features of the signals.
The wellhead water content prediction network based on the depth long-short time memory network splices the obtained water content multi-time sequence feature vectors according to the time sequence, and the obtained water content multi-time sequence feature vectors are used as input vectors of a depth long-short time memory (LSTM) neural network, and the structure of the prediction network is shown in fig. 4. The depth long-short time memory (LSTM) neural network is internally provided with an LSTM unit, and the unit is internally provided with an input gate, a forgetting gate and an output gate which are three functions respectively. The LSTM units have 6 layers in total. And outputting the predicted value by adopting a Softmax classification function as an output function. The prediction truth value is a water content test value of the wellhead and is used for reversely correcting depth long-time memory (LSTM) network internal parameters to achieve the purpose of prediction.
A method for predicting the water content of a low gas production oil well wellhead based on a depth long-time memory (LSTM) neural network comprises the following steps:
the double-ring high-frequency capacitive sensor mounting and working parameter setting method includes
The sensor is arranged on a wellhead descending pipeline and is connected with an access pipeline through a DN50 flange. The sensor is then swept to determine the optimum operating frequency of the sensor. The sweep frequency band of the sensor is set to be 0.8Ghz-10GHz, which is a microwave band. After the optimal working frequency of the sensor is determined, the annular measuring electrode is excited at the frequency, and the amplitude attenuation and the phase attenuation of the microwave signal after passing through the sensor are measured to serve as the raw measurement information of the water content. The double-ring high-frequency capacitance sensor adopts a continuous measurement mode for measuring the wellhead content, the sampling frequency is set to be 10 times per minute, the measured data is a typical time sequence for reflecting the content change, and the measured time sequence value of the sensor can be uploaded to a server in a wireless transmission mode for storage and analysis operation.
Preprocessing of sensor acquisition signals
Windowing and dividing signals, as shown in the right side of the figure 2, setting the window size of a window function dividing signal to be 1000, setting non-overlapping windows between the windows, dividing the signals to obtain a one-dimensional time sequence of the current time period, extracting the water content multivariate characteristic sequences of different time periods through multiple division, and taking out numerical values in each serial port dividing signal according to the time direction to obtain the water content multivariate characteristic sequence. The characteristic extraction module performs time-frequency joint distribution and recursive graph analysis on the obtained water content fluctuation sequence segments as shown in FIG. 3 to obtain a time-frequency graph matrix and a recursive graph matrix, and corresponding characteristic vectors of each segment are obtained through calculation of corresponding formulas; the multivariate time sequence feature vector comprises 7 dimensions, namely time-frequency energy, time-frequency entropy, recursion rate, recursion certainty, recursion average diagonal length, recursion hierarchy and time irreversible capacity. The 7-dimensional feature extraction method is as follows:
firstly, time-frequency domain analysis is carried out on the collected and processed signals, and Wigner-Ville distribution (WVD) is carried out on each time sequence segment after windowing and segmentation. The signal is first subjected to a Hilbert transform (Hilbert transform) and then to a transform by the formula:
(where f is frequency, t is time, τ is time delay, and z (t) is the analytic form of the original signal)
And obtaining time-frequency graphs under different time slices, and then solving time-frequency energy and time-frequency entropy for the time-frequency graph matrix.
Wherein:
time-frequency energy: calculating the time-frequency distribution of the windowed time segment as P (t, f), the time-frequency energy E can pass through
Calculated as follows:
2. time-frequency entropy: calculating the time-frequency distribution of the windowed time segments as P (t, f), and dividing the time-frequency plane into
N rectangles with equal size are set that the energy of each block is P i The energy of the whole time-frequency plane is E, then
The frequency entropy can be calculated by:
aiming at the time-frequency joint distribution characteristic of the vertical oil-water two-phase flow with low flow rate and high water content in quantitative analysis, quadratic time-frequency distribution can more reasonably and visually reflect the fluid characteristics, wherein time-frequency energy and time-frequency entropy can directly reflect the characteristics of a time-frequency graph, and the quadratic time-frequency distribution is an important characteristic of time-frequency distribution.
