CN114693005B - Three-dimensional underground oil reservoir dynamic prediction method based on convolution Fourier neural network - Google Patents

Three-dimensional underground oil reservoir dynamic prediction method based on convolution Fourier neural network Download PDF

Info

Publication number
CN114693005B
CN114693005B CN202210603016.6A CN202210603016A CN114693005B CN 114693005 B CN114693005 B CN 114693005B CN 202210603016 A CN202210603016 A CN 202210603016A CN 114693005 B CN114693005 B CN 114693005B
Authority
CN
China
Prior art keywords
pressure
convolution
fourier
saturation
dimensional
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
Application number
CN202210603016.6A
Other languages
Chinese (zh)
Other versions
CN114693005A (en
Inventor
张凯
左袁德
王晓雅
张黎明
刘丕养
严侠
张华清
杨勇飞
孙海
张文娟
姚军
樊灵
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China University of Petroleum East China
Original Assignee
China University of Petroleum East China
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by China University of Petroleum East China filed Critical China University of Petroleum East China
Priority to CN202210603016.6A priority Critical patent/CN114693005B/en
Publication of CN114693005A publication Critical patent/CN114693005A/en
Application granted granted Critical
Publication of CN114693005B publication Critical patent/CN114693005B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
    • G06F17/141—Discrete Fourier transforms
    • G06F17/142—Fast Fourier transforms, e.g. using a Cooley-Tukey type algorithm
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/23—Design optimisation, verification or simulation using finite element methods [FEM] or finite difference methods [FDM]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/28—Design optimisation, verification or simulation using fluid dynamics, e.g. using Navier-Stokes equations or computational fluid dynamics [CFD]
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/10—Numerical modelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2113/00—Details relating to the application field
    • G06F2113/08—Fluids
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/14—Force analysis or force optimisation, e.g. static or dynamic forces
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A10/00—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE at coastal zones; at river basins
    • Y02A10/40—Controlling or monitoring, e.g. of flood or hurricane; Forecasting, e.g. risk assessment or mapping

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Business, Economics & Management (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Strategic Management (AREA)
  • Computer Hardware Design (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Human Resources & Organizations (AREA)
  • Economics (AREA)
  • Geometry (AREA)
  • Marketing (AREA)
  • Algebra (AREA)
  • General Business, Economics & Management (AREA)
  • Tourism & Hospitality (AREA)
  • Computational Mathematics (AREA)
  • Discrete Mathematics (AREA)
  • Animal Husbandry (AREA)
  • Databases & Information Systems (AREA)
  • Mining & Mineral Resources (AREA)
  • Marine Sciences & Fisheries (AREA)
  • Primary Health Care (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)

Abstract

The invention discloses a three-dimensional underground oil reservoir dynamic prediction method based on a convolution Fourier neural network, which belongs to the technical field of oil reservoir engineering, and dynamically predicts the residual oil saturation or pressure of the three-dimensional underground oil reservoir by utilizing a constructed 3D convolution Fourier neural network model, and specifically comprises the following steps: collecting three-dimensional underground oil reservoir data; constructing a 3D convolution Fourier neural network, and considering time and space information simultaneously by combining LSTM; setting hyper-parameters of a convolution Fourier network model, and training a 3D convolution Fourier network model; evaluating the performance of the 3D convolutional Fourier neural network; and after the training is finished, outputting a 3D convolution Fourier network model with good test indexes. The method realizes high-precision rapid prediction of the three-dimensional oil reservoir model, well aims at the characteristics of the space-time property and the physical system of the three-dimensional oil reservoir, and better meets the actual field requirements.