And then carrying out recursive domain quantitative analysis on the acquired and processed signals, wherein the indexes of the recursive quantitative analysis comprise a recursion rate, a certainty, an average diagonal length, a hierarchy and a time irreversible amount. Wherein:
recursion rate: calculating a recursion matrix RR of the windowed time segment, wherein the recursion rate is the percentage of recursion points in the recursion graph plane to the total number of receivable points of the plane, and can be calculated in the following way:
it shows the proportion of phase space points close to each other in the m-dimensional phase space to the total number of points;
certainty: calculating the recursive matrix RR of the windowed time segment, the certainty is the percentage of all recursive points of the recursive points forming the line segment along the diagonal direction, and can be calculated as follows:
in the formula, the number of segments having a length of l is shown. Only the length of the diagonal line segment is greater than a predetermined lower limit l min The counting is started. l min And is generally selected to be an integer of not less than 2. DET will isolate in recursion mapThe discrete recursion points are distinguished from organized recursion points forming successive diagonal line segments. The more developed the line texture along the main diagonal in the recursive graph, the stronger the certainty of the system is shown;
average diagonal length: calculating a recursive matrix RR of the windowed time segments, wherein the certainty is a weighted average of the lengths of the diagonal line segments, and the method can be calculated by the following steps:
the average diagonal length L represents the time length of two phase trajectories close to each other in the phase space trajectory, or represents the average period of the system, without accounting for the main diagonal. The larger L, the more deterministic the system is.
Layering: calculating a recursion matrix RR of the windowing time segment, wherein the hierarchy is the percentage of recursion points forming a vertical line segment to all recursion points, and the recursion matrix RR can be calculated in the following mode:
in the formula, P (v) is the number of segments having a length v. Only the length of the diagonal line segment is greater than a predetermined lower limit v min The counting is started. v. of min And is generally selected to be an integer of not less than 2. LAM represents the probability of recursion points of the hierarchical state in the system, and when the recursion points are isolated more than the line segment structure in the vertical direction in the recursion graph, LAM is reduced. Setting v of the invention min Is 2.
Time irreversible amount: the original time series x (t) is first converted to an incremental time series y (t), which represents
The following were used:
y(i)＝Δu(i)＝x(i+1)-x(i)，1＜i≤N
the time irreversibility can be calculated by:
wherein A represents the time irreversibility of the nonlinear dissipative system, y i The incremental time series of the original time series, N is the length of the signal and H (×) is a sign function.
The recursive graph quantitative analysis is beneficial to supplement and explore for disclosing a two-phase flow pattern conversion mechanism which has complexity and uncertainty and is difficult to accurately describe by a mathematical model.
Performing characteristic vector splicing and depth long-term memory (LSTM) neural network prediction
(1) Splicing the feature vectors of different signal segments according to the time direction (t) 1 ，t 2 ......t n ) And forming a water content multi-element time sequence feature vector as shown on the left side of the figure 4. The characteristic vector is formed by splicing in the time direction, so that the time dimension characteristic is reserved, and the vector can visually reflect the sequence characteristic.
(2) And then, inputting the water content multi-element time sequence feature vector as training data of a deep long-short time memory (LSTM) neural network into a network model for training. The depth long-short time memory (LSTM) neural network totally adopts 6 layers of LSTM units, the super-parameter of the depth long-short time memory (LSTM) neural network is set, the training is finished through 10,000 times of maximum iteration, wherein the batch size is 100, the time step is 150, and the number of the LSTM units is 128. Three functions are arranged in each LSTM unit and are respectively an input gate function, a forgetting gate function and an output gate function, wherein the input gate determines how much input value information at the current moment is added into the state of the LSTM unit, the forgetting gate determines how much information is discarded from the state of the LSTM unit, and the output gate determines what value needs to be output according to the current state of the LSTM unit. The formulas are respectively 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 )
wherein W i 、W f And W O Respectively representing the weight parameters corresponding to the input gate, the forgetting gate and the output gate, b i 、 b f And b o Respectively corresponding to the bias terms, h t-1 Internal state of LSTM cell, x, at the previous moment t Is the input value at the current moment.
Water cut multivariate timing characteristic vector t 1 After the input of the first layer LSTM unit, the calculation of the three gate functions is carried out, and the LSTM output is determined. After calculating the characteristic sequence of the current moment, the LSTM unit goes to the next moment t 2 And moving, repeating the above processes and calculating output. And after the first-layer LSTM unit is calculated, taking the output vector of the first layer as the input vector of the second-layer LSTM unit, and the process is the same as the above. The output of each layer of LSTM cells is the input to the next layer.
In the training process, the multi-dimensional characteristic time sequence signal is according to time (t) 1 ,t 2 ......t n ) And sequentially inputting the depth long-short time memory neural network into LSTM units in the depth long-short time memory network for training, predicting classification values through the depth long-short time memory neural network in the training process, and comparing the classification values with the actual water content test values at the well head.
(3) And (4) judging through a Softmax function, reversely transmitting the judging result back to the depth long-time memory neural network and updating the network parameters layer by layer. The Softmax function "compresses" a K-dimensional vector Z containing arbitrary real numbers into another K-dimensional real vector σ (Z) such that each element ranges between (0, 1) and the sum of all elements is 1, the Softmax form:
wherein j =1, \8230;, K, j represents a certain class in K, z j A value representing the classification.
(4) The trained model can be used for predicting the water content.
During prediction, after the multi-dimensional characteristic time sequence signal is input into a depth long-short time memory network, the output value of the Softmax function is the water content of the current signal.
Claims (2)
1. The utility model provides a low yield gas oil well head moisture content prediction system based on network is recalled to degree of depth length time which characterized in that: the water content multi-element time sequence characteristic extraction device comprises a double-ring high-frequency capacitor sensor, a water content multi-element time sequence characteristic extraction module and a wellhead water content prediction network based on a long-time and short-time memory network, wherein the double-ring high-frequency capacitor sensor is used for acquiring wellhead water content information, a high-frequency sine excitation signal source of the water content multi-element time sequence characteristic extraction module generates an excitation signal, the excitation signal is sent to an annular measuring electrode of the double-ring capacitor sensor through a power divider to carry out frequency sweeping, the annular measuring electrode excites water content data measured by the frequency sweeping and then enters a mixer to carry out signal mixing, and the mixed signal is subjected to adder and voltage bias to obtain the water content multi-element time sequence characteristic; the wellhead water content prediction network based on the long and short term memory network splices the obtained water content multi-element time sequence characteristic vectors according to the time sequence, the obtained water content multi-element time sequence characteristic vectors are used as input vectors of a depth long and short term memory neural network, an LSTM unit is arranged in the depth long and short term memory neural network, three functions of an input gate, a forgetting gate and an output gate are respectively arranged in the unit, the depth long and short term memory neural network has 6 layers, and a Softmax classification function is used as an output function to output a predicted value;
the double-ring high-frequency capacitance sensor consists of a stainless steel metal protective shell and an internal sensor pipeline, wherein a left flange and a right flange are arranged at two ends of the stainless steel metal protective shell, the right flange and the metal protective shell where the left flange is arranged are in threaded connection, two ends of the metal protective shell are connected with a wellhead pipeline, a lead hole is radially formed in the side wall of the stainless steel metal protective shell, the internal sensor pipeline made of wool fabric is coaxially embedded in the stainless steel metal protective shell, two annular measuring electrodes are installed on the outer wall of the internal sensor pipeline at intervals, an electromagnetic shielding layer is installed on the outer side of each annular measuring electrode, and the internal sensor pipeline is tightly pressed and sealed with the metal shell through O-shaped rings on the end faces of two sides;
the window function of the water content multivariate timing characteristic extraction module adopts a non-overlapping window with the window size of 1000 to repeatedly segment water content signals, water content multivariate characteristic sequences of different time periods are extracted, segments of the water content multivariate timing characteristic sequences are distributed by adopting WVD (WVD) to obtain a time-frequency domain matrix, the signals are processed by adopting a recursive graph analysis method to obtain a recursive graph matrix, time-frequency energy and time-frequency entropy characteristics are respectively extracted from the time-frequency domain matrix, recursive rate, certainty, average diagonal length, hierarchy and time irreversible characteristic are respectively extracted from the recursive graph matrix, and the extracted characteristic vectors account for seven characteristic parameters in total.