Description

Three-dimensional underground oil reservoir dynamic prediction method based on convolution Fourier neural network
Technical Field
The invention belongs to the technical field of oil reservoir engineering, and particularly relates to a three-dimensional underground oil reservoir residual oil saturation or pressure production dynamic prediction method based on a 3D convolution Fourier neural network model.
Background
Numerical simulation of multiphase flow in porous media is critical to many geoscience applications. In the field of oil reservoir engineering, an oil-water two-phase partial differential equation can accurately describe an underground flow system, and a large amount of partial differential equation solution calculation is needed for simulating underground flow, which is also the basis of a numerical simulation technology. The traditional numerical simulation solves partial differential equations based on numerical calculation methods such as finite element method, finite difference method and the like. The reservoir numerical simulation can accurately calculate reservoir states of future reservoirs such as pressure, saturation, yield and the like for dynamic evaluation of production development. The number of real three-dimensional oil reservoir models is large, the models are large and the boundaries are complex, and a large amount of time and computing resources are consumed for solving.
With the deepening of informatization and the continuous development of hardware level, artificial intelligent methods such as deep learning and the like are widely applied to scenes such as image recognition, target detection and the like. In the field of oil reservoir engineering, a data-driven machine learning method provides a faster alternative scheme for a traditional simulator by constructing a neural network proxy model. The mapping between the machine learning agent model and the numerical simulation data provides a faster choice for the traditional simulator, and the prediction time can be effectively saved. Most of the current popular agent model methods are directed at two-dimensional oil reservoirs, more grids are formed for a three-dimensional complex geological model, and meanwhile, the gravity influence is considered, and an oil well penetrates through a plurality of areas. Therefore, the three-dimensional oil reservoir prediction is more challenging and more practical than the two-dimensional oil reservoir prediction. The existing methods applied to the three-dimensional model only consider local information among images, but cannot consider physical information and space-time characteristics, and have the limitations of low precision, poor generalization performance, slow training time and the like.
Disclosure of Invention
Aiming at the problems that the residual oil saturation or pressure prediction time is too long when numerical reservoir simulation is used at present, a common machine learning agent model cannot effectively predict a three-dimensional complex oil reservoir and the like, the invention provides a production dynamic prediction method of the residual oil saturation or pressure of the three-dimensional oil reservoir based on a convolution Fourier neural network model, and the influence of the time-space property, the physical information and the gravity factor of the three-dimensional oil reservoir is considered, so that the prediction precision is effectively improved, the method is suitable for different oil reservoir conditions, and the calculation time is saved.
The technical scheme of the invention is as follows:
a three-dimensional underground oil reservoir dynamic prediction method based on a convolution Fourier neural network utilizes a built 3D convolution Fourier neural network model to dynamically predict the residual oil saturation or pressure of the three-dimensional underground oil reservoir, and specifically comprises the following steps:
step 1, collecting three-dimensional underground oil reservoir data, generating a three-dimensional permeability field according with geological characteristics by using an SGeMS geological statistical tool, solving saturation or pressure within a period of time corresponding to different permeabilities by using a numerical simulator as a sample library, dividing a data set according to a proportion, taking the permeability as input, and taking the saturation or pressure as output;
step 2, constructing a 3D convolution Fourier neural network, combining 3D convolution and Fourier transform, extracting image local information by using a 3D convolution operator, extracting a physical information approximate differential operator by using Fourier transform, and adding gravity constraint in the network; mapping from permeability to an oil reservoir state is realized by utilizing a coding and decoding network, and time and space information is considered simultaneously by combining with an LSTM;
step 3, setting hyper-parameters of the convolution Fourier network model, and training the 3D convolution Fourier network model under a training set;
step 4, evaluating the performance of the 3D convolution Fourier neural network by using the test set, and calculating the relative error between RMSE and the quantized saturation or pressure;
step 5, outputting a 3D convolution Fourier network model with good test indexes after the training is finished; the method comprises the steps of collecting three-dimensional underground oil deposit data in real time by monitoring equipment installed on the basis of an oil deposit, generating a permeability field of the three-dimensional oil deposit on line by using a geological statistical tool, predicting oil deposit saturation or pressure field distribution of the oil deposit within a future period of time on line by using a 3D convolution Fourier network model, further calculating yield through pressure and saturation, and providing visual reference for formulating a production strategy.
Further, the specific process of step 1 is as follows:
step 1.1, collecting three-dimensional underground oil reservoir data, generating a three-dimensional permeability field by using an SGeMS geological statistical model, wherein the grid size is 40 multiplied by 20, and generating 2000 permeability samples;
step 1.2, calculating an oil-water two-phase flow equation of the three-dimensional oil reservoir system by using a finite element method through a numerical simulator to obtain saturation or pressure results in all grid blocks; the oil reservoir underground oil-water flow system is three-dimensional, needs to consider the gravity effect and consists of a mass conservation equation (1) and a phase Darcy velocity equation (2);
the mass conservation equation is expressed as (1),
Figure 393993DEST_PATH_IMAGE001
（1）
wherein the content of the first and second substances,
Figure 933558DEST_PATH_IMAGE002
representing a gradient operator;jindicating different phases, including oil and water phases,oby which is meant an oil,wrepresents water;
Figure 331173DEST_PATH_IMAGE003
represents the density of the phase;v j represents the phase darcy speed;
Figure 639794DEST_PATH_IMAGE004
representing source/sink items, superscriptlRepresenting a parameter at the well;
Figure 713930DEST_PATH_IMAGE005
represents porosity;S j represents the saturation of the phase;trepresents time;
the expression of the facies darcy velocity equation is shown as (2),
Figure 317955DEST_PATH_IMAGE006
（2）
wherein the content of the first and second substances,k rj which represents the relative permeability of the phases,
Figure 378315DEST_PATH_IMAGE007
the absolute permeability vector is represented by the absolute permeability vector,
Figure 666077DEST_PATH_IMAGE008
which represents the viscosity of the phase(s),p j the pressure of the phase is indicated and,gwhich is indicative of the force of gravity,zrepresenting a depth;
step 1.3, according to 8: 2, dividing the data set and the test set in proportion; the permeability of input data is a tensor of (n, nx, ny, nz,1), the saturation or pressure of output data is a tensor of (n, nx, ny, nz,1, T), n is the number of samples, nx is the number of horizontal grids, ny is the number of vertical grids, nz is the number of grids in the depth direction, and T is the number of time steps.
Further, the specific process of step 2 is as follows:
the 3D convolution Fourier neural network mainly comprises an encoding network and a decoding network, and a cyclic neural module which takes the time sequence influence into consideration and increases the processing time; the input parameter is a permeability field of a known observed value, and the output parameter is saturation or pressure in a future period of time;
the coding network structure comprises M3D convolutional Fourier layers, each of which is structured as follows,
Figure 52059DEST_PATH_IMAGE009
（3）
Figure 441583DEST_PATH_IMAGE010
（4）
wherein the content of the first and second substances,xis the initial input of the parameters that are,Wis the weight of the image,
Figure 305634DEST_PATH_IMAGE011
is an activation function, activated using Relu;v N+1 (x) Is a convolution Fourier moduleNThe output result of the +1 layer is,v N (x) Is thatNOutputting the result of the layer;
Figure 447902DEST_PATH_IMAGE012
the product is a three-dimensional Fourier integral operator used for extracting physical information;