2. A method for predicting wellhead water content of a low gas production well based on a depth time memory neural network by using the system as claimed in claim 1, wherein the method comprises the following steps: the method comprises the following steps:
the method comprises the following steps of double-ring type high-frequency capacitance sensor installation and working parameter setting:
installing a sensor on a wellhead descending pipeline, and carrying out frequency sweeping operation on the sensor so as to determine the optimal working frequency of the sensor; after the optimal working frequency of the sensor is determined, exciting the annular measuring electrode by adopting a high-frequency sinusoidal excitation signal source, and taking the amplitude attenuation and the phase attenuation of the measured microwave signal after passing through the sensor as the original measurement information of the water content; the dual-ring high-frequency capacitance sensor adopts a continuous measurement mode for measuring the wellhead content, the sampling frequency is set to be 10 times per minute, and the measurement data is a typical time sequence of reaction content change;
preprocessing of sensor acquisition signals
Windowing and dividing signals, setting the window size of window function dividing signals to be 1000, enabling the windows to be free of overlapping windows, obtaining a one-dimensional time sequence of the current time period by dividing, extracting multiple water content time sequence feature sequences of different time periods by multiple times, and taking out numerical values in serial port dividing signals according to the time direction to obtain the multiple water content time sequence feature sequences; the characteristic extraction module carries out time-frequency joint distribution and recursive graph analysis on the obtained water content multi-element time sequence characteristic sequence fragments to obtain a time-frequency graph matrix and a recursive graph matrix, and corresponding characteristic vectors of each fragment are obtained through calculation; the water content multi-element time sequence feature vector comprises 7 dimensions which are time-frequency energy, time-frequency entropy, recursion rate, recursion certainty, recursion average diagonal length, recursion hierarchy and time irreversible; the 7-dimensional feature extraction method is as follows:
firstly, carrying out time-frequency domain analysis on the collected and processed signals, and carrying out Wigner-Ville distribution on each windowed and segmented time sequence segment; the signal is first subjected to a hilbert transform and then, by the formula:
wherein WVD (t, f) is the water content multivariate time sequence characteristic, f is the frequency, t is time, τ is time delay, z (t) is the analytic form of the original signal, z * (t) taking a conjugate function form of z (t), obtaining time-frequency graphs under different time segments, and then solving time-frequency energy and time-frequency entropy for a time-frequency graph matrix; wherein:
time-frequency energy: and calculating the time-frequency distribution of the windowing time segment to be P (t, f), and calculating the time-frequency energy E in the following way:
time-frequency entropy: calculating the time frequency distribution of the windowing time segment as P (t, f), dividing the time frequency plane into N rectangles with equal size, and setting the energy of each block as P i If the energy of the whole time-frequency plane is E, the time-frequency entropy is calculated in the following manner:
then carrying out recursion domain quantitative analysis on the acquired and processed signals, wherein the recursion domain quantitative analysis indexes comprise recursion rate, determinacy, average diagonal length, hierarchy and time irreversible amount; wherein:
recursion rate: calculating a recursion matrix RR of the windowed time segment, wherein the recursion rate is the percentage of recursion points in a recursion graph plane to the total containable points of the plane, and the recursion rate is calculated in the following way:
in the above formula, RR represents a recursive matrix, R i,j The method comprises the steps of calculating the dimension of a recursive matrix, wherein the dimension of the recursive matrix is the element value of the ith row and the jth column in the recursive matrix, i is the ith row index of the recursive matrix, j is the jth column index of the recursive matrix, and N is the dimension of the recursive matrix;
the RR recursion rate indicates a proportion of phase space points close to each other in the m-dimensional phase space to the total number of points;
certainty: calculating the recursion matrix RR of the windowed time segment, wherein the certainty is the percentage of recursion points forming a line segment along the diagonal direction in all recursion points, and the recursion matrix RR is calculated in the following mode:
wherein P (l) is the number of segments having a length of l, and only the length of the segments in the diagonal direction is greater than a predetermined lower limit of l min Then starting counting; l. the min Is selected to be an integer not less than 2; DET distinguishes isolated recursion points in the recursion graph from organized recursion points forming 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 is shown;
average diagonal length: calculating a recursive matrix RR of the windowed time segment, wherein the average diagonal length is a weighted average of lengths of the diagonal line segments, and the average diagonal length is calculated by the following method:
the average diagonal length L represents the time length of two adjacent phase tracks in the phase space track or represents the average period of the system, the main diagonal is not calculated, and the larger L is, the stronger the certainty of the system is;
layering: calculating a recursion matrix RR of the windowing time segment, wherein the hierarchy is the ratio of recursion points forming a line segment in the vertical direction to all recursion points, and the recursion matrix RR is calculated in the following mode:
wherein P (v) is the number of segments having a length v, and only the length of a segment in the diagonal direction is greater than a predetermined lower limit v min The counting is started at the moment v min The number of the selected integers is not less than 2, LAM represents the probability of recursion points of a hierarchical state in the system, and when isolated recursion points in a recursion graph are more than a line segment structure in the vertical direction, LAM is reduced;
time irreversible amount: the original time series x (t) is first converted into an incremental time series y (t), which is expressed as follows:
y(i)＝Δu(i)＝x(i+1)-x(i)，1＜i≤N
the time irreversibility is calculated by:
wherein A represents the time irreversibility of the nonlinear dissipative system, y i An incremental time sequence which is an original time sequence, wherein N is the length of a signal, and H (#) is a sign function;
thirdly, splicing characteristic vectors and predicting depth long-time and short-time memory neural network
(1) Splicing the feature vectors of different signal segments according to the time direction to form a water content multi-element time sequence feature vector;
(2) the water content multi-element time sequence feature vector is used as training data of a deep long-short time memory neural network and is input into a network model for training; the depth long-short time memory neural network adopts 6 layers of LSTM units, the super parameters of the depth long-short time memory neural network are set, the training is finished through 10 to 000 maximum iteration times, wherein the batch size is 100, the time step is 150, and the number of LSTM units is 128; three functions are arranged in each LSTM unit and are respectively an input gate function, a forgetting gate function and an output gate function, wherein the input gate determines how much input value information at the current moment is added into the state of the LSTM unit, the forgetting gate determines how much information is discarded from the state of the LSTM unit, and the output gate determines what value needs to be output according to the current state of the LSTM unit; the formulas are respectively 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 )
wherein W i 、W f And W o Respectively represent the corresponding weight parameters of the input gate, the forgetting gate and the output gate, b i 、b f And b o Respectively corresponding to the bias terms, h t-1 Internal state of LSTM cell, x, at the previous moment t Is the input value at the current moment;