Figure 270365DEST_PATH_IMAGE013
is a fast fourier transform of the signal to be processed,
Figure 6239DEST_PATH_IMAGE014
is an inverse fourier transform of the signal,Cthe operation of cutting off Fourier series exceeding the maximum threshold value is carried out, and the Fourier transform operation can approximate the physical rule which is met by data through the Fourier series, namely a series of sine and cosine curves;Conv3Dthe convolution method is a 3D convolution result and is used for extracting local information, the size and the step of each layer of convolution kernel can be adjusted, the sizes of the convolution kernels are 3 multiplied by 3, and the convolution steps are 1 multiplied by 1 and 2 multiplied by 2; because the Fourier transform is infinite dimension mapping, the dimension of the result extracted by using the Fourier transform is required to be consistent with that of the convolution result, the original dimension cannot be changed after the addition, and the obtained final result enters a decoding network through five layers of Fourier layers;
the decoding network structure is composed of a deconvolution layer, the input of a part of layers of the decoding network is formed by copying and splicing the output matrix corresponding to the coding network and the current input, and the purpose is to store more information and prevent the loss of characteristics in the convolution process; the decoding network is a process of decoding the intermediate result to an original data space, and the obtained result is a pressure value or saturation of a certain time step;
considering the influence of time sequence, enabling intermediate results of a plurality of time steps input through a decoding network to pass through an LSTM structure, combining time sequence information, and obtaining an output result of each time step by using a corresponding decoding network for each feature map after LSTM processing;
LSTM attThe time of day is input three: state of cellC t-1 In a hidden layer stateh t-1 Current input characteristicsx t ；
C t Carrying space-time information, the calculation formula is as follows,
Figure 50812DEST_PATH_IMAGE015
(5)
wherein, the first and the second end of the pipe are connected with each other,f t representing forgetting gate, controlling the state of discarded cellsC t-1 Some of the information in (1);
Figure 922953DEST_PATH_IMAGE016
represents a new candidate cell state that is,i t representative update gate, selecting to preserve candidate cell state
Figure 775372DEST_PATH_IMAGE016
Some of the information in (1);
hidden layer output resulth t From the output gateo t And calculated in the previous stepC t Calculated by the tanh activation function, as follows,
Figure 608330DEST_PATH_IMAGE017
(6)
forgetting doorf t Updating doori t Output dooro t And novel candidate cell states
Figure 814183DEST_PATH_IMAGE016
The calculation formula is as follows, each gate has its own weight and offset, in LSTM
Figure 931044DEST_PATH_IMAGE018
Is a sigmoid activation function, the activation result is between (0, 1), and the tanh activation result is between (-1, 1);
Figure 829729DEST_PATH_IMAGE019
(7)
Figure 382939DEST_PATH_IMAGE020
(8)
Figure 392484DEST_PATH_IMAGE021
(9)
Figure 895009DEST_PATH_IMAGE022
(10)
wherein, the first and the second end of the pipe are connected with each other,W f 、W i 、W o andW c weights corresponding to the forgetting gate, the updating gate, the output gate and the new candidate cell state are respectively set;b f 、b i 、b o andb c biases corresponding to the forgetting gate, the updating gate, the output gate and the new candidate cell state respectively;
inputting a permeability field, obtaining hidden layer variables through a convolution Fourier coding layer, and obtaining a prediction result of saturation or pressure simultaneously containing time information and space information through an LSTM structure and a decoding network
Figure 699017DEST_PATH_IMAGE023
。
Further, the specific process of step 3 is as follows:
step 3.1, setting input and output of a 3D convolution Fourier network model; the input parameter is the permeability of the three-dimensional oil reservoir; setting an output value, and predicting the saturation or pressure distribution of the oil reservoir in a future period of time;
step 3.2, initializing the hyper-parameters, wherein the batch processing size is 8, the activation function uses a ReLu function, and randomly initializing the weightW(ii) a The super-parameters are optimized and adjusted according to network evaluation performance, the loss function uses an MSE function to calculate loss, and the MSE loss function value is as follows during pressure training: the coding and decoding network uses the same super parameter setting and trains to converge in 200 epochs;
loss function of saturationloss sat_MSE The following were used:
Figure 772147DEST_PATH_IMAGE024
（11）
wherein the content of the first and second substances,y sat_i and
Figure 319803DEST_PATH_IMAGE025
are respectively the first in the training setiThe true value of the output saturation of each sample and the saturation calculation result of the 3D convolution Fourier neural network model,n train total number of samples in training set;
for vertical wells drilled through multiple zones, at depthdWellbore pressure at formation
Figure 411255DEST_PATH_IMAGE026
The junction attraction force affected by the change in depth z; the wellbore pressure at the uppermost perforation, called bottom hole pressure, continues down with depthd+1 formation depth wellbore pressure
Figure 386165DEST_PATH_IMAGE027
The calculation is as in equation (12),
Figure 336803DEST_PATH_IMAGE028
is the difference in depth of the image,
Figure 64981DEST_PATH_IMAGE029
is the fluid mix density;
when the 3D convolution Fourier network model predicts the pressure, considering that the pressure in the depth direction is influenced by the gravity action, the pressure satisfies equation (12),
Figure 886306DEST_PATH_IMAGE030
（12）
loss function of pressureloss press_MSE The following were used:
Figure 422330DEST_PATH_IMAGE031
（13）
wherein the content of the first and second substances,y press_i and with
Figure 594685DEST_PATH_IMAGE032
Are respectively the first in the training setiThe real value of the output pressure of each sample and the pressure calculation result of the 3D convolution Fourier neural network model;
Figure 359510DEST_PATH_IMAGE033
at the grid of wellsdThe actual value of wellbore pressure at +1 formation depth,
Figure 300921DEST_PATH_IMAGE034
at the grid of wellsdA predicted value of wellbore pressure at the depth of layer;
Figure 742267DEST_PATH_IMAGE035
are the coefficients of the data constraint and,
Figure 136339DEST_PATH_IMAGE036
is the coefficient of the equation constraint;
the network carries out forward propagation according to the set hyper-parameters and the network iterative structure;
step 3.3, performing back propagation according to the loss error calculated in the step 3.2, and updating the weight by using an ADAM optimizerWGradually reducing the MSE loss; the initial learning rate lr =0.003 of the ADAM optimization algorithm, the batch size is 8, and the mixing weight is 1000; this process is repeated and after 200 epochs the error falls within the acceptance range and the training is complete.
Further, the specific process of step 4 is as follows:
calculating relative Root Mean Square Error (RMSE) for all test samples, and quantifying the relative error of the predicted saturation or pressure;
the RMSE calculation formula is:
Figure 203390DEST_PATH_IMAGE037
（14）
wherein the content of the first and second substances,n test the total number of samples in the test set;
Figure 999308DEST_PATH_IMAGE038
predicting results for the model, including
Figure 877134DEST_PATH_IMAGE039
And
Figure 758502DEST_PATH_IMAGE040
；y i solving true values for numerical simulations, includingy sat_i Andy press_i ；
the formula for calculating the saturation or pressure relative error of the three-dimensional oil reservoir is as follows:
Figure 865130DEST_PATH_IMAGE041
（15）
the relative error of saturation is
Figure 515554DEST_PATH_IMAGE042
The relative error of the pressure is
Figure 298702DEST_PATH_IMAGE043
，n test Is the total number of samples tested and,n M is the total number of the grids,n t is the number of time steps that the user can take,
Figure 667367DEST_PATH_IMAGE044
is the first to testiA sample ofmA gridtThe result of the prediction of the saturation at the time,
Figure 67431DEST_PATH_IMAGE045
is the true value for the corresponding saturation;
Figure 572362DEST_PATH_IMAGE046
is the first to testiA sample ofmA gridtThe result of the prediction of the pressure at the moment,
Figure 791990DEST_PATH_IMAGE047
is the true value of the corresponding pressure; the pressure normalization is performed by subtracting the minimum pressure from the maximum pressure,
Figure 647951DEST_PATH_IMAGE048
is the firstiA test specimentThe maximum pressure at the moment of time is,
Figure 830802DEST_PATH_IMAGE049
is the firstiA test specimentThe minimum pressure at the moment.
Further, the specific process of step 5 is as follows:
storing a 3D convolution Fourier neural network model with better performance index evaluation, storing the trained weight and the super-parameter, and rapidly predicting the future saturation or pressure state of the three-dimensional oil reservoir by using the 3D convolution Fourier neural network model; the input is the permeability field of the three-dimensional reservoir and the output is a period of timetThe internal pressure and saturation distribution are input and output can be obtained through one-time forward propagation calculation;
the yield of the oil-water well is calculated by utilizing the pressure and the saturation through a peakeman model, the calculation formula is as follows,
Figure 580452DEST_PATH_IMAGE050
（16）
wherein the content of the first and second substances,
Figure 846348DEST_PATH_IMAGE051
is the firstmThe rate of injection/recovery of oil/water phase from wells in each grid,k m is as followsmThe absolute permeability of the individual grids is,k r j, is composed ofk m The corresponding oil/water relative permeability,S w m, is as followsmThe value of the water saturation at each grid,
Figure 297927DEST_PATH_IMAGE052
is the thickness of the grid or grid,
Figure 940261DEST_PATH_IMAGE053
by default, to
Figure 544418DEST_PATH_IMAGE054
The reservoir is in both x and y directionsThe width of the two-way pipe is equal,r w is the radius of the borehole,
Figure 981215DEST_PATH_IMAGE055
is the viscosity of the water or oil and,P m andP l are respectively the firstmThe pressure at each grid and the wellbore pressure.
The invention has the following beneficial technical effects:
the invention provides a production dynamic prediction method based on residual oil saturation or pressure of a 3D (three-dimensional) convolution Fourier model, which is characterized in that a convolution Fourier layer is utilized to simultaneously extract local information and physical information, a coding and decoding network is utilized to realize the conversion from permeability to saturation or pressure, an LSTM (least squares) structure is adopted in the network to consider the time sequence problem, and spatial information and time information are simultaneously considered in a neural network. And adding constraints in the loss function by using a physical equation when predicting the pressure, and considering the influence of gravity factors on pressure change in a neural network. The method realizes high-precision rapid prediction of the three-dimensional oil reservoir model, well aims at the characteristics of the space-time property and the physical system of the three-dimensional oil reservoir, and better meets the actual field requirements.
Drawings
FIG. 1 is a schematic flow chart of a three-dimensional reservoir residual oil saturation or pressure production dynamic prediction method based on a 3D convolution Fourier neural network according to the invention;
FIG. 2 is a schematic structural diagram of a 3D convolutional Fourier neural network model of the present invention;
FIG. 3 is a schematic diagram of the structure of the convolutional Fourier layer of the present invention;
FIG. 4 is a schematic diagram of the combination of the encoding and decoding structure of the present invention and LSTM;
FIG. 5 is a detailed schematic diagram of the LSTM structure of the present invention;
FIG. 6 is a graph of relative permeability in an experiment of the present invention;
FIG. 7 is a diagram illustrating the effect of mean square error RMSE in the experiments of the present invention;
FIG. 8 is a comparison of the actual and measured values of well production for production well 1 and production well 2 in the experiment of the present invention;
FIG. 9 is a comparison of the actual and measured well production for production wells 3 and 4 in the present invention experiment.
Detailed Description
The invention is described in further detail below with reference to the following figures and detailed description:
according to the invention, a neural network proxy model is constructed through a 3D convolution Fourier neural network, local information and physical information are simultaneously extracted by utilizing a convolution Fourier layer, and the conversion from permeability to saturation or pressure is realized by utilizing a coding and decoding network. The method can well replace a numerical simulation method for the three-dimensional oil reservoir which accords with the field reality, and can quickly carry out the dynamic prediction of oil reservoir production. Compared with the existing proxy model, the method can better process the three-dimensional oil reservoir model, and has better prediction precision and generalization.
Referring to fig. 1, the invention provides a dynamic prediction method of a three-dimensional underground oil reservoir based on a convolution Fourier neural network, which dynamically predicts the residual oil saturation or pressure of the three-dimensional underground oil reservoir by using a constructed 3D convolution Fourier neural network model, and specifically comprises the following steps:
step 1, collecting three-dimensional underground oil reservoir data, generating a three-dimensional permeability field according with geological characteristics by using an SGeMS geological statistical tool, solving oil reservoir states, namely saturation or pressure, corresponding to different permeabilities within a period of time by using a numerical simulator to serve as a sample library, dividing a training test set data set in proportion, wherein the permeabilities serve as input, and the saturation or pressure serves as output.
The specific process is as follows:
step 1.1, collecting three-dimensional underground oil reservoir data, generating a three-dimensional permeability field by using an SGeMS geological statistical model, wherein the grid size is 40 multiplied by 20, and generating 2000 permeability samples.
And step 1.2, calculating an oil-water two-phase flow equation of the three-dimensional oil reservoir system by using a finite element method through a numerical simulator to obtain saturation or pressure results in all grid blocks. The oil reservoir underground oil-water flow system is three-dimensional, needs to consider the gravity effect, and consists of a mass conservation equation (1) and a phase Darcy velocity equation (2).
The mass conservation equation is expressed as (1),
Figure 421555DEST_PATH_IMAGE056
（1）
wherein the content of the first and second substances,
Figure 336421DEST_PATH_IMAGE057
representing a gradient operator;jrepresenting different phases, including two phases of oil and water,oby which is meant an oil,wrepresents water;
Figure 60664DEST_PATH_IMAGE058
represents the density of the phase;v j represents the phase darcy speed;
Figure 668362DEST_PATH_IMAGE059
representing source/sink items (superscript)lRepresenting a parameter at the well);
Figure 97463DEST_PATH_IMAGE060
represents porosity;S j represents the saturation of the phase;trepresents time;
the expression of the facies darcy velocity equation is shown as (2),
Figure 816020DEST_PATH_IMAGE061
（2）
wherein the content of the first and second substances,k rj which represents the relative permeability of the phases,
Figure 394769DEST_PATH_IMAGE062
the absolute permeability vector is represented by the absolute permeability vector,
Figure 173370DEST_PATH_IMAGE063
which represents the viscosity of the phase(s),p j the pressure of the phase is indicated and,gwhich is indicative of the force of gravity,zindicating the depth.
Step 1.3, according to 8: 2 to divide the data set and test set. The permeability of input data is a tensor of (n, nx, ny, nz,1), the saturation or pressure of output data is a tensor of (n, nx, ny, nz,1, T), n is the number of samples, nx is the number of horizontal grids, ny is the number of vertical grids, nz is the number of grids in the depth direction, and T is the number of time steps.
And 2, constructing a 3D convolution Fourier neural network shown in figures 2-5, combining 3D convolution and Fourier transformation, extracting image local information by using a 3D convolution operator, extracting a physical information approximate differential operator by using Fourier transformation, and adding gravity constraint in the network. And mapping from permeability to reservoir state is realized by utilizing a coding and decoding network, and finally, the saturation or pressure of a certain period of time in the future is output by combining with the LSTM and considering time and space information. The specific process is as follows:
the 3D convolution Fourier neural network mainly comprises an encoding network and a decoding network, and a recurrent neural module (RNN) which increases the processing time in consideration of the influence of time series. The input parameter is a known observed permeability field. The output parameter is the saturation or pressure for a future period of time.
The coding network structure contains M3D convolutional fourier layers, as shown in fig. 4, each of which is structured as follows.
Figure 588301DEST_PATH_IMAGE064
（3）
Figure 110550DEST_PATH_IMAGE065
（4）
Wherein the content of the first and second substances,xis the initial input of the parameters that are,Wis the weight of the image,
Figure 278226DEST_PATH_IMAGE066
is an activation function, activated using Relu;v N+1 (x) Is a convolution Fourier moduleNThe output result of the +1 layer is,v N (x) Is thatNAnd outputting the result of the layer.
Figure 227727DEST_PATH_IMAGE067