after the water content multivariate timing characteristic vector t1 is input into the first layer of LSTM units, all the calculation of the three gate functions is carried out, and the LSTM output is determined; after calculating the characteristic sequence of the current moment, the LSTM unit moves to the next moment t2, repeats the above process and calculates the output; after the first layer of LSTM units are calculated, the output vector of the first layer is used as the input vector of the second layer of LSTM units, and the process is the same as the above; the output of each layer of LSTM units is the input of the next layer;
in the training process, the water content multi-element time sequence characteristic signals are sequentially input into LSTM units in the depth long-short time memory network according to time to be trained, and the classification values are predicted through the depth long-short time memory neural network in the training process and are compared with the actual wellhead water content test values;
(3) the evaluation is carried out by a Softmax function, and the evaluation result is reversedTransmitting the depth time and memorizing the neural network and updating the network parameters layer by layer; the Softmax function can compress a K-dimensional vector Z containing arbitrary real numbers into another K-dimensional real vector σ (Z) such that each element ranges between (0, 1) and the sum of all elements is of the form 1, softmax:
wherein, sigma (z) is an output value of a Softmax function in the moisture content prediction network and is also a predicted value of the moisture content; j =1, \ 8230;, K, i denotes a certain class in K, zj denotes the value of this class;
(4) moisture content prediction by trained model
During prediction, after the water content multivariate timing characteristic signals are input into the depth long-short time memory network, the output value of the Softmax function is the water content of the current signals.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910613947.2A CN110630256B (en) | 2019-07-09 | 2019-07-09 | Low-gas-production oil well wellhead water content prediction system and method based on depth time memory network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910613947.2A CN110630256B (en) | 2019-07-09 | 2019-07-09 | Low-gas-production oil well wellhead water content prediction system and method based on depth time memory network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110630256A CN110630256A (en) | 2019-12-31 |
| CN110630256B true CN110630256B (en) | 2022-12-02 |
Family
ID=68969589
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910613947.2A Active CN110630256B (en) | 2019-07-09 | 2019-07-09 | Low-gas-production oil well wellhead water content prediction system and method based on depth time memory network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110630256B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111324990A (en) * | 2020-03-19 | 2020-06-23 | 长江大学 | Porosity prediction method based on multilayer long-short term memory neural network model |
| CN112814649B (en) * | 2021-01-04 | 2023-03-10 | 天津科技大学 | Intelligent metering device and method for water content of oil well produced liquid |
| CN112922582B (en) * | 2021-03-15 | 2022-03-11 | 西南石油大学 | Gas well wellhead choke tip gas flow analysis and prediction method based on Gaussian process regression |
| CN113622880B (en) * | 2021-07-28 | 2022-07-22 | 陕西埃菲克能源科技有限公司 | Non-stop interval pumping method based on frequency conversion technology |
| CN114048932A (en) * | 2022-01-13 | 2022-02-15 | 西南石油大学 | Coal-bed gas well gas production rate prediction method based on LSTM |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104189968A (en) * | 2008-06-26 | 2014-12-10 | 甘布罗伦迪亚股份公司 | Method and device for processing a time-dependent measurement signal |
| CN106650929A (en) * | 2016-10-11 | 2017-05-10 | 天津大学 | Recursive-graph-based deep learning model and its application in oil-water phase rate measurement |
| CN108344564A (en) * | 2017-12-25 | 2018-07-31 | 北京信息科技大学 | A kind of state recognition of main shaft features Testbed and prediction technique based on deep learning |
| CN109915113A (en) * | 2019-04-11 | 2019-06-21 | 燕山大学 | Well head oil-gas-water three-phase flow impedance capacitance moisture percentage measuring apparatus and method |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10184320B2 (en) * | 2014-09-02 | 2019-01-22 | Saudi Arabian Oil Company | Systems, methods, and computer medium to enhance hydrocarbon reservoir simulation |
| US10832152B2 (en) * | 2017-11-29 | 2020-11-10 | Saudi Arabian Oil Company | Wet well prediction using real time data |
-
2019