The product is a three-dimensional fourier integration operator to extract the physical information.
Figure 362911DEST_PATH_IMAGE068
Is a fast fourier transform of the signal to be processed,
Figure 688850DEST_PATH_IMAGE069
is the inverse of the fourier transform,Cthe operation of cutting off Fourier series exceeding the maximum threshold value, and the Fourier transform operation can approximate the physical law which is met by data through the Fourier series, namely a series of sine and cosine curves.Conv3DThe convolution kernel size and the step size of each layer are adjusted possibly, the convolution kernel size is 3 × 3 × 3, and the convolution step size is 1 × 1 × 1 and 2 × 2 × 2. Since the fourier transform is infinite dimension mapped, the result of using fourier extraction here needs to be consistent with the dimension of the convolution result, without changing the original dimension after addition, through five fourier layers, i.e. layers that do not contain the information in equation (3)Conv3DThe network layer of this operation will get the final result into the decoding network.
The decoding network is realized by deconvolution transformation, and combines the advantages of LSTM in analog time series transformation. The decoding network structure is composed of a deconvolution layer, the input of a part of layers of the decoding network is formed by copying and splicing the output matrix corresponding to the coding network and the current input, and the purpose is to store more information and prevent the loss of characteristics in the convolution process. The decoding network is a process of decoding the intermediate result to the original data space, and the obtained result is the pressure value or the saturation of a certain time step. Considering the influence of time series, the intermediate results of multiple time steps input through the decoding network are passed through the LSTM structure, which is shown in fig. 2, and the output result of each time step is obtained by applying the corresponding decoding network to each feature map after the LSTM processing, in combination with the time series information.
The iterative structure of the LSTM is shown in FIG. 5, the LSTM is attThe time of day is input three: state of the cellC t-1 In a hidden layer stateh t-1 Current input characteristicsx t 。
C t Carrying space-time information, the calculation formula is as follows,
Figure 976612DEST_PATH_IMAGE070
(5)
wherein the content of the first and second substances,f t representing forgetting gate, controlling the state of discarded cellsC t-1 Some of the information in (1);
Figure 97015DEST_PATH_IMAGE071
represents a new candidate cell state that is,i t representative update gate, selecting to preserve candidate cell state
Figure 486539DEST_PATH_IMAGE072
Some of the information in (1);
hidden layer output resulth t From the output gateo t And calculated in the previous stepC t Calculated by the tanh activation function, as follows,
Figure 350590DEST_PATH_IMAGE073
(6)
forgetting doorf t Updating doori t Output dooro t And novel candidate cell states
Figure 492858DEST_PATH_IMAGE074
The calculation formula is as follows, each gate has its own weight and offset, unlike other cases in the network, in LSTM
Figure 49742DEST_PATH_IMAGE075
Is a sigmoid activation function, the activation result is between (0, 1), and the tanh activation result is between (-1, 1).
Figure 896868DEST_PATH_IMAGE076
(7)
Figure 423665DEST_PATH_IMAGE077
(8)
Figure 561385DEST_PATH_IMAGE078
(9)
Figure 554749DEST_PATH_IMAGE079
(10)
Wherein the content of the first and second substances,W f 、W i 、W o andW c weights corresponding to the forgetting gate, the updating gate, the output gate and the new candidate cell state are respectively set;b f 、b i 、b o andb c respectively corresponding biases of a forgetting gate, an updating gate, an output gate and a new candidate cell state;
inputting a permeability field, obtaining hidden layer variables through a convolution Fourier coding layer, and obtaining a prediction result of saturation or pressure simultaneously containing time information and space information through an LSTM structure and a decoding network
Figure 122127DEST_PATH_IMAGE080
。
Fig. 3 shows the details of the specific coding and decoding structure of a single time step, firstly, local and physical information is extracted through 6 layers of fourier convolution layers, the dimension of the intermediate variable is changed, and then, the dimension is not changed through five layers of fourier layers. In the decoding structure, the output of the corresponding coding layer with the same size and the output of the last decoding layer are copied and spliced together. The final intermediate variables are decoded to obtain the prediction result at each time step. The specific change in the shape and size values is shown in Table 1.
And 3, setting hyper-parameters of the convolution Fourier network model, and training the 3D convolution Fourier network model under a training set. The specific process is as follows:
step 3.1, setting input and output of a 3D convolution Fourier network model; the input parameter is the permeability of the three-dimensional oil reservoir; setting an output value, and predicting the oil reservoir saturation or pressure distribution in a future period of time;
step 3.2, initializing the hyper-parameters, wherein the batch processing size is 8, the activation function uses a ReLu function, and randomly initializing the weightW(ii) a The super-parameters are optimized and adjusted according to network evaluation performance, the loss function uses an MSE function to calculate loss, and the MSE loss function value is as follows during pressure training: the coding and decoding network uses the same super parameter setting and trains to converge in 200 epochs;
loss function of saturationloss sat_MSE The following were used:
Figure 718194DEST_PATH_IMAGE081
（11）
wherein the content of the first and second substances,y sat_i and
Figure 710421DEST_PATH_IMAGE082
are respectively the first in the training setiThe true value of the output saturation of each sample and the saturation calculation result of the 3D convolution Fourier neural network model,n train is the total number of samples in the training set.
For vertical wells drilled through multiple zones, at depthdWellbore pressure at formation
Figure 717429DEST_PATH_IMAGE083
Influenced by the junction attraction varying in depth z. The wellbore pressure at the uppermost perforation, referred to as Bottom Hole Pressure (BHP), continues down with depthd+1 formation depth wellbore pressure
Figure 427896DEST_PATH_IMAGE084
The calculation is as in equation (12),
Figure 562074DEST_PATH_IMAGE085
is the difference in depth of the image,
Figure 408807DEST_PATH_IMAGE086
is the fluid mix density.
When the 3D convolution Fourier network model predicts the pressure, in order to improve the prediction accuracy, the influence of gravity on the pressure in the depth direction needs to be considered, the pressure satisfies equation (12),
Figure 353760DEST_PATH_IMAGE087
（12）
loss function of pressureloss press_MSE The following were used:
Figure 285944DEST_PATH_IMAGE088
（13）
wherein the content of the first and second substances,y press_i and
Figure 223813DEST_PATH_IMAGE089
are respectively the first in the training setiThe real value of the output pressure of each sample and the pressure calculation result of the 3D convolution Fourier neural network model;
Figure 190632DEST_PATH_IMAGE090
at the grid of wellsdThe actual value of wellbore pressure at +1 formation depth,
Figure 745635DEST_PATH_IMAGE091
at the grid of wellsdA predicted value of wellbore pressure at the depth of layer;
the invention takes the gravity constraint as a penalty phase in the loss function of the pressure and embeds the physical significance. The equations are trained so that the output meets both the data constraints and the equation constraints,
Figure 555328DEST_PATH_IMAGE092
is a coefficient of data constraint，
Figure 906675DEST_PATH_IMAGE093
The coefficients are constrained by an equation, are hyper-parameters of the network, are respectively initialized to 0.9 and 0.1, and can be automatically updated in the training process.
The network carries out forward propagation according to the set hyper-parameters and the network iterative structure;
step 3.3, performing back propagation according to the loss error calculated in the step 3.2, and updating the weight by using an ADAM optimizerWThe MSE loss is gradually reduced. The initial learning rate lr =0.003 of the ADAM optimization algorithm, the batch size is 8, and the mixing weight is 1000; the training is completed by repeating the process, and the error generally falls within the acceptable range after 200 epochs.
And 4, evaluating the performance of the 3D convolution Fourier neural network by using the test set, and calculating the relative error of RMSE and the quantized saturation or pressure. The specific process is as follows:
for all test samples, the relative root mean square error RMSE is calculated and the relative error of the predicted saturation or pressure is quantified; the RMSE calculation formula is:
Figure 603367DEST_PATH_IMAGE094
（14）
wherein the content of the first and second substances,n test the total number of samples in the test set;
Figure 749177DEST_PATH_IMAGE095
to evaluate the predicted solution, including
Figure 780587DEST_PATH_IMAGE096
And
Figure 935625DEST_PATH_IMAGE097
；y i solving true values for numerical simulations, includingy sat_i Andy press_i . The closer the RMSE value is to 0, the higher the similarity between the two parameters.
The formula for calculating the saturation or pressure relative error of the three-dimensional oil reservoir is as follows:
Figure 985358DEST_PATH_IMAGE041
（15）
the relative error of saturation is
Figure 567649DEST_PATH_IMAGE098
The relative error of the pressure is
Figure 820776DEST_PATH_IMAGE099
，n test Is the total number of test samples 400,n M is the total number of grids 40 x 20=16000,n t it is the number of time steps that is 20,
Figure 779505DEST_PATH_IMAGE100
is the first to testiA sample ofmA gridtThe result of the prediction of the saturation at the time,
Figure 450789DEST_PATH_IMAGE101
is the true value for the corresponding saturation;
Figure 203981DEST_PATH_IMAGE102
is the first to testiA sample ofmA gridtThe result of the prediction of the pressure at the moment,
Figure 944404DEST_PATH_IMAGE103
is the true value of the corresponding pressure. The saturation range is not normalized to 0-1, the pressure is not normalized to this range, the maximum pressure minus the minimum pressure is used for pressure normalization,
Figure 441244DEST_PATH_IMAGE104
is the firstiA test specimentThe maximum pressure at the moment of time is,
Figure 214639DEST_PATH_IMAGE105
is the firstiA test specimentThe minimum pressure at the moment. The closer the relative error is to 0, the more accurate the prediction of the saturation or pressure of the three-dimensional reservoir.
And 5, outputting a 3D convolution Fourier network model with good test indexes after the training is finished. The method comprises the steps of collecting three-dimensional underground oil deposit data in real time by monitoring equipment installed on the basis of an oil deposit, generating a permeability field of the three-dimensional oil deposit on line by using a geological statistical tool, predicting oil deposit saturation or pressure field distribution of the oil deposit within a future period of time on line by using a 3D convolution Fourier network model, further calculating yield through pressure and saturation, and providing visual reference for formulating a production strategy. The specific process is as follows:
and (3) storing the 3D convolution Fourier neural network model with better performance index evaluation, namely storing the trained weight and the hyper-parameter, and rapidly predicting the future saturation or pressure state of the three-dimensional oil reservoir by using the 3D convolution Fourier neural network model. The input is the permeability field of the three-dimensional oil reservoir and the output is a period of timetThe internal pressure and saturation distribution are input and output can be obtained through one-time forward propagation calculation.
The yield of the oil-water well is calculated by utilizing the pressure and the saturation through a peakeman model, the calculation formula is as follows,
Figure 263367DEST_PATH_IMAGE106
（16）
wherein the content of the first and second substances,
Figure 632031DEST_PATH_IMAGE107
is the firstmThe oil/water phase injection/recovery rate of the wells in each grid, which is also the injection and recovery result to be calculated,k m is as followsmThe absolute permeability of the individual grids is,k r j, is composed ofk m The corresponding oil/water relative permeability,S w m, is as followsmThe value of the water saturation at each grid,
Figure 542349DEST_PATH_IMAGE108
is the thickness of the grid or grid,
Figure 47280DEST_PATH_IMAGE109
by default, to
Figure 266909DEST_PATH_IMAGE110
The reservoir has equal width in both x and y directions,r w is the radius of the borehole,
Figure 857290DEST_PATH_IMAGE111
is the viscosity of the water or oil and,P m andP l are respectively the firstmPressure at each grid and wellbore pressure. According to the prediction result, the future state of the oil reservoir can be further analyzed, and visual reference is provided for formulating a production strategy.
To demonstrate the feasibility of the present invention, the following experiments were performed.
The experimental data was derived from a three-dimensional field block in a region, and the model contained 40 x 20 grid blocks, each grid block having dimensions of 20m in the lower x, y, z directions. Four producing wells, two water injection wells, one water injection well and two producing wells jet the upper 10 layers of the reservoir, and the other water injection well and two producing wells jet the bottom 10 layers of the reservoir. All wells were produced under constant pressure conditions, with the bottom pressure of the producing well set at 33.5Mpa and the bottom pressure of the water injection well set at 31.0 Mp. The permeability in each direction x, y, z is assumed to be equal.
Figure 69834DEST_PATH_IMAGE112
Initially 0.25, initial oil saturation was set to 0.90 and initial water saturation was set to 0.10. The relative permeability curves represent the relationship between water saturation and relative permeability, as shown in fig. 6, showing the relative permeability curves for the water phase and the oil phase, respectively. The oil-water flow state of the oil reservoir in 1000 days in the future is simulated, the production period of each section is 50 days, and 20 time steps are total. Firstly, utilizing sequential Gauss simulation to generate random three-dimensional osmotic field whose sample number is 2000 and grid number is 40X 20, utilizing numerical modeThe simulator calculates an oil-water two-phase partial differential equation to calculate a high-fidelity residual oil saturation or pressure field result in a period of time as a data set, 1600 samples are randomly selected as a training set, and the rest 400 samples are used as a test set. The network comprises an encoding structure, a decoding structure and an LSTM structure, the specific structure is shown in table 1, and the sizes of convolution kernels are all 3 x 3. Setting a network hyper-parameter initial value, initializing hyper-parameters (weight and bias of a Fourier convolutional coding decoding network and an LSTM network, contribution coefficient in a loss function and the like) in a network, and using an ADAM optimization algorithm, wherein the initial learning rate lr =0.003, the batch size is 8, and the mixing weight is 1000. The neural network model trains 200 epochs. And testing the neural network model after the training is completed. In the experiment, a saturation field and pressure field map with 50 time steps and 30 days of each time step is output by training a 3D convolution Fourier neural network agent model.
Table 1 fourier convolution network architecture details table
Figure 694851DEST_PATH_IMAGE113
As shown in fig. 7, the mean square error RMSE of the test for the 3D convolved nerves after 200 epoch training was 0.0651, which is close to the training loss and did not produce an overfitting.
Table 2 shows the saturation field and pressure prediction results of the three-dimensional oil reservoir produced by using a 40 × 40 × 20 grid and constant pressure, and it can be seen that the RMSE values of the saturation field and the pressure prediction results are both close to 0, the fitting regression effect is good, and the accuracy is high. Relative error of saturation quantified by equation (15)
Figure 85381DEST_PATH_IMAGE114
Relative error with pressure
Figure 897479DEST_PATH_IMAGE099
Figure 897479DEST_PATH_IMAGE099
5% and 0.67%, respectively, and the relatively low error performance indicates that the neural network has high prediction precision on the three-dimensional reservoir model. The Fourier convolution structure well extracts spatial information in three-dimensional space and simultaneouslyThe LSTM structure takes timing characteristics into good account.
TABLE 2 three-dimensional reservoir saturation or pressure prediction results
Figure 149600DEST_PATH_IMAGE115
FIGS. 8 (production well 1 and production well 2) and 9 (production well 3 and production well 4) show the comparison of the yield calculated by the peakeman equation to the true yield, the predicted yield and the true yield being very close to each other throughout the production cycle, and the well yield being 750m 3 Day to 1000m 3 The fitting between the/day ranges is good, which shows that the 3D convolution Fourier neural network model has good prediction effect.
The experiment can effectively prove that the method can effectively improve the residual oil and pressure prediction speed of the three-dimensional oil reservoir. A Fourier convolution structure and an LSTM structure are adopted in the network to combine the space-time property and the physical significance, and meanwhile, the gravity constraint is considered. The method has better effect on treating the three-dimensional oil reservoir, higher training speed, higher precision, better generalization and more on-site practical value.
It is to be understood that the above description is not intended to limit the present invention, and the present invention is not limited to the above examples, and those skilled in the art may make various changes, modifications, additions and substitutions within the spirit and scope of the present invention.

Claims (4)

1. A three-dimensional underground oil reservoir dynamic prediction method based on a convolution Fourier neural network is characterized in that a built 3D convolution Fourier neural network model is used for dynamically predicting the residual oil saturation or pressure of the three-dimensional underground oil reservoir, and the method specifically comprises the following steps:
step 1, collecting three-dimensional underground oil reservoir data, generating a three-dimensional permeability field according with geological characteristics by using an SGeMS geological statistical tool, solving saturation or pressure within a period of time corresponding to different permeabilities by using a numerical simulator as a sample library, dividing a data set according to a proportion, taking the permeability as input, and taking the saturation or pressure as output; the specific process is as follows:
step 1.1, collecting three-dimensional underground oil reservoir data, generating a three-dimensional permeability field by using an SGeMS geological statistical model, wherein the grid size is 40 multiplied by 20, and generating 2000 permeability samples;
step 1.2, calculating an oil-water two-phase flow equation of the three-dimensional oil reservoir system by using a finite element method through a numerical simulator to obtain saturation or pressure results in all grid blocks; the oil reservoir underground oil-water flow system is three-dimensional, needs to consider the gravity effect and consists of a mass conservation equation (1) and a phase Darcy velocity equation (2);
the mass conservation equation is expressed as (1),
Figure DEST_PATH_IMAGE002
（1）
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE004
representing a gradient operator;jindicating different phases, including oil and water phases,oby which is meant an oil,wrepresents water;
Figure DEST_PATH_IMAGE006
represents the density of the phase;v j represents the phase darcy speed;
Figure DEST_PATH_IMAGE008
representing source/sink items, superscriptlRepresenting a parameter at the well;
Figure DEST_PATH_IMAGE010
represents porosity;S j represents the saturation of the phase;trepresents time;
the expression of the facies darcy velocity equation is shown as (2),
Figure DEST_PATH_IMAGE012
（2）
wherein the content of the first and second substances,k rj which represents the relative permeability of the phases,
Figure DEST_PATH_IMAGE014
the absolute permeability vector is represented by the absolute permeability vector,
Figure DEST_PATH_IMAGE016
the viscosity of the phase is expressed in terms of,p j the pressure of the phase is indicated and,gwhich is indicative of the force of gravity,zrepresenting a depth;
step 1.3, according to 8: 2, dividing the data set and the test set in proportion; the method comprises the steps that a tensor with (n, nx, ny, nz,1) permeability is input, a tensor with (n, nx, ny, nz,1, T) saturation or pressure of output data is output, n is a sample number, nx is a transverse grid number, ny is a longitudinal grid number, nz is a grid number in the depth direction, and T is a time step number;
step 2, constructing a 3D convolution Fourier neural network, combining 3D convolution and Fourier transform, extracting image local information by using a 3D convolution operator, extracting a physical information approximate differential operator by using Fourier transform, and adding gravity constraint in the network; mapping from permeability to an oil reservoir state is realized by utilizing a coding and decoding network, and time and space information is considered simultaneously by combining with an LSTM; the specific process is as follows:
the 3D convolution Fourier neural network comprises an encoding network and a decoding network, and a cyclic neural module which takes the time sequence influence into consideration and increases the processing time; the input parameter is a permeability field of a known observed value, and the output parameter is saturation or pressure in a future period of time;
the coding network structure comprises M3D convolutional Fourier layers, each of which is structured as follows,
Figure DEST_PATH_IMAGE018
（3）
Figure DEST_PATH_IMAGE020
（4）
wherein the content of the first and second substances,xis the initial input of the parameters that are,Wis the weight of the image,
Figure DEST_PATH_IMAGE022
is an activation function, activated using Relu;v N+1 (x) Is a convolution Fourier moduleNThe output result of the +1 layer is,v N (x) Is thatNOutputting the result of the layer;
Figure DEST_PATH_IMAGE024
the product is a three-dimensional Fourier integral operator used for extracting physical information;
Figure DEST_PATH_IMAGE026
is a fast fourier transform of the signal to be processed,
Figure DEST_PATH_IMAGE028
is an inverse fourier transform of the signal,Cthe operation of cutting off Fourier series exceeding the maximum threshold value is carried out, and the Fourier transform operation can approximate the physical rule which is met by data through the Fourier series, namely a series of sine and cosine curves;Conv3Dthe convolution method is a 3D convolution result and is used for extracting local information, the size and the step of each layer of convolution kernel can be adjusted, the sizes of the convolution kernels are 3 multiplied by 3, and the convolution steps are 1 multiplied by 1 and 2 multiplied by 2; because the Fourier transform is infinite dimension mapping, the dimension of the result extracted by using the Fourier transform is required to be consistent with that of the convolution result, the original dimension is not changed after the Fourier transform is added, and the final result obtained enters a decoding network after five layers of Fourier transform are performed;
the decoding network structure is composed of a deconvolution layer, the input of a part of layers of the decoding network is formed by copying and splicing the output matrix corresponding to the coding network and the current input, and the purpose is to store more information and prevent the loss of characteristics in the convolution process; the decoding network is a process of decoding the intermediate result to an original data space, and the obtained result is a pressure value or saturation of a certain time step;
considering the influence of time sequence, enabling intermediate results of a plurality of time steps input through a decoding network to pass through an LSTM structure, combining time sequence information, and obtaining an output result of each time step by using a corresponding decoding network for each feature map after LSTM processing;
LSTM attThe time of day is input three: state of the cellC t-1 In a hidden layer stateh t-1 Current input characteristicsx t ；
C t Carrying space-time information, the calculation formula is as follows,
Figure DEST_PATH_IMAGE030
(5)
wherein the content of the first and second substances,f t representing forgetting gate, controlling the state of discarded cellsC t-1 Some of the information in (1);
Figure DEST_PATH_IMAGE032
represents a new candidate cell state that is,i t representative update gate, selecting to preserve candidate cell state
Figure 944378DEST_PATH_IMAGE032
Some of the information in (1);
hidden layer output resulth t From the output gateo t And calculated in the previous stepC t Calculated by the tanh activation function, as follows,
Figure DEST_PATH_IMAGE034
(6)
forgetting doorf t Updating doori t Output dooro t And novel candidate cell states
Figure 64781DEST_PATH_IMAGE032
The calculation formula is as follows, each gate has its own weight and offset, in LSTM
Figure DEST_PATH_IMAGE036
Is a sigmoid activation function, the activation result is between (0, 1), and the tanh activation result is between (-1, 1);
Figure DEST_PATH_IMAGE038
(7)
Figure DEST_PATH_IMAGE040
(8)
Figure DEST_PATH_IMAGE042
(9)
Figure DEST_PATH_IMAGE044
(10)
wherein the content of the first and second substances,W f 、W i 、W o andW c weights corresponding to the forgetting gate, the updating gate, the output gate and the new candidate cell state are respectively set;b f 、b i 、b o andb c biases corresponding to the forgetting gate, the updating gate, the output gate and the new candidate cell state respectively;
inputting a permeability field, obtaining hidden layer variables through a convolution Fourier coding layer, and obtaining a prediction result of saturation or pressure simultaneously containing time information and space information through an LSTM structure and a decoding network
Figure DEST_PATH_IMAGE046
；
Step 3, setting hyper-parameters of the convolution Fourier network model, and training the 3D convolution Fourier network model under a training set;
step 4, evaluating the performance of the 3D convolution Fourier neural network by using the test set, and calculating the relative error between RMSE and the quantized saturation or pressure;
step 5, outputting a 3D convolution Fourier network model with good test indexes after the training is finished; the method comprises the steps of collecting three-dimensional underground oil deposit data in real time by monitoring equipment installed on the basis of an oil deposit, generating a permeability field of the three-dimensional oil deposit on line by using a geological statistical tool, predicting oil deposit saturation or pressure field distribution of the oil deposit within a future period of time on line by using a 3D convolution Fourier network model, further calculating yield through pressure and saturation, and providing visual reference for formulating a production strategy.
2. The method for predicting the dynamic of the three-dimensional underground oil reservoir based on the convolutional Fourier neural network as set forth in claim 1, wherein the specific process of step 3 is as follows:
step 3.1, setting input and output of a 3D convolution Fourier network model; the input parameter is the permeability of the three-dimensional oil reservoir; setting an output value, and predicting the oil reservoir saturation or pressure distribution in a future period of time;
step 3.2, initializing the hyper-parameters, wherein the batch processing size is 8, the activation function uses a ReLu function, and randomly initializing the weightW(ii) a The super-parameters are optimized and adjusted according to network evaluation performance, the loss function uses the MSE function to calculate loss, and the MSE loss function value is as follows during pressure training: the coding and decoding network uses the same super parameter setting and trains to converge in 200 epochs;
loss function of saturationloss sat_MSE The following:
Figure DEST_PATH_IMAGE048
（11）
wherein the content of the first and second substances,y sat_i and with
Figure DEST_PATH_IMAGE050
Are respectively the first in the training setiOutput saturation true of individual samplesThe saturation calculation results of the real-valued and 3D convolutional fourier neural network models,n train total number of samples in training set;
for vertical wells drilled through multiple zones, at depthdWellbore pressure at formation
Figure DEST_PATH_IMAGE052
The junction attraction force affected by the change in depth z; the wellbore pressure at the uppermost perforation, called bottom hole pressure, continues down with depthd+1 formation depth wellbore pressure
Figure DEST_PATH_IMAGE054
The calculation is as in equation (12),
Figure DEST_PATH_IMAGE056
is the difference in depth of the image,
Figure DEST_PATH_IMAGE058
is the fluid mix density;
when the 3D convolution Fourier network model predicts the pressure, considering that the pressure in the depth direction is influenced by the gravity action, the pressure satisfies equation (12),
Figure DEST_PATH_IMAGE060
（12）
loss function of pressureloss press_MSE The following were used:
Figure DEST_PATH_IMAGE062
（13）
wherein the content of the first and second substances,y press_i and
Figure DEST_PATH_IMAGE064
are respectively the first in the training setiThe real value of the output pressure of each sample and the pressure calculation result of the 3D convolution Fourier neural network model;
Figure DEST_PATH_IMAGE066
at the grid of wellsdThe actual value of wellbore pressure at +1 formation depth,
Figure DEST_PATH_IMAGE068
at the grid of wellsdA predicted value of wellbore pressure at the depth of layer;
Figure DEST_PATH_IMAGE070
are the coefficients of the data constraint and,
Figure DEST_PATH_IMAGE072
is the coefficient of the equation constraint;
the network carries out forward propagation according to the set hyper-parameters and the network iterative structure;
step 3.3, performing back propagation according to the loss error calculated in the step 3.2, and updating the weight by using an ADAM optimizerWGradually reducing the MSE loss; the initial learning rate lr =0.003, the batch size is 8, and the mixing weight is 1000 for the ADAM optimization algorithm; this process is repeated, and after 200 epochs the error falls within the acceptance range and training is complete.
3. The method for dynamically predicting the three-dimensional underground oil reservoir based on the convolutional Fourier neural network as claimed in claim 2, wherein the specific process of the step 4 is as follows:
calculating relative Root Mean Square Error (RMSE) for all test samples, and quantifying the relative error of the predicted saturation or pressure;
the RMSE calculation formula is:
Figure DEST_PATH_IMAGE074
（14）
wherein the content of the first and second substances,n test the total number of samples in the test set;
Figure DEST_PATH_IMAGE076
predicting results for the model, including
Figure DEST_PATH_IMAGE078
And
Figure DEST_PATH_IMAGE080
；y i solving true values for numerical simulations, includingy sat_i Andy press_i ；
the formula for calculating the saturation or pressure relative error of the three-dimensional reservoir is as follows:
Figure DEST_PATH_IMAGE082
（15）
the relative error of saturation is
Figure DEST_PATH_IMAGE084
The relative error of the pressure is
Figure DEST_PATH_IMAGE086
，n test Is the total number of samples tested and,n M is the total number of the grids,n t is the number of time steps taken,
Figure DEST_PATH_IMAGE088
is the first to testiA sample ofmA gridtThe result of the prediction of the saturation at the time,
Figure DEST_PATH_IMAGE090
is the true value for the corresponding saturation;
Figure DEST_PATH_IMAGE092
is the first to testiA sample ofmA gridtThe result of the prediction of the pressure at the moment,
Figure DEST_PATH_IMAGE094
is the true value of the corresponding pressure; the pressure normalization is performed by subtracting the minimum pressure from the maximum pressure,
Figure DEST_PATH_IMAGE096
is the firstiA test specimentThe maximum pressure at the moment of time is,
Figure DEST_PATH_IMAGE098
is the firstiA test specimentThe minimum pressure at the moment.
4. The method for dynamically predicting the three-dimensional underground oil reservoir based on the convolutional Fourier neural network as claimed in claim 3, wherein the specific process of the step 5 is as follows:
storing a 3D convolution Fourier neural network model with better performance index evaluation, storing the trained weight and the super-parameter, and rapidly predicting the future saturation or pressure state of the three-dimensional oil reservoir by using the 3D convolution Fourier neural network model; the input is the permeability field of the three-dimensional reservoir and the output is a period of timetThe internal pressure and saturation distribution are input and output can be obtained through one-time forward propagation calculation;
the yield of the oil-water well is calculated by utilizing the pressure and the saturation through a peakeman model, the calculation formula is as follows,
Figure DEST_PATH_IMAGE100
（16）
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE102
is the firstmThe rate of injection/recovery of oil/water phase from the wells in the grid,k m is as followsmThe absolute permeability of the individual grids is,k r j, is composed ofk m The corresponding oil/water relative permeability,S w m, is as followsmThe value of the water saturation at each grid,
Figure DEST_PATH_IMAGE104
is the thickness of the grid or grid,
Figure DEST_PATH_IMAGE106
by default, to
Figure DEST_PATH_IMAGE108
The reservoir has equal width in both x and y directions,r w is the radius of the borehole and,
Figure DEST_PATH_IMAGE110
is the viscosity of the water or oil and,P m andP l are respectively the firstmThe pressure at each grid and the wellbore pressure.
CN202210603016.6A 2022-05-31 2022-05-31 Three-dimensional underground oil reservoir dynamic prediction method based on convolution Fourier neural network Active CN114693005B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210603016.6A CN114693005B (en) 2022-05-31 2022-05-31 Three-dimensional underground oil reservoir dynamic prediction method based on convolution Fourier neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210603016.6A CN114693005B (en) 2022-05-31 2022-05-31 Three-dimensional underground oil reservoir dynamic prediction method based on convolution Fourier neural network

Publications (2)

Publication Number Publication Date
CN114693005A CN114693005A (en) 2022-07-01
CN114693005B true CN114693005B (en) 2022-08-26

Family

ID=82131298

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210603016.6A Active CN114693005B (en) 2022-05-31 2022-05-31 Three-dimensional underground oil reservoir dynamic prediction method based on convolution Fourier neural network

Country Status (1)

Country Link
CN (1) CN114693005B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115204530B (en) * 2022-09-16 2023-05-23 中科数智能源科技(深圳)有限公司 Oil reservoir prediction method based on Fourier neural operator and cyclic neural network
CN115204531B (en) * 2022-09-16 2022-12-27 中科数智能源科技(深圳)有限公司 Oil reservoir prediction method, equipment and medium based on Fourier neural operator
CN115618750B (en) * 2022-12-16 2023-04-07 中国石油大学(华东) Underground oil-water seepage agent model based on coupling neural network
CN117494617B (en) * 2023-12-29 2024-04-16 中国石油大学(华东) Carbon dioxide flooding quick simulation method based on embedded physical information neural network

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5251286A (en) * 1992-03-16 1993-10-05 Texaco, Inc. Method for estimating formation permeability from wireline logs using neural networks
CN102061910A (en) * 2010-06-30 2011-05-18 中国石油大学(北京) Method for establishing waterflooding extraction predictable physical model of sandstone oil deposit
CN106291700A (en) * 2016-09-28 2017-01-04 西安交通大学 Based on the earthquake weighted average instantaneous frequency distilling method synchronizing extruding conversion
CN111561313A (en) * 2020-05-15 2020-08-21 中国地质大学（北京） Compact sandstone reservoir parameter prediction method based on physical model and machine learning
CN112541572A (en) * 2020-11-25 2021-03-23 中国石油大学(华东) Residual oil distribution prediction method based on convolutional encoder-decoder network
CN113052371A (en) * 2021-03-16 2021-06-29 中国石油大学(华东) Residual oil distribution prediction method and device based on deep convolutional neural network
CN114492213A (en) * 2022-04-18 2022-05-13 中国石油大学(华东) Wavelet neural operator network model-based residual oil saturation and pressure prediction method

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110967471A (en) * 2019-11-14 2020-04-07 广东电网有限责任公司 Method for predicting concentration of dissolved gas in transformer oil
US20220101494A1 (en) * 2020-09-30 2022-03-31 Nvidia Corporation Fourier transform-based image synthesis using neural networks
CN112100930B (en) * 2020-11-11 2021-02-02 中国石油大学(华东) Formation pore pressure calculation method based on convolutional neural network and Eaton formula
CN114492211B (en) * 2022-04-15 2022-07-12 中国石油大学(华东) Residual oil distribution prediction method based on autoregressive network model

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5251286A (en) * 1992-03-16 1993-10-05 Texaco, Inc. Method for estimating formation permeability from wireline logs using neural networks
CN102061910A (en) * 2010-06-30 2011-05-18 中国石油大学(北京) Method for establishing waterflooding extraction predictable physical model of sandstone oil deposit
CN106291700A (en) * 2016-09-28 2017-01-04 西安交通大学 Based on the earthquake weighted average instantaneous frequency distilling method synchronizing extruding conversion
CN111561313A (en) * 2020-05-15 2020-08-21 中国地质大学（北京） Compact sandstone reservoir parameter prediction method based on physical model and machine learning
CN112541572A (en) * 2020-11-25 2021-03-23 中国石油大学(华东) Residual oil distribution prediction method based on convolutional encoder-decoder network
CN113052371A (en) * 2021-03-16 2021-06-29 中国石油大学(华东) Residual oil distribution prediction method and device based on deep convolutional neural network
CN114492213A (en) * 2022-04-18 2022-05-13 中国石油大学(华东) Wavelet neural operator network model-based residual oil saturation and pressure prediction method

Non-Patent Citations (6)

* Cited by examiner, † Cited by third party
Title
《各向异性油藏水平井变流量压力动态响应计算方法》;孟宪伟等;《石油钻采工艺》;20190331;第41卷(第2期);全文 *
A Novel Preprocessing Scheme to Improve the Prediction of Sand Fraction From Seismic Attributes Using Neural Networks;Soumi Chaki;《IEEE XPlore》;20150325;全文 *
径向基函数神经网络法致密砂岩储层相对渗透率预测与含水率计算;王谦等;《石油地球物理勘探》;20200815(第04期);全文 *
快速的卷积神经网络算法及应用;包志强等;《计算机工程与设计》;20200816(第08期);全文 *
断块油藏高部位剩余油径向钻孔产能预测;张凯等;《油气地质与采收率》;20170930;第24卷(第5期);全文 *
灰岩裂缝储层地震识别技术研究——以四扣地区为例;何欣;《中国优秀硕士学位论文全文数据库 (基础科学辑)》;20110415;全文 *

Also Published As

Publication number Publication date
CN114693005A (en) 2022-07-01

Similar Documents

Publication Publication Date Title
CN114693005B (en) Three-dimensional underground oil reservoir dynamic prediction method based on convolution Fourier neural network
Jiang et al. Deep residual U-net convolution neural networks with autoregressive strategy for fluid flow predictions in large-scale geosystems
CN113052371B (en) Residual oil distribution prediction method and device based on deep convolutional neural network
CN112541572B (en) Residual oil distribution prediction method based on convolutional encoder-decoder network
CN106355003B (en) Markov chain Monte-Carlo automatic history matching method and system based on t distributions
CN114492213B (en) Wavelet neural operator network model-based residual oil saturation and pressure prediction method
CN112818591B (en) Physical constraint-based method for predicting tight oil fracturing range by using DL model
CN106295199A (en) Automatic history matching method and system based on autocoder and multiple-objection optimization
CN111523713A (en) Method and device for predicting residual oil saturation distribution in oil field
Ng et al. Smart proxy modeling of a fractured reservoir model for production optimization: implementation of metaheuristic algorithm and probabilistic application
CN112016212B (en) Reservoir longitudinal heterogeneity evaluation method based on seepage control equation
Bai et al. Accelerating geostatistical modeling using geostatistics-informed machine Learning
Calvette et al. Forecasting smart well production via deep learning and data driven optimization
Han et al. Multiple point geostatistical simulation with adaptive filter derived from neural network for sedimentary facies classification
Chen et al. Physics-guided meta-learning method in baseflow prediction over large regions
Razak et al. Embedding physical flow functions into deep learning predictive models for improved production forecasting
Le et al. Physics-constrained deep learning for production forecast in tight reservoirs
Ju et al. Hydrologic simulations with artificial neural networks
Wen et al. Multiphase flow prediction with deep neural networks
Costa et al. A new methodology to reduce uncertainty of global attributes in naturally fractured reservoirs
Silva et al. Generative network-based reduced-order model for prediction, data assimilation and uncertainty quantification
Wang et al. On the Feasibility of An Ensemble Multi-Fidelity Neural Network for Fast Data Assimilation for Subsurface Flow in Porous Media
Azizzadeh Mehmandoust Olya et al. Hydrocarbon reservoir potential mapping through Permeability estimation by a CUDNNLSTM Deep Learning Algorithm
WO2021160980A1 (en) An improved micro-ct image acquisition method
Zhang et al. Conditional stochastic simulation of fluvial reservoirs using multi-scale concurrent generative adversarial networks

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