- 2019-07-09 CN CN201910613947.2A patent/CN110630256B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104189968A (en) * | 2008-06-26 | 2014-12-10 | 甘布罗伦迪亚股份公司 | Method and device for processing a time-dependent measurement signal |
| CN106650929A (en) * | 2016-10-11 | 2017-05-10 | 天津大学 | Recursive-graph-based deep learning model and its application in oil-water phase rate measurement |
| CN108344564A (en) * | 2017-12-25 | 2018-07-31 | 北京信息科技大学 | A kind of state recognition of main shaft features Testbed and prediction technique based on deep learning |
| CN109915113A (en) * | 2019-04-11 | 2019-06-21 | 燕山大学 | Well head oil-gas-water three-phase flow impedance capacitance moisture percentage measuring apparatus and method |
Non-Patent Citations (5)
| Title |
|---|
| 基于CPN网络井口计量原油含水率预测模型;王丽娜等;《大庆石油学院学报》;20091231;第33卷(第06期);第101-104页 * |
| 基于数据挖掘的长短期记忆网络模型油井产量预测方法;谷建伟等;《特种油气藏》;20190131;第26卷(第02期);第77-81页 * |
| 基于长短时记忆网络的旋转机械状态预测研究;赵建鹏等;《噪声与振动控制》;20170831;第37卷(第04期);第155-159页 * |
| 应用BP神经网络预测原油含水率问题的研究;曾蕾蕾等;《中国测试技术》;20060731;第32卷(第04期);第25-27页 * |
| 自适应时频分析及其时频属性提取方法研究;张宁;《中国优秀硕士学位论文全文数据库 信息科技辑》;20090215(第2期);第I136-26页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110630256A (en) | 2019-12-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110630256B (en) | Low-gas-production oil well wellhead water content prediction system and method based on depth time memory network | |
| Song et al. | GIS partial discharge pattern recognition via deep convolutional neural network under complex data source | |
| CN110336534B (en) | Fault diagnosis method based on photovoltaic array electrical parameter time series feature extraction | |
| CN100504846C (en) | Electrical energy power quality disturbance automatic identification method and system based on information fusion | |
| CN101334434B (en) | Electromagnetic environment test system for extracting electromagnetic leakage signal by utilizing wavelet transformation | |
| CN110630244B (en) | High-yield gas-oil well water content prediction system and method based on multi-sensor measurement and long-time and short-time memory network | |
| CN103941254A (en) | Soil physical property classification recognition method and device based on geological radar | |
| CN110632132A (en) | High-yield gas-oil well wellhead liquid water content prediction method based on multi-sensor measurement and deep convolutional neural network | |
| Mostajabi et al. | Single-sensor source localization using electromagnetic time reversal and deep transfer learning: application to lightning | |
| CN108090462A (en) | A kind of Emitter Fingerprint feature extracting method based on box counting dimension | |
| CN113126038A (en) | High-frequency ground wave radar working frequency optimization method, system, storage medium and application | |
| Reolon et al. | Unlocking data analytics for the automatic evaluation of cement bond scenarios | |
| CN110243885A (en) | A kind of low yield gas well mouth of oil well hydrated comples ion method based on time-frequency characteristics | |
| CN106788817A (en) | A kind of frequency spectrum sensing method based on bayesian criterion and energy measuring method | |
| CN107966687B (en) | MIMO radar signal modulation type identification method based on partial autocorrelation spectrum | |
| CN110243886A (en) | A kind of low yield gas well mouth of oil well hydrated comples ion method based on nonlinear characteristic | |
| CN104865508A (en) | Partial discharge recognition method based on data grouping quantification | |
| CN117172601A (en) | Non-invasive load monitoring method based on residual total convolution neural network | |
| CN111339986A (en) | Frequency law mining method and system for equipment based on time domain/frequency domain analysis | |
| CN115586406A (en) | GIS partial discharge fault diagnosis method and system based on ultrahigh frequency signal | |
| CN113095215B (en) | Solar radio filtering method and system based on improved LSTM network | |
| Wang et al. | Estimation of soil organic matter by in situ Vis-NIR spectroscopy using an automatically optimized hybrid model of convolutional neural network and long short-term memory network | |
| CN115078327A (en) | Rapid detection method for dangerous chemicals based on neural network architecture search | |
| Zhao et al. | Research on Transformer Oil Multi-frequency Ultrasonic Monitoring Technology Based on Convolutional Neural Network | |
| Mishra et al. | Diagnostics analysis of partial discharge events of the power cables at various voltage levels using ramping behavior analysis method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |